Conference Agenda

During registration, you will receive your nametag and your lunch voucher, to be redeemed for a meal at the Helmholtz Munich Mensa. The registration/info desk will be open in the Helmholtz Munich Conference Center from 8:00 until 16:45, when the last session starts. Limited lockers are also available for your small luggage in the building.

Brief Description and Outline:\nModern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. Learning to distribute data across multiple GPUs during deep learning model training makes possible an incredible wealth of new applications utilizing deep learning. \nAdditionally, the effective use of systems with multiple GPUs reduces training time, allowing for faster application development and much faster iteration cycles. Teams who are able to perform training using multiple GPUs will have an edge, building models trained on more data in shorter periods of time and with greater engineer productivity. \nThis workshop teaches you techniques for data-parallel deep learning training on multiple GPUs to shorten the training time required for data-intensive applications. Working with deep learning tools, frameworks, and workflows to perform neural network training, you’ll learn how to decrease model training time by distributing data to multiple GPUs, while retaining the accuracy of training on a single GPU. \n-\nOutline:\n- Introduction \n- Stochastic Gradient Descent and the Effects of Batch Size \n- Training on Multiple GPUs with PyTorch Distributed Data Parallel (DDP) \n- Maintaining Model Accuracy when Scaling to Multiple GPUs \n- Workshop Assessment and Final review\n-\nGoals:\n- Understand how data parallel deep learning training is performed using multiple GPUs - Achieve maximum throughput when training, for the best use of multiple GPUs - Distribute training to multiple GPUs using Pytorch Distributed Data Parallel - Understand and utilize algorithmic considerations specific to multi-GPU training performance and accuracy\nPresenters Experience:\nSeverine Habert, Senior Solutions Architect, Nvidia\nPallavi Mohan and Nael Fasfous, Scientist, Nvidia\n-\nTarget Audience:\nthe course is adequate for anyone with experience on training deep learning models in Python. \nAttendees to bring their own laptops and be responsible for running websockets test on their system before arriving. http://websocketstest.courses.nvidia.com/\n-\nKeywords:\nMulti-GPU Deep Learning, PyTorch Distributed Data Parallel (DDP), GPU Computing, AI Model Optimization.

Brief Description and Outline:\nModern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. Learning to distribute data across multiple GPUs during deep learning model training makes possible an incredible wealth of new applications utilizing deep learning. \nAdditionally, the effective use of systems with multiple GPUs reduces training time, allowing for faster application development and much faster iteration cycles. Teams who are able to perform training using multiple GPUs will have an edge, building models trained on more data in shorter periods of time and with greater engineer productivity. \nThis workshop teaches you techniques for data-parallel deep learning training on multiple GPUs to shorten the training time required for data-intensive applications. Working with deep learning tools, frameworks, and workflows to perform neural network training, you’ll learn how to decrease model training time by distributing data to multiple GPUs, while retaining the accuracy of training on a single GPU. \n-\nOutline:\n- Introduction \n- Stochastic Gradient Descent and the Effects of Batch Size \n- Training on Multiple GPUs with PyTorch Distributed Data Parallel (DDP) \n- Maintaining Model Accuracy when Scaling to Multiple GPUs \n- Workshop Assessment and Final review\n-\nGoals:\n- Understand how data parallel deep learning training is performed using multiple GPUs - Achieve maximum throughput when training, for the best use of multiple GPUs - Distribute training to multiple GPUs using Pytorch Distributed Data Parallel - Understand and utilize algorithmic considerations specific to multi-GPU training performance and accuracy\nPresenters Experience:\nSeverine Habert, Senior Solutions Architect, Nvidia\nPallavi Mohan and Nael Fasfous, Scientist, Nvidia\n\nTarget Audience:\nthe course is adequate for anyone with experience on training deep learning models in Python. \nAttendees to bring their own laptops and be responsible for running websockets test on their system before arriving. http://websocketstest.courses.nvidia.com/\n-\nKeywords:\nMulti-GPU Deep Learning, PyTorch Distributed Data Parallel (DDP), GPU Computing, AI Model Optimization.

Brief Description and Outline:\nModern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. Learning to distribute data across multiple GPUs during deep learning model training makes possible an incredible wealth of new applications utilizing deep learning. \nAdditionally, the effective use of systems with multiple GPUs reduces training time, allowing for faster application development and much faster iteration cycles. Teams who are able to perform training using multiple GPUs will have an edge, building models trained on more data in shorter periods of time and with greater engineer productivity. \nThis workshop teaches you techniques for data-parallel deep learning training on multiple GPUs to shorten the training time required for data-intensive applications. Working with deep learning tools, frameworks, and workflows to perform neural network training, you’ll learn how to decrease model training time by distributing data to multiple GPUs, while retaining the accuracy of training on a single GPU. \n-\nOutline:\n- Introduction \n- Stochastic Gradient Descent and the Effects of Batch Size \n- Training on Multiple GPUs with PyTorch Distributed Data Parallel (DDP) \n- Maintaining Model Accuracy when Scaling to Multiple GPUs \n- Workshop Assessment and Final review\n-\nGoals:\n- Understand how data parallel deep learning training is performed using multiple GPUs - Achieve maximum throughput when training, for the best use of multiple GPUs - Distribute training to multiple GPUs using Pytorch Distributed Data Parallel - Understand and utilize algorithmic considerations specific to multi-GPU training performance and accuracy\nPresenters Experience:\nSeverine Habert, Senior Solutions Architect, Nvidia\nPallavi Mohan and Nael Fasfous, Scientist, Nvidia\n---\nTarget Audience:\nthe course is adequate for anyone with experience on training deep learning models in Python.\nAttendees to bring their own laptops and be responsible for running websockets test on their system before arriving. http://websocketstest.courses.nvidia.com/\n-\nKeywords:\nMulti-GPU Deep Learning, PyTorch Distributed Data Parallel (DDP), GPU Computing, AI Model Optimization.

Brief Description and Outline:\nModern deep learning challenges leverage increasingly larger datasets and more complex models. As a result, significant computational power is required to train models effectively and efficiently. Learning to distribute data across multiple GPUs during deep learning model training makes possible an incredible wealth of new applications utilizing deep learning. \nAdditionally, the effective use of systems with multiple GPUs reduces training time, allowing for faster application development and much faster iteration cycles. Teams who are able to perform training using multiple GPUs will have an edge, building models trained on more data in shorter periods of time and with greater engineer productivity. \nThis workshop teaches you techniques for data-parallel deep learning training on multiple GPUs to shorten the training time required for data-intensive applications. Working with deep learning tools, frameworks, and workflows to perform neural network training, you’ll learn how to decrease model training time by distributing data to multiple GPUs, while retaining the accuracy of training on a single GPU. \n-\nOutline:\n- Introduction \n- Stochastic Gradient Descent and the Effects of Batch Size \n- Training on Multiple GPUs with PyTorch Distributed Data Parallel (DDP) \n- Maintaining Model Accuracy when Scaling to Multiple GPUs \n- Workshop Assessment and Final review\n-\nGoals:\n- Understand how data parallel deep learning training is performed using multiple GPUs - Achieve maximum throughput when training, for the best use of multiple GPUs - Distribute training to multiple GPUs using Pytorch Distributed Data Parallel - Understand and utilize algorithmic considerations specific to multi-GPU training performance and accuracy\nPresenters Experience:\nSeverine Habert, Senior Solutions Architect, Nvidia\nPallavi Mohan and Nael Fasfous, Scientist, Nvidia\n--\nTarget Audience:\nthe course is adequate for anyone with experience on training deep learning models in Python. \nAttendees to bring their own laptops and be responsible for running websockets test on their system before arriving. http://websocketstest.courses.nvidia.com/\n-\nKeywords:\nMulti-GPU Deep Learning, PyTorch Distributed Data Parallel (DDP), GPU Computing, AI Model Optimization.

Brief Description and Outline:\nLarge language models and agentic AI systems are increasingly being adopted as tools in scientific workflows —from literature review and data analysis to experimental design and code generation.\nYet building effective agentic tools for research remains as much craft as science: architectural choices around retrieval strategies, tool-use protocols, embedding models, and human-in-the-loop interfaces can make or break a system’s practical utility. At the same time, the rapid pace of development means that hard-won lessons about what works (and what fails) are rarely shared systematically across projects.\n\nThis workshop brings together developers and users of agentic ML tools for science to openly exchange practical knowledge and experience. Rather than focusing on algorithmic novelty, we emphasize engineering reality: Which design patterns lead to reliable agentic behavior? Where do current LLM-based tools fall short in scientific contexts? How do we evaluate whether an agentic tool genuinely accelerates research or merely creates an illusion of productivity?\n\nAs a basis for discussion, we present concrete example projects that tackle different facets of agentic scientific tooling —from semantic literature exploration with retrieval-augmented generation and clustering MCP tools to machine learning project planning systems. By dissecting these systems in detail, including their failures and limitations, we aim to distill transferable principles for the Community.\n\nThe workshop combines presentation-driven discussion in the morning with hands-on collabora- tive work in the afternoon. Morning sessions feature short project presentations interleaved with structured discussion. The afternoon hackathon lets participants turn ideas into code or concrete plans in small groups. Participants are encouraged to bring their own experiences, tools, and “war stories” to contribute throughout.\n-\nPresentations — Architecture walk-through, live demo, design trade-offs, and lessons learned & Discussion (2 h)\n-\nTime Session\n~30 min Introduction to Agentic AI for Science\n~30 min Project Presentation: Abstracts Explorer\n~30 min Project Presentation: Voucher Canvas Agent\n~15 min Break\n~15 min General Discussion —Cross-cutting themes: What design patterns were transferred? Where did things break? What would we do differently?\n-\nHackathon (2 h)\nParticipants form small groups to build or plan agentic ML tools hands-on. Groups self-organize around topics proposed in the morning discussion. Possible tracks include:\n• Feature sprints —Implement a concrete feature in one of the presented projects (e.g., a new MCP tool, a conference downloader plugin, an improved clustering pipeline). Bring a laptop and be ready to write code.\n• Project planning —Sketch the architecture and roadmap for a new agentic tool addressing a participant’s own research domain. Output: a brief design document or project proposal.\n• Integration experiments —Wire up existing tools via MCP or other protocols and stress-test them with real scientific queries. Time Session ~15 min Group Formation —Pitch topics, form groups (3–5 people). ~90 min Hands-on Work —Small-group hacking / planning with roaming support from organizers. ~15 min Lightning Reports —Each group presents what they built, learned, or planned (~2 min each).\n-\nExample Projects\n\n\nAbstracts Explorer — An open-source Python toolkit combining LLM-based semantic search, unsupervised clustering, and RAG to explore conference proceedings (NeurIPS, ICLR, ICML) at scale. Features native MCP integration for LLM-driven topic and trend analysis, a Flask web interface, and Docker-based deployment. (github.com/thawn/abstracts-explorer)\n\n\nVoucher Canvas Agent — an AI-powered exploration agent designed to help users plan machine learning projects. It uses a multi-container architecture with a LangGraph agent and a custom exploration server\n\n\n-\nGoals:\n• Architecture & Integration: How should agentic tools expose capabilities to LLMs (e.g., via the Model Context Protocol)? What are effective patterns for combining retrieval, structured analysis, and generation?\n• Evaluation & Trust: How do we measure whether an agentic science tool is actually helpful?\nWhen should we trust LLM-generated analyses of scientific literature?\n• Practical Pitfalls: What are common failure modes —hallucinated references, embedding model drift, brittle tool-use chains —and how can they be mitigated?\n• Human-in-the-Loop: What level of autonomy is appropriate? Where must human oversight remain, and how do we design interfaces that support it?\n-\nPresenters Experience:\nTill Korten is a Helmholtz AI Consultant and developer of Abstracts Explorer.\nHaider Khan is a Helmholtz AI Consultant and developer of Voucher Canvas Agent.\n-\nTarget Audience:\nResearchers, research software engineers, and data scientists who are building, evaluating, or considering agentic ML tools in their scientific workflows. No prior experience with agent frameworks is required — practical curiosity and a willingness to share experiences are the main prerequisites. For the afternoon hackathon, a laptop with Python 3.11+ and Git installed is recommended.\n-\nKeywords:\nAgentic AI; scientific workflows; human–AI interaction; evaluation and benchmarking; research infrastructure; AI governance

Brief Description and Outline:\nLarge language models and agentic AI systems are increasingly being adopted as tools in scientific workflows —from literature review and data analysis to experimental design and code generation.\nYet building effective agentic tools for research remains as much craft as science: architectural choices around retrieval strategies, tool-use protocols, embedding models, and human-in-the-loop interfaces can make or break a system’s practical utility. At the same time, the rapid pace of development means that hard-won lessons about what works (and what fails) are rarely shared systematically across projects.\n\nThis workshop brings together developers and users of agentic ML tools for science to openly exchange practical knowledge and experience. Rather than focusing on algorithmic novelty, we emphasize engineering reality: Which design patterns lead to reliable agentic behavior? Where do current LLM-based tools fall short in scientific contexts? How do we evaluate whether an agentic tool genuinely accelerates research or merely creates an illusion of productivity?\n\nAs a basis for discussion, we present concrete example projects that tackle different facets of agentic scientific tooling —from semantic literature exploration with retrieval-augmented generation and clustering MCP tools to machine learning project planning systems. By dissecting these systems in detail, including their failures and limitations, we aim to distill transferable principles for the Community.\n\nThe workshop combines presentation-driven discussion in the morning with hands-on collabora- tive work in the afternoon. Morning sessions feature short project presentations interleaved with structured discussion. The afternoon hackathon lets participants turn ideas into code or concrete plans in small groups. Participants are encouraged to bring their own experiences, tools, and “war stories” to contribute throughout.\n-\nPresentations — Architecture walk-through, live demo, design trade-offs, and lessons learned & Discussion (2 h)\n-\nTime Session\n~30 min Introduction to Agentic AI for Science\n~30 min Project Presentation: Abstracts Explorer\n~30 min Project Presentation: Voucher Canvas Agent\n~15 min Break\n~15 min General Discussion —Cross-cutting themes: What design patterns were transferred? Where did things break? What would we do differently?\n-\nHackathon (2 h)\nParticipants form small groups to build or plan agentic ML tools hands-on. Groups self-organize around topics proposed in the morning discussion. Possible tracks include:\n• Feature sprints —Implement a concrete feature in one of the presented projects (e.g., a new MCP tool, a conference downloader plugin, an improved clustering pipeline). Bring a laptop and be ready to write code.\n• Project planning —Sketch the architecture and roadmap for a new agentic tool addressing a participant’s own research domain. Output: a brief design document or project proposal.\n• Integration experiments —Wire up existing tools via MCP or other protocols and stress-test them with real scientific queries. Time Session ~15 min Group Formation —Pitch topics, form groups (3–5 people). ~90 min Hands-on Work —Small-group hacking / planning with roaming support from organizers. ~15 min Lightning Reports —Each group presents what they built, learned, or planned (~2 min each).\n-\nExample Projects\n\n\nAbstracts Explorer — An open-source Python toolkit combining LLM-based semantic search, unsupervised clustering, and RAG to explore conference proceedings (NeurIPS, ICLR, ICML) at scale. Features native MCP integration for LLM-driven topic and trend analysis, a Flask web interface, and Docker-based deployment. (github.com/thawn/abstracts-explorer)\n\n\nVoucher Canvas Agent — an AI-powered exploration agent designed to help users plan machine learning projects. It uses a multi-container architecture with a LangGraph agent and a custom exploration server\n\n\n-\nGoals:\n• Architecture & Integration: How should agentic tools expose capabilities to LLMs (e.g., via the Model Context Protocol)? What are effective patterns for combining retrieval, structured analysis, and generation?\n• Evaluation & Trust: How do we measure whether an agentic science tool is actually helpful?\nWhen should we trust LLM-generated analyses of scientific literature?\n• Practical Pitfalls: What are common failure modes —hallucinated references, embedding model drift, brittle tool-use chains —and how can they be mitigated?\n• Human-in-the-Loop: What level of autonomy is appropriate? Where must human oversight remain, and how do we design interfaces that support it?\n-\nPresenters Experience:\nTill Korten is a Helmholtz AI Consultant and developer of Abstracts Explorer.\nHaider Khan is a Helmholtz AI Consultant and developer of Voucher Canvas Agent.\n-\nTarget Audience:\nResearchers, research software engineers, and data scientists who are building, evaluating, or considering agentic ML tools in their scientific workflows. No prior experience with agent frameworks is required — practical curiosity and a willingness to share experiences are the main prerequisites. For the afternoon hackathon, a laptop with Python 3.11+ and Git installed is recommended.\n-\nKeywords:\nAgentic AI; scientific workflows; human–AI interaction; evaluation and benchmarking; research infrastructure; AI governance

Brief Description and Outline:\nSimulation-Based Inference (SBI) is rapidly emerging as a powerful paradigm for scientific discovery, enabling parameter estimation, uncertainty quantification, and model selection in complex systems. However, significant challenges remain in translating SBI’s potential into widespread practical application. \nThis workshop, “Open Challenges in Simulation-Based Inference,” will bring together domain scientists actively employing SBI and developers of core SBI methodologies to identify and address these critical bottlenecks. \nWe aim to foster a collaborative environment for knowledge exchange, promoting mutual understanding of needs, benefits, and limitations across diverse scientific fields. While SBI methods have advanced rapidly, their application often faces hurdles related to computational cost, robustness to noise, scalability to high-dimensional parameter spaces, and effective integration into existing scientific workflows. Many domain scientists encounter specific challenges unique to their fields that are not adequately addressed by current methodological developments. \nThis workshop directly addresses the need for a dedicated space to bridge this gap, facilitating a dialogue between end-users and method developers. Addressing these challenges is crucial for unlocking the full potential of SBI and accelerating scientific progress across disciplines. \nRegistered participants (on ConfTool) can also submit an abstract for our poster session (deadline April 29th EOD):\nhttps://events.hifis.net/event/3955/abstracts/\nThis workshop aligns with the HAICON’s focus on scientific interdisciplinary research using machine learning.\nThis workshop will employ a highly interactive format designed to maximize engagement and knowledge sharing:\n\n\n[60 min] Keynote Address: A leading expert will present a comprehensive overview of the current state of SBI methods, outlining recent advancements and highlighting open research questions. (We are actively pursuing [Keynote Speaker Name/Area of Expertise] as a potential speaker.)\n\n\n[30 min] Lightning Talks: Selected poster presenters will deliver concise (3-5 minute) lightning talks to introduce their challenges and research to the broader audience.\n\n\n[120 min] Poster Session & Challenge Forum: The core of the workshop, dedicated to in-depth discussion and exploration of submitted posters. We will utilize structured community interviews during the poster session to systematically identify common challenges and potential solutions.\n\n\n[30 min] Wrap-up & Summary: A facilitated discussion will synthesize key findings from the poster session and lightning talks, leading to a concise summary of identified challenges and potential research directions.\n\n\n-\nGoals:\nSimulation-based inference (SBI) has become an essential tool across a wide and rapidly growing range of scientific domains — from biology and astronomy to geology and physics. Yet despite this breadth, practitioners across these fields frequently encounter the same fundamental challenges: questions about scalability, model misspecification, validation, and the gap between methodological advances and practical applicability. These challenges are rarely addressed in a cross-domain setting, leaving communities to rediscover solutions in isolation. This workshop aims to change that. By bringing together SBI method developers, domain scientists, and practitioners under one roof, we seek to create a dedicated space for structured exchange that is rarely possible within the boundaries of a single-domain conference. Our concrete objectives are to foster the adoption and continued development of SBI by connecting those who build these methods with those who apply them; to sharpen the community\'s collective understanding of the limitations of SBI and the mitigations available; and to provide an accessible, representative overview of the field that serves both newcomers and experienced practitioners.\n-\nPresenters Experience:\nThe workshop is organized by:\n\n\nDr Alina Bazarova, Helmholtz AI consultant at Forschungszentrum Jülich and a subgroup leader\n\n\nPeter Steinbach, head of AI consultants in Helmholtz Zentrum Dresden Rossendorf\n\n\n\nGiuseppe Viterbo, IMPRS PhD Candidate, University of Heidelberg\n\nWe hope to attract a seasoned sbi scientist to deliver an in-depth keynote to kick-off our workshop.\n- \nTarget Audience:\nSBI users from any scientific domain and SBI method developers\n- \nKeywords:\nInverse Problems, Generative AI, Simulation-Based Inference, Community, Machine Learning

Simulation-Based Inference (SBI) is rapidly emerging as a powerful paradigm for scientific discovery, enabling parameter estimation, uncertainty quantification, and model selection in complex systems. However, significant challenges remain in translating SBI’s potential into widespread practical application. \nThis workshop, “Open Challenges in Simulation-Based Inference,” will bring together domain scientists actively employing SBI and developers of core SBI methodologies to identify and address these critical bottlenecks. \nWe aim to foster a collaborative environment for knowledge exchange, promoting mutual understanding of needs, benefits, and limitations across diverse scientific fields. While SBI methods have advanced rapidly, their application often faces hurdles related to computational cost, robustness to noise, scalability to high-dimensional parameter spaces, and effective integration into existing scientific workflows. Many domain scientists encounter specific challenges unique to their fields that are not adequately addressed by current methodological developments. \nThis workshop directly addresses the need for a dedicated space to bridge this gap, facilitating a dialogue between end-users and method developers. Addressing these challenges is crucial for unlocking the full potential of SBI and accelerating scientific progress across disciplines. \nRegistered participants (on ConfTool) can also submit an abstract for our poster session (deadline April 29th EOD):\nhttps://events.hifis.net/event/3955/abstracts/\nThis workshop aligns with the HAICON’s focus on scientific interdisciplinary research using machine learning.\nThis workshop will employ a highly interactive format designed to maximize engagement and knowledge sharing:\n\n\n[60 min] Keynote Address: A leading expert will present a comprehensive overview of the current state of SBI methods, outlining recent advancements and highlighting open research questions. (We are actively pursuing [Keynote Speaker Name/Area of Expertise] as a potential speaker.)\n\n\n[30 min] Lightning Talks: Selected poster presenters will deliver concise (3-5 minute) lightning talks to introduce their challenges and research to the broader audience.\n\n\n[120 min] Poster Session & Challenge Forum: The core of the workshop, dedicated to in-depth discussion and exploration of submitted posters. We will utilize structured community interviews during the poster session to systematically identify common challenges and potential solutions.\n\n\n[30 min] Wrap-up & Summary: A facilitated discussion will synthesize key findings from the poster session and lightning talks, leading to a concise summary of identified challenges and potential research directions.\n\n\n-\nGoals:\nSimulation-based inference (SBI) has become an essential tool across a wide and rapidly growing range of scientific domains — from biology and astronomy to geology and physics. Yet despite this breadth, practitioners across these fields frequently encounter the same fundamental challenges: questions about scalability, model misspecification, validation, and the gap between methodological advances and practical applicability. These challenges are rarely addressed in a cross-domain setting, leaving communities to rediscover solutions in isolation. This workshop aims to change that. By bringing together SBI method developers, domain scientists, and practitioners under one roof, we seek to create a dedicated space for structured exchange that is rarely possible within the boundaries of a single-domain conference. Our concrete objectives are to foster the adoption and continued development of SBI by connecting those who build these methods with those who apply them; to sharpen the community\'s collective understanding of the limitations of SBI and the mitigations available; and to provide an accessible, representative overview of the field that serves both newcomers and experienced practitioners.\n-\nPresenters Experience:\nThe workshop is organized by:\n\n\nDr Alina Bazarova, Helmholtz AI consultant at Forschungszentrum Jülich and a subgroup leader\n\n\nPeter Steinbach, head of AI consultants in Helmholtz Zentrum Dresden Rossendorf\n\n\n\nGiuseppe Viterbo, IMPRS PhD Candidate, University of Heidelberg\n\nWe hope to attract a seasoned sbi scientist to deliver an in-depth keynote to kick-off our workshop.\n- \nTarget Audience:\nSBI users from any scientific domain and SBI method developers\n- \nKeywords:\nInverse Problems, Generative AI, Simulation-Based Inference, Community, Machine Learning

Brief Description and Outline:\nThis workshop will focus on standardized benchmarking and reproducible evaluation of multiomics integration methods in bioinformatics, using Multiverse as a unifying framework rather than as the sole focus. The session will begin with a 15-minute conceptual overview of multimodal integration challenges, model diversity, and current limitations in benchmarking practices.\nWe will then dedicate 15-minutes to a structured review of representative integration approaches, emphasizing their assumptions, strengths, and limitations in biological contexts.\nThe core of the workshop will consist of two hands-on blocks. The first 30-minute session will guide participants through running a standardized benchmarking workflow: selecting curated datasets, configuring preprocessing pipelines, training multiple integration models, performing hyperparameter optimization, and evaluating results using multiple integration metrics. The second 60-minute session will focus on extensibility and community benchmarking: integrating a custom model within a containerized setup and performing comparative analysis across methods.\nWe will explicitly demonstrate how containerization resolves dependency conflicts and ensures\nreproducibility across heterogeneous tools. Overall, the tutorial balances theoretical context, coverage of representative methodological families, and practical reproducible workflows, ensuring relevance beyond a single framework while providing concrete skills for multi-omics integration\nresearch.\n\nGoals:\nThe primary goal of this tutorial is to equip participants with a rigorous and practical understanding of how to benchmark multi-omics integration methods in a standardized and reproducible manner. Participants will learn how to design fair comparisons across heterogeneous models, select appropriate evaluation metrics aligned with biological objectives, and interpret performance trade-offs between modality alignment, clustering quality, and biological signal preservation.\nA second goal is to provide hands-on experience with reproducible computational workflows. Participants will gain practical skills in running containerized pipelines, managing model dependencies, performing systematic hyperparameter optimization, and executing end-to-end benchmarking experiments across multiple datasets and methods.\nA third goal is to foster extensibility and community-driven benchmarking. Attendees will learn how to integrate new datasets, add custom models, and implement additional evaluation metrics within a unified framework, enabling them to adapt benchmarking workflows to their own research questions. By the end of the tutorial, participants will be able to critically assess integration methods and build reproducible, extensible benchmarking pipelines for multi-omics data analysis.\n\nPresenters Experience:\nAnis Ismail is a second-year PhD researcher at the Laboratory of Multi-Omic Integrative Bioinformatics working on explainable multi-omic representation learning models. He is currently President of the ISCB Regional Student Group Belgium and has co-organized the Interuniversity Belgian Biohackathon (September 2025). He also leads the Data for Good Challenge at Emergent, designing and organizing Leuven’s biggest Data Science Challenge. He also has substantial teaching and outreach experience, including workshops on AI and data science at Beirut AI, Zaka, SE Factory, and Lebanese American University. He will be leading the workshop, and will focus on multi-omics representation learning concepts, benchmarking methodology and demonstrating containerized workflows.\nSaptarshi Chakrabarti is a data engineer at the Leuven Institute of Single Cell Omics where he develops scalable machine learning infrastructure for computational biology. He contributed to the MLOps and workflow automation components of the Multiverse framework, with expertise in reproducible research software engineering, pipeline deployment, and maintainable ML systems in academic environments. He has mentored undergraduate programming courses at KU Leuven and presented a technical demo at RSE Day 2025 at KU Leuven. His role in the workshop will be focusing on the practical implementation aspects of benchmarking multi-omics models and technical support of attendees.\n\nTarget Audience:\nThe target audience includes master’s students, PhD students, and researchers working in single-cell and multi-modal bioinformatics. Participants are expected to have a basic understanding of machine learning concepts and familiarity with biological data analysis. Practical experience with Python and computational workflows is recommended, as the tutorial will involve hands-on benchmarking, model training, and evaluation of multi-omics integration methods. Prior knowledge of deep learning is helpful but not strictly required, as core methodological ideas will be introduced during the session.\n\nKeywords:\nMulti-omics integration, Benchmarking, Deep learning, Containerization, Reproducibility, Model evaluation, Bioinformatics tools

Brief Description and Outline:\nJoint introduction (~12 min)\nElla Bahry and Deborah Schmidt will give a short overview of PixelPatrol, and Hans Werners and Philipp Heuser will introduce the Helmholtz Model Zoo. The two tools address opposite ends of the AI pipeline in scientific imaging: PixelPatrol helps you understand and curate your data before training, while the Helmholtz Model Zoo lets you share and run inference with the resulting trained models.\nTrack A: PixelPatrol\nPixelPatrol is an open-source quality control and data exploration tool for scientific image datasets. It systematically profiles imaging data and generates interactive dashboard reports with statistics and visualizations, making it a useful first step before any computationally intensive analysis or AI model training.\nParticipants will work with their own datasets or a provided example (aqQua foundation model training data). The session covers generating comprehensive quality reports; interpreting visualizations including file stats, metadata consistency, and image statistics; interactive filtering and grouping; identifying outliers, artifacts, and acquisition inconsistencies; and an introduction to developing custom PixelPatrol packages and plugins.\nTrack B: Helmholtz Model Zoo\nThe Helmholtz Model Zoo (HMZ) is a cloud-based platform for sharing and running inference on deep learning models across the Helmholtz Association. It is integrated with Helmholtz infrastructure, including Helmholtz ID, dCache, and DESY\'s HPC cluster with NVIDIA L40S GPUs. Users can run inference via both a web interface and a REST API, with support for datasets from gigabytes to terabytes. NVIDIA Triton Inference Server and Slurm manage GPU resources, and virtual organizations enable fine-grained access control across scientific domains.\nParticipants will go through the steps required to prepare and deploy a model on the HMZ, including an overview of available deployment templates. Participants are encouraged to bring their own model or an open-source model they would like to deploy (for example, from Hugging Face); an example model will also be provided. Those who want to prepare in advance are welcome to reach out at support@helmholtz-imaging.de.\n\nGoals:\nPixel Patrol: This hands-on workshop introduces participants to systematic data curation practices for scientific image datasets using PixelPatrol, an open-source tool for dataset exploration and quality control.\nParticipants will learn to identify data quality issues before computationally expensive analysis, explore dataset characteristics interactively, and implement quality control workflows that prevent common pitfalls in imaging pipelines and AI model training.\nWith PixelPatrol it will be possible to generate fingerprints of the training data, which can be used in the context of the HMZ to support users to identify trained models where the fingerprint of the training data suggests successful inference to the users data\nWe will also provide an outlook on how both Pixel Patrol and the Model Zool will be integrated in the near future.\n\nPresenters Experience:\nElla Bahry and Deborah Schmidt (Helmholtz Imaging, MDC Berlin) co-lead PixelPatrol development and both teach at HIDA.\nHans Werners (Helmholtz Imaging, DESY Hamburg) is the primary developer of the Helmholtz Model Zoo.\nPhilipp Heuser leads the Helmholtz Imaging Support and Engineering team at DESY and is deeply involved in HMZ development.\n \nTarget Audience:\nImaging researchers working with large or heterogeneous datasets, scientists preparing data for AI/deep learning workflows, data curators, research software engineers, and ML practitioners wanting to make their models accessible to the Helmholtz community. Participants should bring a laptop and, optionally, a dataset or model they would like to work with.\n \nKeywords:\ndata curation, image quality control, dataset exploration, deep learning workflows, data integrity, metadata validation, FAIR research, model deployment, Helmholtz Model Zoo, inference

Brief Description and Outline:\nThis tutorial aims to provide attendees with a foundational introduction to quantum computing and quantum machine learning (QML). It will address key challenges in QML, including data en- coding strategies, model trainability [1], the barren plateau phenomenon [2], optimization issues such as local minima [3], dequantization [4], benchmarking [5], current hardware limitations [6], etc. In addition, the tutorial will highlight recent and promising research directions in the field.\nThe tutorial will conclude with a project demonstrating a simple quantum application using Py- Torch/TensorFlow.\n-\n1 Thanasilp, S., Wang, S., Nghiem, N.A. et al. Subtleties in the trainability of quantum machine learning models. Quantum Mach. Intell. 5, 21 (2023). https://doi.org/10.1007/s42484-023-00103-6\n2 M. Larocca et al., “Barren Plateaus in Variational Quantum Computing,”Nat Rev Phys, vol. 7, no. 4, pp. 174–189, Mar. 2025, doi: 10.1038 s42254-025-00813-9.\n3 X. You and X. Wu, “Exponentially Many Local Minima in Quantum Neural Networks,”Oct. 06, 2021, arXiv: arXiv:2110.02479. doi: 10.48550/arXiv.2110.02479.\n4 R. Sweke et al., “Potential and limitations of random Fourier features for dequantizing quantum machine learning,”Quantum, vol. 9, p. 1640, Feb. 2025, doi: 10.22331/q-2025-02-20-\n1640.\n5 J. Bowles, S. Ahmed, and M. Schuld, “Better than classical? The subtle art of benchmarking quantum machine learning models,”Mar. 14, 2024, arXiv: arXiv:2403.07059. Accessed: Apr. 08, 2024. [Online]. Available: http://arxiv.org/abs/2403.07059 6 J. Preskill, “Quantum Computing in the NISQ era and beyond,”Quantum, vol. 2, p. 79, Aug. 2018, doi: 10.22331/q-2018-08-06-79.\n-\nGoals:\nThe goal of the tutorial is to give participants an overview of current challenges, limitations and possibilities in the field of QML. After the tutorial, participants should be comfortable assessing current QML publications and should be able to deduce themselves which projects might benefit from integrating quantum computers in the short-, mid-, and long-term.\nQuantum computers provide support for complex problems and can potentially offer exponentially faster solutions for some use cases. Although current quantum computers are still limited in their applicability in real-world projects, the technological and algorithmical outlook offer promising options. Furthermore,  awareness and knowledge needs to be built to enable good integration and use of quantum computers along their improving capabilities.\n-\nPresenters Experience:\nEileen Kuhn is a team leader on quantum machine learning at KIT. She received her PhD in computer science from the Karlsruhe Institute of Technology in 2017. She is used to working in close collaboration with diverse domains including for example High Energy Physics, Climatology, or Muselogy. Her research activities focus on trainability and efficiency of QML models, hybrid models and workflows, applied quantum computing, and (quantum) research software engineering. She has particular experience in training, teaching and supervision.\nSince 2009 she has regular teaching experience and since 2022 she is teaching QML and Quantum Computing at KIT. She also has experience in giving tutorials for various audiences including primary pupils, pupils and participants at bachelors/masters/doctoral degrees at various occasions including summer schools.\nGabriel Ruiz is a second-year PhD researcher at KIT focusing on the trainability of quantum machine learning models and the design of quantum algorithms for practical applications. Prior to his doctoral studies, he worked as a research assistant in numerical simulations of fluid dynamics and wave propagation. Since 2019, he has been involved in teaching algorithms, discrete mathematics, programming, parallel computing, and linear algebra. Since 2025, he has co-taught with Eileen the QML course at KIT.\n-\nTarget Audience:\nKnowledge of linear algebra is recommended.\n-\nKeywords:\nQuantum Computing, Quantum Algorithms, Quantum Machine Learning

Brief Description and Outline:\nThis is a hands-on tutorial to create a hybrid quantum-classical QML workflow for training a Quantum Neural Network. We will first introduce different strategies on how to include quantum computers into machine learning workflows. Further, we will not only show how an optimization/training process looks like, but will also discuss different data encoding options and how to choose an appropriate QML model. Furthermore, we will look into how to interpret the output of quantum computers.\nThe project itself will be realized with Pennylane, a differentiable programming framework that is optimized for integration in high performance clusters. We will further interface with common tools such as PyTorch and TensorFlow.\n\nGoals:\nGoal of the tutorial is to lower the barriers for starting to develop and research hybrid quantum-classical workflows. As a current and upcoming technology with the potential to revolutionize machine learning, it is critical for researchers to understand both the limitations and well as potential of this field.\n\nPresenters Experience:\nEileen Kuhn is a team leader on quantum machine learning at KIT. She received her PhD in computer science from the Karlsruhe Institute of Technology in 2017. She is used to working in close collaboration with diverse domains including for example High Energy Physics, Climatology, or Museology. Her research activities focus on trainability and efficiency of QML models, hybrid models and workflows, applied quantum computing, and (quantum) research software engineering. She has particular experience in training, teaching and supervision.\nSince 2009 she has regular teaching experience and since 2022 she is teaching QML and Quantum Computing at KIT. She also has experience in giving tutorials for various audiences including primary pupils, pupils and participants at bachelors/masters/doctoral degrees at various occasions including summer schools.\nGabriel Ruiz is a second-year PhD researcher at KIT, focusing on the trainability of quantum machine learning models and the design of quantum algorithms for practical applications. Prior to his doctoral studies, he worked as a research assistant in numerical simulations of fluid dynamics and wave propagation. Since 2019, he has been involved in teaching algorithms, discrete mathematics, programming, parallel computing, and linear algebra. Since 2025, he has co-taught with Eileen the QML course at KIT.\n\nTarget Audience:\nParticipants should have experience in programming in python. We expect participants to either have a basic understanding of quantum computing already or to participate in the first part of the tutorial.\nEach participant should bring a laptop.\n\nKeywords:\nQuantum Machine Learning, Quantum Variational Algorithms, Quantum Neural Networks

Brief Description and Outline:\nThis 4-hour hands-on tutorial introduces practical methods for building Generative AI (Gen AI) applications using prompt engineering and Retrieval-Augmented Generation (RAG). The session moves from controlling large language model (LLM) behavior to grounding models in external documents and assembling a simple AI system using Python.\n-\nThe tutorial includes three modules:\n● Module 1 (60 minutes): Prompt Engineering Fundamentals.\n● Module 2 (90 minutes): Retrieval-Augmented Generation (RAG) and LangChain.\n● Module 3 (90 minutes): Assembling a Simple AI Application.\n-\nGoals:\nThis tutorial equips participants with practical skills for building reliable AI-powered applications.\nBy the end of the session, participants will be able to:\n● Design structured and controllable prompts\n● Explain and implement a basic RAG pipeline\n● Connect language models to external documents\n● Assemble a minimal AI application using Python\n–> The tutorial supports HAICON26’s focus on applied AI by bridging theoretical understanding and real-world implementation. As Gen AI becomes central to research workflows, grounding models in real data and ensuring output reliability is increasingly important.\n-\nPresenters Experience:\nDr. Osama Hamed is a postdoctoral researcher specializing in applied artificial intelligence and machine learning. He has several years of experience working in the fields of natural language processing (NLP) and AI. Dr. Hamed earned his PhD in NLP from the University of Duisburg–Essen in 2019. Additionally, he accumulated several years of teaching experience as both an assistant professor and while holding a master’s degree.\nEng. Fatima Rajab is a Computer Systems Engineering graduate and QA Automation Engineer with a strong focus on Gen AI and RAG. She has experience integrating AI techniques into software testing workflows to improve reliability and traceability.\n-\nTarget Audience:\nResearchers, developers, data scientists, and graduate students interested in building practical GenAI applications.\nFamiliarity with Python programming is required. No prior experience with LLMs, RAG or LangChain is necessary. The tutorial begins with foundational concepts.\nParticipants should bring their own laptops.\n-\nKeywords:\nPrompt Engineering, LLMs, RAG, LangChain, Gen AI

Brief Description and Outline:\nThis 4-hour hands-on tutorial introduces practical methods for building Generative AI (Gen AI) applications using prompt engineering and Retrieval-Augmented Generation (RAG). The session moves from controlling large language model (LLM) behavior to grounding models in external documents and assembling a simple AI system using Python.\nThe tutorial includes three modules:\n● Module 1 (60 minutes): Prompt Engineering Fundamentals.\n● Module 2 (90 minutes): Retrieval-Augmented Generation (RAG) and LangChain.\n● Module 3 (90 minutes): Assembling a Simple AI Application.\n-\nGoals:\nThis tutorial equips participants with practical skills for building reliable AI-powered applications.\nBy the end of the session, participants will be able to:\n● Design structured and controllable prompts\n● Explain and implement a basic RAG pipeline\n● Connect language models to external documents\n● Assemble a minimal AI application using Python\n–> The tutorial supports HAICON26’s focus on applied AI by bridging theoretical understanding and real-world implementation. As Gen AI becomes central to research workflows, grounding models in real data and ensuring output reliability is increasingly important.\n-\nPresenters Experience:\nDr. Osama Hamed is a postdoctoral researcher specializing in applied artificial intelligence and machine learning. He has several years of experience working in the fields of natural language processing (NLP) and AI. Dr. Hamed earned his PhD in NLP from the University of Duisburg–Essen in 2019. Additionally, he accumulated several years of teaching experience as both an assistant professor and while holding a master’s degree.\nEng. Fatima Rajab is a Computer Systems Engineering graduate and QA Automation Engineer with a strong focus on Gen AI and RAG. She has experience integrating AI techniques into software testing workflows to improve reliability and traceability.\n-\nTarget Audience:\nResearchers, developers, data scientists, and graduate students interested in building practical GenAI applications.\nFamiliarity with Python programming is required. No prior experience with LLMs, RAG or LangChain is necessary. The tutorial begins with foundational concepts.\nParticipants should bring their own laptops.\n-\nKeywords:\nPrompt Engineering, LLMs, RAG, LangChain, Gen AI

Brief Description and Outline:\nFoundation models can be adapted in many ways, but it’s often unclear how these approaches differ beyond their descriptions. In this interactive workshop, we start from a fixed pretrained check-point and walk through several prominent adaptation strategies (linear probing, LoRA, prompt-based methods, and related techniques). Using a shared base experiment, participants will run hands-on notebooks that visualize parameter updates, representational shifts, and performance changes across methods.\n-\nGoals:\nBuild practical intuition for major adaptation strategies\n-\nPresenters Experience:\nDasha Trofimova is a ML engineer at DKFZ with over 10 years of experience applying DL methods to biomedical and scientific problems. She did the workshop last year at HAICON and has teaching experience.\nJeremiasAI Traub is a Research Engineer at DKFZ, Helmholtz Imaging\n-\nTarget Audience:\nresearchers with data analytics experience\n-\nKeywords:\nModel Adaptation, Parameter-Efficient Fine-Tuning, Foundation Model

Brief Description and Outline:\nThis tutorial sheds the light on the small and locally deployed large language models (LLMs) and vision language models (VLMs). It starts by deploying them on our local machines with the help of Ollama. This is followed by the implementation of API endpoints – using Python – that utilize the VLM for image captioning via zero-shot prompting. Finally, the tutorial concludes with an evaluation of the AI-generated captions using the dedicated metrics, such as BLEU and ROUGE\n\nGoals:\nFamiliarize the audience with the importance of locally deployed LLM or VLM models, especially for implementing services that don’t require API keys or the disclosure of organizational data outside its network\n\nPresenters Experience:\nDr. Osama Hamed is a postdoctoral researcher specializing in applied artificial intelligence and machine learning. He has several years of experience working in the fields of natural language processing (NLP) and AI. Dr. Hamed earned his PhD in NLP from the University of Duisburg–Essen in 2019. Additionally, he accumulated several years of teaching experience as both an assistant professor and while holding a master’s degree.\n\nTarget Audience:\nResearchers who are interested in local LLMs, Computer Vision or NLP.\nBasic to good understanding of Gen AI, LLMs as well as text generation evaluation metrics.\nParticipants should bring their own laptops.\n\nKeywords:\nNatural Language Processing, Computer Vision, Image Captioning, LLMs, Vision-Language Models (VLMs).

Brief Description and Outline:\nDesigning GDPR-compliant projects is challenging for many researchers, resulting in long approval cycles, repeated changes to the project design and suboptimal technical setups. This hands-on workshop introduces a six-step process for translating research concepts into lean, GDPR-compliant project designs. Participants will then practice by applying the process on three cases involving healthcare data. The workshop is suitable for all participants working with personal data in the EU. No GDPR expertise is required, basic data protection concepts (e.g. pseudonymization vs anonymization) should be familiar.\nAgenda:\n(1) Introduction, Objectives, Agenda, Expectations (10 min): We introduce the structure of the session and gather input from participants on common GDPR-related challenges in their projects.\n(2) GDPR-compliant design in six steps (20 min): Participants are introduced to a six-step process which takes them from a research concept to a GDPR-compliant project design. The six steps cover GDPR scope assessment, data classification, processing options, roles and responsibilities, international data transfers, and safeguards. In preparation for the group exercises, the three cases and supporting material for the case practice are presented. The cases are based on real-world projects and involve common challenges encountered in research projects, e.g. combining personal data from several countries.\n(3) Group Case Exercises (40 min): In small groups of two to six, participants apply the six-step process to one of the three cases. They analyze requirements, explore data processing options, and develop GDPR-compliant project designs.\n(4) Group Presentations & Discussion (40 min): Each group presents its solution, key decisions, and underlying reasoning. Common patterns, trade-offs of different approaches and open questions are discussed.\n(5) Wrap-Up & Takeaways (10 min): The workshop concludes with a summary of key insights, pitfalls, and practical guidance. Participants will get a one-pager hand-out, which describes the six steps and provides practical examples how to apply them.\n\nGoals:\nBy the end of the workshop, participants will be able to:\n(1) Apply a six-step process to design GDPR-compliant research projects,\n(2) Evaluate GDPR-related options early in the project,\n(3) Discuss GDPR-related issues confidently with data protection officers or IT specialists. Designing projects in a GDPR-compliant way from the outset helps avoid delays, unnecessary iterations, and compliance risks, thereby speeding up research projects.\n\nPresenters Experience:\nJulian Beimes is Associate Principal at idalab GmbH, where he supports medtech and biopharma companies in conceptualizing and implementing AI solutions.\n\nTarget Audience:\nThe workshop is aimed at scientists and research professionals who design, lead, or contribute to data-driven projects involving personal data. No GDPR expertise is required, basic data protection concepts (e.g. pseudonymization vs anonymization) should be familiar.\n\nKeywords:\nGDPR, Project Design, Data Processing

Brief Description and Outline:\nThe goals are here dual:\n1) Inform and engage on AI developments and initiatives at the EU level (e.g RAISE), from a strategical point of view:\n\n\nWhy were such developments/ initatives chosen?\n\n\nHow did they materialize?\n\n\nHow can we, scientists, influence this process?\n\n\nAre we satisfied with those?\n\n\nDo you think these calls bring AI forward?\n\n\n-\n2) Enhance one’s career / skills: \n\n\nWhat could science policy bring to my current or future work (e.g CoARA)?\n\n\nHow do people weigh in policy-making in Brussels or with other collaborators\n\n\nRetrospective on the Helmholtz Brussel\'s habilitation opportunity to engage temporarily on science policy in their office, for anyone within the Helmholtz Association.\n\n\n-\nThis session will be divided in two short presentations (~20min) intertwined with interactive / role-play aspects and concludes into a general, open discussion.\n-\nGoals:\nBroaden scientists\' perspectives regarding:\n1) AI strategic development within EU and\n2) what science policy looks like.\n-\nPresenters Experience:\nIoannis Legouras is the Head of the Helmholtz Brussels Office, long term experience on career perspective talks.\n-\nTarget Audience:\nYoung scientists with an interest to understand science from another angle.

Brief Description and Outline:\nThis workshop introduces TwinWeaver, a generative artificial intelligence framework for building digital twins of longitudinal systems. Participants will learn how generative models can move beyond static prediction toward dynamic trajectory simulation, counterfactual reasoning, and intervention modelling. The session combines conceptual foundations with practical design principles and a guided walkthrough of the TwinWeaver pipeline.\n-\nHour 1 Conceptual Foundations\n\nIntroduction to digital twins and longitudinal modelling\nLimitations of discriminative prediction models\nPrinciples of generative modelling for temporal systems\nRepresentation of trajectories as structured token sequences\nEthical, governance, and reproducibility considerations\n\nHour 2 TwinWeaver Framework\n\nArchitecture of the TwinWeaver generative model\nTokenisation of longitudinal events and state transitions\nTraining pipeline and data preprocessing\nConditioning and intervention modelling\nEvaluation of trajectory forecasts and simulation realism\n\nHour 3 Practical Walkthrough and Discussion\n\nDemonstration of converting longitudinal data into TwinWeaver format\nOverview of model training workflow\nGeneration of simulated future trajectories\nCounterfactual scenario modelling\nDiscussion of scaling, infrastructure requirements, and deployment strategies\nInteractive discussion on adapting TwinWeaver to participant use cases.\n\n-\nGoals:\nParticipants will understand the conceptual foundations of digital twins and how they differ from traditional predictive models. They will learn how generative transformer architectures can represent structured temporal data. They will gain insight into how to construct longitudinal token representations, train generative models, and simulate future trajectories. They will leave with a clear roadmap for implementing digital twins in their own research domain. Participants will understand how to design and implement a generative digital twin framework for longitudinal modelling. They will gain practical insight into training and evaluating generative trajectory models and will be equipped to initiate digital twin projects within their own research or applied context.\n-\nPresenters Experience:\nThis workshop has already been tested in a hackathon setting in Melbourne and internally at Roche.\nThe underlying paper has led to invitations to present at AstraZeneca, attracted industry funding, and contributed to the launch of a women’s health start up supported by venture capital. We have published the code under an Apache 2.0 licence, making it freely available for use in both academic and commercial settings.\n\nNikita Makarov did his PhD on generative AI and digital twins for clinical trials & real-world data, in a collaboration between Roche, Helmholtz Munich, and the Ludwig Maximilian University of Munich. He is the lead author on the TwinWeaver project.  He now continues to advance this research at Valinor Discovery as senior scientist, working on how multi-modality can enhance predictive modeling in healthcare.\n\n\n\nDaksh Pamar and Alina Ameth are respectively PhD student and PhD graduate in the group of Michael Menden, within the Faculty of Medicine, Dentistry and Health Sciences of the University of Melbourne.\n\n-\nTarget Audience:\nThis workshop is designed for researchers and practitioners working with longitudinal data, including computational scientists, data scientists, clinicians, biomedical researchers, and doctoral candidates. Basic familiarity with machine learning concepts is recommended, but prior experience with generative modelling is not required.\nParticipants should bring a laptop capable of running Python notebooks. A prepared example dataset and code templates will be provided.\n-\nKeywords:\nDigital Twins

Brief Description and Outline:\nThis workshop introduces TwinWeaver, a generative artificial intelligence framework for building digital twins of longitudinal systems. Participants will learn how generative models can move beyond static prediction toward dynamic trajectory simulation, counterfactual reasoning, and intervention modelling. The session combines conceptual foundations with practical design principles and a guided walkthrough of the TwinWeaver pipeline.\n-\nHour 1 Conceptual Foundations\n\nIntroduction to digital twins and longitudinal modelling\nLimitations of discriminative prediction models\nPrinciples of generative modelling for temporal systems\nRepresentation of trajectories as structured token sequences\nEthical, governance, and reproducibility considerations\n\nHour 2 TwinWeaver Framework\n\nArchitecture of the TwinWeaver generative model\nTokenisation of longitudinal events and state transitions\nTraining pipeline and data preprocessing\nConditioning and intervention modelling\nEvaluation of trajectory forecasts and simulation realism\n\nHour 3 Practical Walkthrough and Discussion\n\nDemonstration of converting longitudinal data into TwinWeaver format\nOverview of model training workflow\nGeneration of simulated future trajectories\nCounterfactual scenario modelling\nDiscussion of scaling, infrastructure requirements, and deployment strategies\nInteractive discussion on adapting TwinWeaver to participant use cases.\n\n-\nGoals:\nParticipants will understand the conceptual foundations of digital twins and how they differ from traditional predictive models. They will learn how generative transformer architectures can represent structured temporal data. They will gain insight into how to construct longitudinal token representations, train generative models, and simulate future trajectories. They will leave with a clear roadmap for implementing digital twins in their own research domain. Participants will understand how to design and implement a generative digital twin framework for longitudinal modelling. They will gain practical insight into training and evaluating generative trajectory models and will be equipped to initiate digital twin projects within their own research or applied context.\n-\nPresenters Experience:\nThis workshop has already been tested in a hackathon setting in Melbourne and internally at Roche.\nThe underlying paper has led to invitations to present at AstraZeneca, attracted industry funding, and contributed to the launch of a women’s health start up supported by venture capital. We have published the code under an Apache 2.0 licence, making it freely available for use in both academic and commercial settings.\n\nNikita Makarov did his PhD on generative AI and digital twins for clinical trials & real-world data, in a collaboration between Roche, Helmholtz Munich, and the Ludwig Maximilian University of Munich. He is the lead author on the TwinWeaver project.  He now continues to advance this research at Valinor Discovery as senior scientist, working on how multi-modality can enhance predictive modeling in healthcare.\n\nDaksh Pamar and Alina Ameth are respectively PhD student and PhD graduate in the group of Michael Menden, within the Faculty of Medicine, Dentistry and Health Sciences of the University of Melbourne.\n\n-\nTarget Audience:\nThis workshop is designed for researchers and practitioners working with longitudinal data, including computational scientists, data scientists, clinicians, biomedical researchers, and doctoral candidates. Basic familiarity with machine learning concepts is recommended, but prior experience with generative modelling is not required.\nParticipants should bring a laptop capable of running Python notebooks. A prepared example dataset and code templates will be provided.\n-\nKeywords:\nDigital Twins

Brief Description and Outline:\nThis workshop introduces the HAICORE HPC Cluster at HZDR, a high-performance computing resource available to Helmholtz AI projects. This session focuses on how researchers can effectively utilize our Open OnDemand web portal to run AI and data-intensive workloads and our comprehensive web application suite for Machine Learning workflows.\nWe highlight how HAICORE integrates established HPC technologies with modern web services and AI oriented tooling, to enable reproducible and scalable research across Helmholtz centers.\n\nWorkshop Outline (2 hours)\n- Introduction to HAICORE at HZDR (10 min) Overview of HAICORE resource allocation portal (ColdFront) Allocation process, account provisioning, and available resources.\n- Getting Started: HAICORE Web Portal (20 min) Accessing HPC resources through Web-portal (Open OnDemand). Overview of available application suite for Coding, MLOps, Data Transfer etc. Introduction to HAICORE Cluster Software stack.\n- Running Machine Learning Workloads (40 min) Training Scientific Deep Learning models. Using MLOps tools for tracking and reproducibility.\n- GenAI Workflows on HAICORE (40 min) Deploying Language Models on the Cluster. Using MLFlow for Generative AI workflows.\n- Q&A and Discussion (10 min)\n\nGoals:\nWorkshop Goals\n\nIntroduce Helmholtz AI researchers to brand new HAICORE HPC resources at HZDR.\nDemonstrate the utility of HZDR HPC-labs web portal and its comprehensive application suite.\nEncourage reproducible and scalable AI workflows by leveraging best MLOps practices.\n\nImportance for HAICON26:\nAs Helmholtz AI initiatives increasingly require scalable GPU resources, reproducible workflows, and support for emerging areas such as Generative AI, there is a strong need to make these capa-bilities accessible and practical for researchers.\nThis workshop directly addresses that need by guiding participants through the full lifecycle of working on HAICORE —from allocation via ColdFront, to interactive access through Open On-Demand, to executing machine learning and GenAI workflows with MLOps best practices. The workshop supports the broader HAICON 2026 mission of enabling scalable and reproducible AI research across Helmholtz centers.\n\nPresenters Experience:\nVarun Sudharshanam:\nScientific Staff at HPC‑Labs, HZDR, focusing on HPC infrastructure and machine‑learning workflows on HPC clusters.\nPrior Teaching Experience:\n\nConducted multiple High Performance Computing Workshops at HZDR and UFZ.\nInstructor for Multi‑GPU Workloads at the HPC‑Gateway Fall School 2025, HZDR.\nHIDA Instructor for Software Carpentries.\n\n\nDr. Kushal Ramakrishna:\nResearch Scientist at HPC‑Labs, HZDR, advancing ab initio electronic‑structure simulations and\nmultiscale materials modelling through data‑driven and machine‑learning approaches for materials science.\nPrior Teaching Experience:\n\nInstructor for the lab component of the “Highly Parallel Programming of GPUs” course at the Faculty of Computer Science, TU Dresden.\nConducted multiple High Performance Computing Workshops at HZDR and UFZ.\n\n\nTarget Audience:\nThis workshop is aimed at researchers, PhD students, postdocs, and research software engineers\nfrom Helmholtz centers who:\n\nDevelop or apply machine learning / AI methods in scientific research\nRequire scalable GPU resources for training or deploying models\nAre interested in reproducible MLOps or LLM-Ops workflows\nWant to transition from local experimentation to production-grade HPC environments.\n\n\nParticipants are expected to have:\n\nBasic familiarity with Python and machine learning concepts\nSome experience working in Linux environments (helpful but not mandatory)\nNo prior experience with HAICORE or HPC systems is required\n\nThe session is designed to be accessible to both AI practitioners new to HPC and experienced HPC users looking to adopt modern AI and MLOps workflows.\n\nKeywords:\nHAICORE, HPC, Scientific ML, MLOps, AI Workflows

Brief Description and Outline:\nThis workshop addresses how ethical, fairness-related, and sustainability considerations can be systematically embedded into industrial AI projects. While industrial AI systems increasingly rely on predictive models and uncertainty-aware decision support, ethical aspects are often addressed late, implicitly, or inconsistently. The workshop focuses on practical, method-oriented approaches that help integrate responsibility into AI projects without slowing down innovation. \nRather than presenting specific software or algorithms, the session emphasizes transferable concepts such as uncertainty-aware decision-making, data governance, and Corporate Digital Responsibility (CDR). Participants will work with representative industrial AI scenarios involving tabular and time-series data, such as predictive models in production contexts. The workshop treats Responsible AI as a set of practices and methods that shape how predictive models are designed, deployed, and embedded in organizational contexts.\nOutline (2 hours):\n\n\nIntroduction & framing (20 min) - Responsible AI challenges in industrial decision-making\n\n\nImpulse: Ethics, uncertainty, and governance (20 min) - Key concepts and methodological perspectives\n\n\nGroup work: Case-based analysis (40 min) - Identification of ethical risks, uncertainties, and responsibility gaps\n\n\nPlenary discussion (20 min) - Comparing approaches and trade-offs\n\n\nSynthesis & takeaways (20 min) - Translating ethical reflection into project practice\n\n\n-\nGoals:\nThe goal of this workshop is to provide participants with practical orientation on how to integrate ethics, fairness, and sustainability into industrial AI projects. The session aims to demonstrate that ethical reflection can function as a structuring element for clarity, robustness, and trust in data-driven decision-making. The workshop contributes to HAICON26 by bridging conceptual frameworks with real-world AI practice and fostering dialogue between researchers and industry practitioners. Responsible AI is conceptualized not as a checklist, but as a structured approach to shaping practices and methods in industrial AI projects.\n-\nPresenters Experience:\nDr. Christopher Koska, Senior Researcher and coordinator of the work package “Ethics” within the transfer project VoMoPro at the Centre for Future Production, University of Augsburg.\nHis work focuses on Responsible AI as an application-oriented framework for shaping ethical and organizational practices in socio-technical contexts and supporting knowledge transfer between research and practice. He holds a PhD in philosophy with a monograph on the ethics of algorithms and has over 15 years of experience in designing and facilitating workshops and training formats on Corporate Digital Responsibility, with a focus on data and algorithmic ethics, in both academic and industry settings.\nMarkus Schatzl, Director of the Innovation Lab at senswork, responsible for innovation processes and the professional development of employees. \nHis work focuses on translating AI-based technologies into organizational practice, with particular attention to innovation processes, responsibility, and human-centered design. In addition to internal training formats, he regularly designs and conducts workshops and educational programs for external audiences, including students and pupils, addressing digital technologies, innovation, and responsible technology use.\nTarget Audience:\nThe workshop targets researchers and practitioners working on industrial AI systems, including machine learning researchers, data scientists, engineers, and project managers. A basic understanding of AI concepts and data-driven decision-making is expected; no prior expertise in AI ethics is required.\n \nKeywords:\nResponsible AI; Ethically Aligned Design; Embedded Ethics; Corporate Digital Responsibility; Industrial AI; Uncertainty Quantification; Data Governance; Decision Support.

Brief Description and Outline:\nHow AI research is communicated plays a central role in shaping how AI is understood, trusted, and governed. AI researchers therefore have a unique opportunity to actively contribute to how their work is perceived and discussed beyond the lab. \nEngaging with journalists, policymakers, and interdisciplinary audiences allows researchers to communicate not only technical results, but also the nuances of uncertainty, limitations, and ongoing scientific disagreement.\n\nDeveloping communication skills strengthens both individual research impact and the field as a whole. This 2-hour workshop introduces core principles of science communication that are broadly applicable across scientific domains, while incorporating AI-specific examples. It equips researchers with practical tools to explain their work clearly to non-expert audiences. \n\nThe workshop is structured around the following themes: why science communication matters, the role of researchers in shaping narratives, core communication skills, and a practical discussion session.\n-\nGoals:\nThe workshop aims to: Participants will learn evidence-based insights from science communication and media studies about the communication of AI and equip AI researchers with practical tools to communicate their work responsibly to non-exper\n-\nPresenters Experience:\nVerena Albrecht is a communication expert at MCML researching public AI communication, with a focus on framing, emotions, and how European AI researchers engage with the public.\n-\nTarget Audience:\nPhD students, postdocs, and senior researchers working on AI or AI-related topics; no formal background in science communication is required.\n-\nKeywords:\nScience Communication, Artificial Intelligence, Responsible AI, Public Engagement, AI Narratives, Trust in AI, Research Impact

Brief Description and Outline:\nDuring this 2-hour interactive workshop participants will get an introduction to rapid prototyping. They will define an idea and explore what makes an idea work, followed by building a prototype and presenting it to the other participants.\nGoals:\nParticipants will explore PROTOTYPING, they will:\n\nLearn about the problem-solution fit and critical function of an idea\nGet to know different types of prototypes\nTrain their creativity and spontaneity by building a prototype\n\nPresenters Experience:\nHi, my name is Mirjam Rohrmann! My expertise spans from designing and managing training programs, organizing and moderating events, and delivering training, to outreach and communication. I\'m a certified YES (Young Entrepreneurs in Science) Trainer, offering workshops on exploring entrepreneurial skills, mindsets and career paths as well as user-centerd methods for generating and developing ideas.\nTarget Audience:\nThis workshop is only open to doctoral researchers and postdocs. It aims at researchers who have not started a business, but want to explore their entrepreneurial potential. No prior knowledge or experience in the areas of entrepreneurship, startups or other business ventures is necessary. A genuine interest in these topics is sufficient.\nKeywords:\nentrepreneurship, prototyping, design thinking

Brief Description and Outline:\nWhile HFMI and Helmholtz AI provide active fora to discuss machine learning from a methodological perspective, and there are dedicated conferences on weather AI (e.g., MLESM) and remote sensing (e.g., ESA ML for EO), there are few opportunities to establish connections between domain science questions and ML methods. In this workshop, we want to explore cross-links between different environmental application areas and compartments and identify opportunities for bringing some of these together with state-of-the-art ML methods. To complement this, we will look at cutting-edge ML methods including foundation models, world models, agentic models, and discuss if these offer potential for new discoveries across environmental application areas.\n\nGoals:\n\n\nBring together the disparate community of Helmholtz researchers who work on environmental problems and explore the use of AI\n\n\n Establish a forum (also beyond Helmholtz) to discuss cross-pollination from domain-specific application problems and AI methods\n\n\nDiscuss domain-specific methodological aspects like spherical geometry, multi-channel images, patterns in remote sensing images and weather maps\n\n\nBridge the current gap between remote sensing and weather AI communities and include other interesting applications\n\n\n\nPresenters Experience:\nProf. Dr. Martin Schultz Head of research group Earth System Data Exploration and co-lead of division Large Scale Data Science. More info here: https://www.fz-juelich.de/profile/schultz_m\n\nTarget Audience:\nResearchers from Helmholtz and beyond with some experience in AI and at least one environmental application area.\n\nKeywords:\nenvironmental science, earth system modelling, remote sensing, weather, climate, air pollution

Brief Description and Outline:\nWhile HFMI and Helmholtz AI provide active fora to discuss machine learning from a methodological perspective, and there are dedicated conferences on weather AI (e.g., MLESM) and remote sensing (e.g., ESA ML for EO), there are few opportunities to establish connections between domain science questions and ML methods. In this workshop, we want to explore cross-links between different environmental application areas and compartments and identify opportunities for bringing some of these together with state-of-the-art ML methods. To complement this, we will look at cutting-edge ML methods including foundation models, world models, agentic models, and discuss if these offer potential for new discoveries across environmental application areas.\n\nGoals:\n\n\nBring together the disparate community of Helmholtz researchers who work on environmental problems and explore the use of AI\n\n\n Establish a forum (also beyond Helmholtz) to discuss cross-pollination from domain-specific application problems and AI methods\n\n\nDiscuss domain-specific methodological aspects like spherical geometry, multi-channel images, patterns in remote sensing images and weather maps\n\n\nBridge the current gap between remote sensing and weather AI communities and include other interesting applications\n\n\n\nPresenters Experience:\nProf. Dr. Martin Schultz: Head of research group Earth System Data Exploration and co-lead of division Large Scale Data Science. More info here: https://www.fz-juelich.de/profile/schultz_m\nDr. Florentine Weber: Project Manager for the WeatherGenerator and RAINA, Science Coordinator for the Center for Earth System Observation and Computational Analysis (CESOC), Research Center Jülich and University of Cologne.\n\nTarget Audience:\nResearchers from Helmholtz and beyond with some experience in AI and at least one environmental application area.\n\nKeywords:\nenvironmental science, earth system modelling, remote sensing, weather, climate, air pollution

Brief Description and Outline:\nFoundation models are rapidly reshaping medical AI, enabling large‑scale representation learning across imaging, clinical text, biosignals, and multimodal health data. While recent advances demonstrate impressive performance across a wide range of downstream tasks, significant challenges remain in translating medical foundation models from pretraining to reliable clinical impact. These challenges include data heterogeneity, domain shift, limited annotation, evaluation biases, robustness, interpretability, and regulatory constraints. \n\nTo set the stage, the workshop will begin with a 45‑minute presentation by the organizers, showcasing ongoing collaborative efforts between Helmholtz Munich and German Cancer Research Center (DKFZ) toward the development, evaluation, and adaptation of medical foundation models. This session will highlight joint projects across multimodal learning, scalable pretraining, domain adaptation, and clinically aligned benchmarking, with the aim of fostering cross‑community exchange and new collaborations. \n\nThe workshop will also feature two invited talks by leading experts Ewa Szczurek (Helmholtz Munich) and Fabian Isensee (DKFZ), providing perspectives on scalable pretraining strategies and pathways to clinical translation. \n\nIn addition, registered participants (on ConfTool) can submit 300‑word extended abstracts (until May 3rd, EOD, see https://medfm-workshop.github.io/MedFM26/), from which a subset will be selected for short oral presentations, while remaining contributions will be presented during a poster session. To maximize interaction and visibility, poster presenters will have the opportunity to give one‑minute lightning pitches. The program includes a combined 45‑minute session of short talks and poster pitches, followed by a dedicated 45‑minute poster session to encourage in‑depth discussion and networking. The workshop will conclude with closing remarks and awards, recognizing outstanding contributions and strengthening community building around medical foundation models.\n-\nWorkshop Program Outline\n\n\n00:00 – 00:10 Welcome Remarks (10 min): Brief introduction, goals, and overview of the workshop.\n\n\n00:10 – 00:55 Talk: Projects on Medical Foundation Models by Workshop Organizers (45 min) Format: ~40 min presentation + 5 min Q&A Deep dive into the organizers’ ongoing projects, covering multimodal pretraining, domain adaptation, evaluation pipelines, and translational challenges.\n\n\n00:55 – 01:25 Invited Talk #1 (30 min) Format: 25 min talk + 5 min Q&A\n\n\n1:25 – 2:10 Short Talks + 1‑Min Poster Pitches (45 min): Short oral presentations: ~4–6 selected abstracts (5–7 min each. Poster lightning pitches: 1 minute per poster\n\n\n2:10 – 2:25 Coffee Break (15 min)\n\n\n2:25 – 3:10 Poster Session (45 min): Dedicated walk‑through and discussion with all presenters.\n\n\n3:10 – 3:40 Invited Talk #2 (30 min) Format: 25 min talk + 5 min Q&A (This replaces the previous panel discussion slot.)\n\n\n03:40 – 3:55 Closing Remarks + Awards (15 min): Summary, acknowledgments, and recognition of outstanding contributions.\n\n\n\nGoals:\nThis workshop aims to bring together researchers and practitioners working on foundation models for medicine, with a focus on the full lifecycle of model development and deployment: large-scale pretraining, adaptation and fine-tuning, evaluation and benchmarking, and real-world clinical applications. We seek contributions that advance methodological foundations as well as practical insights into deploying and validating medical foundation models in realistic settings. MedFM @ HAICON26 aims to provide a focused yet inclusive forum for advancing medical foundation models and for shaping a shared research agenda that bridges methodological innovation and clinical impact.\n\nPresenters Experience:\nEwa Szczurek: Co-director of Helmholtz Munich’s Institute of AI for Health; AI researcher in probabilistic and generative models for computational medicine.\nFabian Isensee: Senior Scientist at German Cancer Research Center; creator of nnU-Net, advancing AI for medical imaging and segmentation.\n\nTarget Audience:\nOur target audience includes PhD students, postdoctoral researchers, and scientists working in AI for medicine or general AI, as well as clinicians with a basic to intermediate understanding of AI who are interested in how foundation models can be translated into real clinical practice. The workshop is designed for participants who are familiar with fundamental machine learning concepts and who want to deepen their knowledge of large‑scale models, multimodal learning, evaluation, and clinical deployment.\n\nKeywords:\nMedical Foundation models; self‑supervised learning; multimodal learning; imaging AI; generalization; benchmarking; evaluation; clinical translation.

Brief Description and Outline:\nFoundation models are rapidly reshaping medical AI, enabling large‑scale representation learning across imaging, clinical text, biosignals, and multimodal health data. While recent advances demonstrate impressive performance across a wide range of downstream tasks, significant challenges remain in translating medical foundation models from pretraining to reliable clinical impact. These challenges include data heterogeneity, domain shift, limited annotation, evaluation biases, robustness, interpretability, and regulatory constraints. \n\nTo set the stage, the workshop will begin with a 45‑minute presentation by the organizers, showcasing ongoing collaborative efforts between Helmholtz Munich and German Cancer Research Center (DKFZ) toward the development, evaluation, and adaptation of medical foundation models. This session will highlight joint projects across multimodal learning, scalable pretraining, domain adaptation, and clinically aligned benchmarking, with the aim of fostering cross‑community exchange and new collaborations. \n\nThe workshop will also feature two invited talks by leading experts Ewa Szczurek (Helmholtz Munich) and Fabian Isensee (DKFZ), providing perspectives on scalable pretraining strategies and pathways to clinical translation. \n\nIn addition, registered participants (on ConfTool) can submit 300‑word extended abstracts (until May 3rd, EOD, see https://medfm-workshop.github.io/MedFM26/), from which a subset will be selected for short oral presentations, while remaining contributions will be presented during a poster session. To maximize interaction and visibility, poster presenters will have the opportunity to give one‑minute lightning pitches. The program includes a combined 45‑minute session of short talks and poster pitches, followed by a dedicated 45‑minute poster session to encourage in‑depth discussion and networking. The workshop will conclude with closing remarks and awards, recognizing outstanding contributions and strengthening community building around medical foundation models.\n-\nWorkshop Program Outline:\n\n\n00:00 – 00:10 Welcome Remarks (10 min): Brief introduction, goals, and overview of the workshop.\n\n\n00:10 – 00:55 Talk: Projects on Medical Foundation Models by Workshop Organizers (45 min) Format: ~40 min presentation + 5 min Q&A Deep dive into the organizers’ ongoing projects, covering multimodal pretraining, domain adaptation, evaluation pipelines, and translational challenges.\n\n\n00:55 – 01:25 Invited Talk #1 (30 min) Format: 25 min talk + 5 min Q&A\n\n\n1:25 – 2:10 Short Talks + 1‑Min Poster Pitches (45 min): Short oral presentations: ~4–6 selected abstracts (5–7 min each. Poster lightning pitches: 1 minute per poster\n\n\n2:10 – 2:25 Coffee Break (15 min)\n\n\n2:25 – 3:10 Poster Session (45 min): Dedicated walk‑through and discussion with all presenters.\n\n\n3:10 – 3:40 Invited Talk #2 (30 min) Format: 25 min talk + 5 min Q&A (This replaces the previous panel discussion slot.)\n\n\n03:40 – 3:55 Closing Remarks + Awards (15 min): Summary, acknowledgments, and recognition of outstanding contributions.\n\n\n\nGoals:\nThis workshop aims to bring together researchers and practitioners working on foundation models for medicine, with a focus on the full lifecycle of model development and deployment: large-scale pretraining, adaptation and fine-tuning, evaluation and benchmarking, and real-world clinical applications. We seek contributions that advance methodological foundations as well as practical insights into deploying and validating medical foundation models in realistic settings. MedFM @ HAICON26 aims to provide a focused yet inclusive forum for advancing medical foundation models and for shaping a shared research agenda that bridges methodological innovation and clinical impact.\n\nPresenters Experience:\nEwa Szczurek: Co-director of Helmholtz Munich’s Institute of AI for Health; AI researcher in probabilistic and generative models for computational medicine.\nFabian Isensee: Senior Scientist at German Cancer Research Center; creator of nnU-Net, advancing AI for medical imaging and segmentation.\n\nTarget Audience:\nOur target audience includes PhD students, postdoctoral researchers, and scientists working in AI for medicine or general AI, as well as clinicians with a basic to intermediate understanding of AI who are interested in how foundation models can be translated into real clinical practice. The workshop is designed for participants who are familiar with fundamental machine learning concepts and who want to deepen their knowledge of large‑scale models, multimodal learning, evaluation, and clinical deployment.\n\nKeywords:\nMedical Foundation models; self‑supervised learning; multimodal learning; imaging AI; generalization; benchmarking; evaluation; clinical translation.

\nBrief Description and Outline:\nSubtitle:\nBest practices for establishing benchmarks across emerging fields\n\nIn this session, we will explore the state of the art in the benchmarking field, showcasing the most useful tools and summarizing best practices for setting up benchmarks. The workshop program features invited talks by leading contributors in the benchmarking field, a poster session from projects funded by the Helmholtz UNLOCK benchmarking initiative, and concludes with a discussion panel on benchmark design with experts from the Vector Institute and Helmholtz Munich.\n-\n\n14:15 - 14:35 HumaniBench: A Human-Centric Benchmark for Large Multimodal Models Evaluation (Shaina Raza, Vector Institute, Canada)\n14:35 - 14:55 ChemBench: A benchmarking app for chemistry LLMs (TBA, Friedrich Schiller, University Jena)\n14:55 - 15:15 OpenProblems: a platform for benchmarking open problems in single-cell analysis (Robrecht Cannoodt, Edaro)\n15:15 - 15:35 AI Energy Consumption benchmarks (Philipp Huber, KIT)\n15:35 - 15:55 Do’s and don’ts for meaningful benchmarking (Annika Reinke, DKFZ)\n15:55 - 16:15 SPRIND competitions (Erik Schäffner, SPRIND)\n16:15 - 16:30 Coffee Break\n16:30 - 17:15 Poster session: discovering the UNLOCK project\n17:15 - 18:00 Panel discussion: Impact of benchmarks on the (scientific) community. Shaina Raza, Malte Luecken, Marie Piraud, Annika Reike, Erik Schäffner. Moderators: Daria Romanovskaia, Kaleb Phipps\nOfficial end of the workshop\n\n-\nGoals:\nWith growing numbers of AI models developed across scientific fields, establishment of benchmarks is essential. We want to bring together a community of researchers that have expertise in establishing benchmarks in their corresponding fields to exchange experience and provide practical advice. We see three main goals of this workshop:\n- Define a wide range of benchmarking goals in the AI space;\n- Highlight importance of global AI safety benchmarks;\n- Encourage exchange between UNLOCK project participants and broader benchmarking community.\n-\nPresenters Experience:\nShaina Raza, from Vector Institute, one of the leading institute in establishing benchmarks in the field of trustworthy AI; \nAmong Helmholtz Munich scientists: \nMarie Piraud - head of AI consultants in Helmholtz Munich; \nFrancesco Campi, PhD Candidate Helmholtz Munich\nSteffen Schneider’s group builds machine learning algorithms for representation learning and inference of nonlinear system dynamics; \nMalte Lücken and his group are one of the pioneers of benchmarking in single-cell genomics. \nDaria Romanovskaia, Scientist Lücken\'s Lab, Helmholtz Munich.\nWe in addition would invite two-three more scientists with relevant backgrounds. \nInteractive poster session will give an overview of the projects, funded by the UNLOCK benchmarking initiative.\n-\nTarget Audience:\nResearchers working on establishing AI benchmarks across different areas of expertise\n-\nKeywords:\nBenchmarks, Trustworthy AI, AI safety

Brief Description and Outline:\nSubtitle: Best practices for establishing benchmarks across emerging fields\nIn this session, we will explore the state of the art in the benchmarking field, showcasing the most useful tools and summarizing best practices for setting up benchmarks. The workshop program features invited talks by leading contributors in the benchmarking field, a poster session from projects funded by the Helmholtz UNLOCK benchmarking initiative, and concludes with a discussion panel on benchmark design with experts from the Vector Institute and Helmholtz Munich.\n-\n\n14:15 - 14:35 HumaniBench: A Human-Centric Benchmark for Large Multimodal Models Evaluation (Shaina Raza, Vector Institute, Canada)\n14:35 - 14:55 ChemBench: A benchmarking app for chemistry LLMs (TBA, Friedrich Schiller, University Jena)\n14:55 - 15:15 OpenProblems: a platform for benchmarking open problems in single-cell analysis (Robrecht Cannoodt, Edaro)\n15:15 - 15:35 AI Energy Consumption benchmarks (Philipp Huber, KIT)\n15:35 - 15:55 Do\'s and don\'ts for meaningful benchmarking (Annika Reinke, DKFZ)\n15:55 - 16:15 SPRIND competitions (Erik Schäffner, SPRIND)\n16:15 - 16:30 Coffee Break\n16:30 - 17:15 Poster session: discovering the UNLOCK project\n17:15 - 18:00 Panel discussion: Impact of benchmarks on the (scientific) community.Shaina Raza, Malte Luecken, Marie Piraud, Annika Reike, Erik Schäffner. Moderators: Daria Romanovskaia, Kaleb Phipps\nOfficial end of the workshop\n\n-\nGoals:\nWith growing numbers of AI models developed across scientific fields, establishment of benchmarks is essential. We want to bring together a community of researchers that have expertise in establishing benchmarks in their corresponding fields to exchange experience and provide practical advice. We see three main goals of this workshop:\n- Define a wide range of benchmarking goals in the AI space;\n- Highlight importance of global AI safety benchmarks;\n- Encourage exchange between UNLOCK project participants and broader benchmarking community.\n-\nPresenters Experience:\nShaina Raza, from Vector Institute, one of the leading institute in establishing benchmarks in the field of trustworthy AI; \nAmong Helmholtz Munich scientists: \n\nMarie Piraud - head of AI consultants in Helmholtz Munich; \nFrancesco Campi, PhD Candidate Helmholtz Munich\nSteffen Schneider’s group builds machine learning algorithms for representation learning and inference of nonlinear system dynamics; \nMalte Lücken and his group are one of the pioneers of benchmarking in single-cell genomics. \nDaria Romanovskaia, Scientist Lücken\'s Lab, Helmholtz Munich.\nWe in addition would invite two-three more scientists with relevant backgrounds. \nInteractive poster session will give an overview of the projects, funded by the UNLOCK benchmarking initiative.\n-\nTarget Audience:\nResearchers working on establishing AI benchmarks across different areas of expertise\n-\nKeywords:\nBenchmarks, Trustworthy AI, AI safety

Brief Description and Outline:\nThis workshop aims to actively discuss the role of concepts of causality in complex dynamic systems and how the application of these concepts can be challenging to address in medicine and biology.\nUnderstanding cause–and–effect relationships rather than pure correlations is the aim of causal modeling, including causal inference and causal AI. Causal inference and reinforcement learning developed as separate disciplines with distinct terminologies yet address mathematically related problems. Recent work has begun to reveal these fundamental connections: online reinforcement learning inherently captures causal relationships, whereas traditional methods show that advantage functions and mean-cantered blip functions are mathematically equivalent objects under uniform policies, and the inverse probability of treatment weighting represents the same probability weighting principle as importance sampling in contextual bandits.\n\nHowever, causal inference in complex systems must account for limited completeness and contextual causality, particularly in biological and physiological domains where feedback loops and multiscaling create temporal dependencies that static causal diagrams cannot capture. On the other hand, statistical methods, such as synergistic, unique and redundant components (SURD), attempt to solve inherent  problems of causality in complex dynamical systems by determining the specific nature of causal relationships, such as whether two variables are synergistic, i.e., whether one variable only influences another if it is paired with a second variable.\n\nThis workshop explores how recent advances in causal discovery using reinforcement learning, estimation of heterogeneous treatment effects, and adaptive experimental design provide opportunities for practitioners to use methods from both modeling methods, namely, reinforcement learning and causal AI. Additionally, we will thematize how to address, for example, complex biological systems that exhibit persistent dependencies and emergent properties that violate standard assumptions in both causal inference and reinforcement learning.\n-\nBuilding on these observations, the workshop addresses three fundamental questions at the interface of causal inference and reinforcement learning in biomedical systems:\n• First, when do control objectives permit bias that inference objectives cannot tolerate, for instance, when optimizing treatment sequences from observational data?\n• Second, how do we handle systems where traditional Markovian assumptions fail?\n• Third, what new methodological opportunities emerge from the explicit recognition of these translational synergies?\n-\nThis discussion emphasizes practical implications for researchers working across causal inference and machine learning divides, particularly in domains requiring sequential decision-making under uncertainty with observational data constraints. Applications span clinical decision support systems where treatment sequences must be optimized from observational data, adaptive trial designs that balance exploration and exploitation while maintaining statistical validity, digital twins (DTs) where therapies or the toxic effects of substances are tested on in silico models of real organisms, and policy evaluation in public health where experimental manipulation is impossible.\n-\nOutline of the Workshop\n• 20 min Introduction session including presenters and participants\n• 1-hour in-depth presentations of presenters\n• 20 min Introduction group work = Setting scene/clinical vignette/t thematic context: particularly in domains requiring sequential decision-making under uncertainty with observational data constraints.\n• 50 Minutes an active parallel workshop on one of the three fundamental open questions from the perspective of causality and reinforcement learning:\n1. First, when, if ever, do control objectives permit bias that inference objectives cannot tolerate?\n2. Second, how do we handle systems where traditional Markovian assumptions fail?\n3. Third, what new methodological opportunities emerge from the explicit recognition of these translational synergies?\nInput to groupwork: literature, guidance through questions.\n-\nDeliverable: brainstorming on possibilities of algorithmic architectures, limitations.\nDashboard for design of algorithmic solutions to questions.\n• 40 min Presentation & Discussion of developed architectures\n• 20 Minutes wrap up\n-\nGoals:\nDiscussion of the challenges and opportunities in applying causal modeling to biomedical systems, with a focus on the limitations of current frameworks\n• Identify key requirements for integrating causal AI with digital twin architectures in clinical and biomedical settings.\n• Explore how recent advances in causal discovery, heterogeneous treatment effect estimation, and adaptive experimental design can be combined with reinforcement learning.\n• Provide participants with hands-on exposure to a real-world use case: NLP-based patient stratification and its potential extension toward causal modeling in emergency oncology.\n• Findings should be synthesized into actionable recommendations for researchers working at the intersection of causal inference and machine learning.\n• Intended impact: Assessment of the potential for the design of systems that could work autonomously in real-world clinical settings.\n-\nPresenters Experience:\nJuan G. Diaz Ochoa is an astrophysicist (Observatorio Astronómico Nacional de Colombia) and physicist specializing in complex systems, systems biology and systems medicine. After completing his doctorate in physics (solid-state physics with Prof. Kurt Binder, University of Mainz), he held various research and development positions, including at the Institute for Theoretical Physics in Bremen, at the Max Planck Institute for Complex Technical Systems (Magdeburg) and at Insilico AG, where he led and led various national and EU projects. In recent years, Juan G. Diaz Ochoa has been working on the application of machine learning and artificial intelligence in medicine and has led projects to develop graph-based knowledge-based platforms and products for the efficient evaluation of unstructured data and ontologies in nephrology and oncology. Juan G. Diaz Ochoa is the author of several articles and book chapters that have been published in international journals. He is also the author of the book \"Complexity measurements and Causation for Dynamic Complex Systems\", and is responsible for the co-organization of the Solid Symposium and is a lecturer in mathematics and physics at the Duale Hochschule Baden-Württemberg (DHBW) in Germany.\nProf. Dr. Alexander Krohn is a professor at MUC.HEALTH.\nDr. Johann Krocza is CEO at HSE365.at GmbH and Medifina GmbH.\n-\nTarget Audience:\nPractitioners are interested in the application of causal AI and reinforcement learning in a clinical framework and are involved in AI architectures; in particular, professionals working with longitudinal data, machine learning researchers interested in causal methods, clinical trial methodologists, and anyone working on sequential decision problems in observational settings. Given that the methodological challenges addressed in this workshop, such as causal inference in complex dynamic systems and learning from observational data, are shared across scientific domains, researchers from adjacent Helmholtz research areas (e.g., climate, earth systems, or materials science) are equally welcome.\n-\nKeywords:\ncausal inference • reinforcement learning • machine learning • Digital Twins • complex dynamical systems • sequential decision problems • Support, NLP • Biomedical AI

Brief Description and Outline:\nThis workshop aims to actively discuss the role of concepts of causality in complex dynamic systems and how the application of these concepts can be challenging to address in medicine and biology.\nUnderstanding cause–and–effect relationships rather than pure correlations is the aim of causal modeling, including causal inference and causal AI. Causal inference and reinforcement learning developed as separate disciplines with distinct terminologies yet address mathematically related problems. Recent work has begun to reveal these fundamental connections: online reinforcement learning inherently captures causal relationships, whereas traditional methods show that advantage functions and mean-cantered blip functions are mathematically equivalent objects under uniform policies, and the inverse probability of treatment weighting represents the same probability weighting principle as importance sampling in contextual bandits.\n\nHowever, causal inference in complex systems must account for limited completeness and contextual causality, particularly in biological and physiological domains where feedback loops and multiscaling create temporal dependencies that static causal diagrams cannot capture. On the other hand, statistical methods, such as synergistic, unique and redundant components (SURD), attempt to solve inherent  problems of causality in complex dynamical systems by determining the specific nature of causal relationships, such as whether two variables are synergistic, i.e., whether one variable only influences another if it is paired with a second variable.\n\nThis workshop explores how recent advances in causal discovery using reinforcement learning, estimation of heterogeneous treatment effects, and adaptive experimental design provide opportunities for practitioners to use methods from both modeling methods, namely, reinforcement learning and causal AI. Additionally, we will thematize how to address, for example, complex biological systems that exhibit persistent dependencies and emergent properties that violate standard assumptions in both causal inference and reinforcement learning.\n\nBuilding on these observations, the workshop addresses three fundamental questions at the interface of causal inference and reinforcement learning in biomedical systems:\n• First, when do control objectives permit bias that inference objectives cannot tolerate, for instance, when optimizing treatment sequences from observational data?\n• Second, how do we handle systems where traditional Markovian assumptions fail?\n• Third, what new methodological opportunities emerge from the explicit recognition of these translational synergies?\n\nThis discussion emphasizes practical implications for researchers working across causal inference and machine learning divides, particularly in domains requiring sequential decision-making under uncertainty with observational data constraints. Applications span clinical decision support systems where treatment sequences must be optimized from observational data, adaptive trial designs that balance exploration and exploitation while maintaining statistical validity, digital twins (DTs) where therapies or the toxic effects of substances are tested on in silico models of real organisms, and policy evaluation in public health where experimental manipulation is impossible.\n\nOutline of the Workshop\n• 20 min Introduction session including presenters and participants\n• 1-hour in-depth presentations of presenters\n• 20 min Introduction group work = Setting scene/clinical vignette/t thematic context: particularly in domains requiring sequential decision-making under uncertainty with observational data constraints.\n• 50 Minutes an active parallel workshop on one of the three fundamental open questions from the perspective of causality and reinforcement learning:\n1. First, when, if ever, do control objectives permit bias that inference objectives cannot tolerate?\n2. Second, how do we handle systems where traditional Markovian assumptions fail?\n3. Third, what new methodological opportunities emerge from the explicit recognition of these translational synergies?\nInput to groupwork: literature, guidance through questions.\n\nDeliverable: brainstorming on possibilities of algorithmic architectures, limitations.\nDashboard for design of algorithmic solutions to questions.\n• 40 min Presentation & Discussion of developed architectures\n• 20 Minutes wrap up\n\nGoals:\nDiscussion of the challenges and opportunities in applying causal modeling to\nbiomedical systems, with a focus on the limitations of current frameworks\n• Identify key requirements for integrating causal AI with digital twin architectures in clinical and biomedical settings.\n• Explore how recent advances in causal discovery, heterogeneous treatment effect estimation, and adaptive experimental design can be combined with reinforcement learning.\n• Provide participants with hands-on exposure to a real-world use case: NLP-based patient stratification and its potential extension toward causal modeling in emergency oncology.\n• Findings should be synthesized into actionable recommendations for researchers working at the intersection of causal inference and machine learning.\n• Intended impact: Assessment of the potential for the design of systems that could work autonomously in real-world clinical settings.\n\nPresenters Experience:\nJuan G. Diaz Ochoa is an astrophysicist (Observatorio Astronómico Nacional de Colombia) and physicist specializing in complex systems, systems biology and systems medicine. After completing his doctorate in physics (solid-state physics with Prof. Kurt Binder, University of Mainz), he held various research and development positions, including at the Institute for Theoretical Physics in Bremen, at the Max Planck Institute for Complex Technical Systems (Magdeburg) and at Insilico AG, where he led and led various national and EU projects. In recent years, Juan G. Diaz Ochoa has been working on the application of machine learning and artificial intelligence in medicine and has led projects to develop graph-based knowledge-based platforms and products for the efficient evaluation of unstructured data and ontologies in nephrology and oncology. Juan G. Diaz Ochoa is the author of several articles and book chapters that have been published in international journals. He is also the author of the book \"Complexity measurements and Causation for Dynamic Complex Systems\", and is responsible for the co-organization of the Solid Symposium and is a lecturer in mathematics and physics at the Duale Hochschule Baden-Württemberg (DHBW) in Germany.\nProf. Dr. Alexander Krohn is a professor at MUC.HEALTH.\nDr. Johann Krocza is CEO at HSE365.at GmbH and Medifina GmbH.\n\n-\nTarget Audience:\nPractitioners are interested in the application of causal AI and reinforcement learning in a clinical framework and are involved in AI architectures; in particular, professionals working with longitudinal data, machine learning researchers interested in causal methods, clinical trial methodologists, and anyone working on sequential decision problems in observational settings. Given that the methodological challenges addressed in this workshop, such as causal inference in complex dynamic systems and learning from observational data, are shared across scientific domains, researchers from adjacent Helmholtz research areas (e.g., climate, earth systems, or materials science) are equally welcome.\n-\nKeywords:\ncausal inference • reinforcement learning • machine learning • Digital Twins • complex dynamical systems • sequential decision problems • Support, NLP • Biomedical AI

Brief Description and Outline:\nThe tutorial covers both practical tools to obtain uncertainty estimates in deep learning, as well as a theoretical understanding of various deep learning phenomena through a Bayesian lens.\nOutline:\n\nPart 1 -- Foundations of Bayesian Principles for Deep Learning (30 minutes)\nPart 2 -- Practical Tools (60 Minutes)\n2.1: Variational Methods (45 Minutes)\n2.2: Laplace Approximations (15 Minutes)\nPart 3: Understanding Deep Learning via Bayesian Principles (30 Minutes)\n\n\nReferences [1] https://arxiv.org/abs/2107.04562 [2.1] https://arxiv.org/abs/2402.17641 https://github.com/team-approx-bayes/ivon [2.2] https://arxiv.org/abs/2106.14806 https://github.com/aleximmer/Laplace [3] https://arxiv.org/abs/2110.11216 https://arxiv.org/abs/2503.02113\n\nGoals:\nAfter completion of the tutorial, participants will understand: \n\n\nLearn how to use recent practical tools (Variational Learning and Laplace Approximations) to obtain uncertainty estimates for deep neural networks\n\n\nUnderstand how these uncertainty estimates can be used in practice \n\n\nA better understanding of generalization in deep learning and phenomena such as overfitting, double descent, grokking, which helps to answer questions such as: \"How big should my model be\"?\n\n\nDeep learning has become the go-to choice for many practical problems (including ones that HAICON26 attendants might be interested in).\n\n\nThe tutorial teaches\n1) practical tools that enables practitioners to obtain uncertainty estimates\n2) an understanding how these uncertainty estimates can be used to improve models\n3) obtain a better understanding of various mysterious phenomena in deep learning.\n\nPresenters Experience:\nThomas Moellenhoff is a Senior Research Scientist at the RIKEN Center for Advanced Intelligence Project in Tokyo, Japan. He received his PhD in 2020 from the Technical University of Munich, under the supervision of Daniel Cremers. After that, he was a post-Doc from 2020-2023 working with Emtiyaz Khan at RIKEN AIP. His research lies in the intersection of optimization and Bayesian deep learning. He has co-organized several workshops, including the ICML 2023 workshop on duality principles in modern machine learning. His homepage is: https://moellenh.github.io\n\nTarget Audience:\nThe target audience are students (both undergraduate and master/PhD level) as well as researchers and practitioners from applied fields and industry. Prerequisites are basic mathematics (linear algebra, probability, calculus) and some previous experience with machine learning and deep learning.\n\nKeywords:\ndeep learning, uncertainty quantification, bayesian deep learning, variational inference

Brief Description and Outline:\nAgentic AI is becoming ubiquitous, but how an agent learns to navigate an environment efficiently remains an open problem. This session traces the exploration problem from its classical formulation in reinforcement learning (where the tension between exploiting known rewards and discovering better ones is unavoidable) through to contemporary approaches rooted in intrinsic motivation, meta-learning, and the free energy principle. The arc moves from hand-designed exploration heuristics toward increasingly principled accounts of curiosity, culminating in active inference as a unifying framework that grounds exploration in Bayesian brain theory and statistical mechanics. Relevance: Efficient exploration is a bottleneck in deploying RL to many real-world problems where rewards are sparse, environments are complex, and data is expensive, making it central to both theoretical and applied AI research. Beyond engineering, the topic sits at the intersection of machine learning, dynamical systems, and computational neuroscience, offering a lens through which mathematical structure can illuminate biological as well as artificial intelligence. Researchers across these fields may find direct connections to open problems in their own work, whether in algorithm design, cognitive modelling, or the foundations of learning theory.\n\nOutline:\n1) Reinforcement Learning and the Exploration Problem (20 min)\n2) Curiosity and Intrinsic Motivation in RL (30-40 min)\n3) Meta-Reinforcement Learning and Learning-to-Explore (20 min)\n4) Active Inference Frameworks (20 min)\n5) Discussion and Open Questions (20-30 min)\n\nGoals:\nTo provide a conceptual map of modern exploration and curiosity-driven methods in reinforcement learning. To clarify the relationship between intrinsic motivation, meta-learning, and adaptive behaviour and to highlight connections between reinforcement learning, Bayesian inference, and active inference frameworks. Also, to equip participants with principled ways to reason about exploration strategies beyond ε-greedy or ad-hoc heuristics.\n\nPresenters Experience:\nJonathan Shock is an Associate Professor in the Department of Maths and Applied Maths at the University of Cape Town, where he directs the UCT AI Initiative. He is also an Adjunct Professor at INRS Montréal. His work spans theoretical physics, reinforcement learning, and computational neuroscience, with a focus on understanding intelligence as a dynamical and physical process. He is particularly interested in reinforcement learning, multi-agent systems, and theory-driven AI for science. Much of his research explores how mathematical structure — from statistical mechanics to dynamical systems — can inform the design and analysis of learning systems. He completed his PhD at the University of Southampton in 2005 on applications of string theory to quantum chromodynamics, followed by postdoctoral appointments in Beijing, Santiago de Compostela, and Munich before joining UCT in 2013.\n\nTarget Audience:\na) PhD students, postdoctoral researchers, and early-to mid-career researchers\nb) Participants with a basic familiarity with machine learning or reinforcement learning\nc) Researchers from computer science, applied mathematics, neuroscience or robotics.\nThis tutorial is designed as a 2-hour session focusing at a semi-technical level introducing a range of topics but not expecting mastery. No coding or specialised software is required, but some mathematics background is expected.

Brief Description and Outline:\nIn machine learning, the ability to make reliable predictions is paramount. Yet, standard ML models and pipelines provide only point predictions without accounting for model confidence (or the lack thereof). Uncertainty in model outputs, especially when faced with out-of-distribution (OOD) data, is essential when deploying models in production. This tutorial serves as an introduction to the concepts and techniques for quantifying uncertainty in machine learning models. We will explore the different sources of uncertainty and cover various methods for estimating these uncertainties effectively.\n\nGoals:\nProvide participants with basic knowledge and go-to tools and methods for applying UQ in their Work.\n\nPresenters Experience:\nSteve Schmerler, AI consultant @HZDR, has hosted an UQ workshop and world cafe table at previous HAICONs. Further, he has been teaching neural net intro courses such as https://github.com/elcorto/2024-thrill-school-machine-learning and has developed educational material for Gaussian processes (https://elcorto.github.io/gp_playground).\nJakob Gawlikowski is a PhD Student at German Aerospace Center (DLR) and Technical University of Munich.\nChristoph Schweden is a postdoctoral fellow at German Aerospace Center (DLR).\n\nTarget Audience:\nFamiliarity with neural nets, basic probability (ML textbook level), Bayesian stats would be ideal but not required.\n\nKeywords:\nUncertainty, approximate Bayesian, calibration

You had until 26 May to indicate on your registration form if you required childcare. We look forward to welcoming those of you who opted for this service.\nWe\'ll look after your little scientists while you enjoy HAICON26! :)

Dinner is planned from 19:30 h. The reservation is for the beer garden, and by mentioning the Helmholtz AI booking upon arrival, the staff will show guests to the correct seating area.\nPlease note that dinner costs are to be covered by participants.\nHow to get there:\nThe most straightforward way to get there is to take bus 295 from Helmholtz to \"Am Hart\", and then the U2 from \"Am Hart\" to \"Hauptbahnhof\" / main station. It is then a short walk from there to the Augustiner-Keller. \nAddress: Augustiner-Keller, Arnulfstraße 52, 80335, München\nGoogle maps link: https://maps.app.goo.gl/3VMXhNnpjN5xtHFk6

Coffee breaks take place at the Conference Center and the Helmholtz Pioneer Campus (HPC) simultaneously. Bear in mind the time to walk from one location to another (max 7 min) and the fact that workshops will start on time.\nDrinks, but not food, are allowed to be taken with you into the rooms. Please bring anything you brought into the room back with you at the end of your session.

Coffee breaks take place at the Conference Center and the Helmholtz Pioneer Campus (HPC) simultaneously. Bear in mind the time to walk from one location to another (max 7 min) and the fact that workshops will start on time.\nDrinks, but not food, are allowed to be taken with you into the rooms. Please bring anything you brought into the room back with you at the end of your session.