Conference Agenda

Description

For the first time in history, knowledge, analysis, and decision support are being generated not only by humans, but alongside them. AI systems design experiments, write code, generate hypotheses, optimize processes, and increasingly contribute directly to measurable output. Artificial work is scaling, but our economic models, incentive systems, and institutional structures still assume human labor as the sole source of value creation.

This World Café explores the economic consequences of that shift:
When AI contributes to productivity, who captures the value?
How do business models evolve when “labor” becomes partially autonomous?
Do we need new ownership, taxation, or compensation frameworks for artificial agents?
How should research institutions adapt funding logic, IP models, and performance metrics?
What happens to cost structures and competitive advantage when cognitive work scales at near-zero marginal cost?

Rather than discussing tools, we will examine structural transformation: from platform economics and AI-native research models to the redesign of governance and incentive systems.
The objective is to explore how Europe, and institutions such as Helmholtz, can proactively shape sustainable economic architectures for the age of artificial work, rather than merely reacting to technological acceleration.

Description

AI tools are rapidly transforming research practices, yet institutional governance often struggles to keep pace. Science managers face a complex challenge: balancing regulatory compliance (data protection, EU AI Act, DFG guidelines, internal regulations...), a rapidly evolving tool landscape, heterogeneous skill levels across professional groups, and significant uncertainty about risks and opportunities. Drawing on real-world experiences of participants, this table invites to explore how AI governance can move beyond static policies towards an agile, practice-oriented platform approach. The main goal would be to foster a shared understanding of what "good (enough) governance" could look like under conditions of uncertainty, and to explore what kind of approach could address challenges many research institutions are confronted with right now.

Core questions include:
--------------------------------------------
- Regulation vs. usability
- How do we establish clear boundaries without governance becoming so restrictive that it drives informal workarounds? What does "good (enough) governance” look like under regulatory uncertainty?
- Enablement and skills
- Researchers, administrators, and IT staff have fundamentally different AI literacy and needs. How do we build training and support structures that work across these differences?
- Organisational implementation

How do institutions keep track when tools evolve faster than policy cycles? Who owns AI governance, who enforces it, and what are realistic lightweight mechanisms that don't slow down practical use cases in science management and research?

Description

We would like to bring together colleagues from across Helmholtz who want to strengthen Data Readiness for AI. The focus is on extending an existing criteria list so that it works across many domains, adding requirements from different research areas, developing short and accessible training materials, and exploring how these elements can support a simple AI readiness checker that offers tailored improvement suggestions.

We are especially looking for colleagues who can contribute cross‑domain perspectives and domain‑specific knowledge to ensure that the criteria and the tool reflect the diversity of data and analytical settings in Helmholtz. If this aligns with your interests, we would be glad to have you join.

Description

Large particle accelerators such as DESY generate extremely bright X-ray beams, which are then used for imaging and analysis of living and non-living objects, which furthers our understanding of the world. However, the detectors and setups used in the experiments result in noisy, incomplete, or intensity-only information. This invites algorithms to be developed that tackle these inverse problems through optimization methods, including deep learning.
The complete reconstruction pipeline at our phase-contrast imaging entails a chain of ill-posed inverse problems: (i) flat-field correction; (ii) denoising without erasing physical fine structure; (iii) phase and absorption retrieval from the measured propagated intensity; (iv) tomographic reconstruction; and (v) segmentation.

In this table, we delve deeper into unified self-supervised framework that tackles these challenges without the need for training data. The opportunities and challenges it has for other challenges will be discusses and methods on how to accelerate and deploy this solution.

Keywords: inverse problems, physics-informed, tomography, phase-retrieval, flat-field correction, deconvolution, uncertainty quantification

Description

I would like to open a discussion on the role of Large Vision-Language Models (LVLMs) in Earth Observation and environmental applications.

In particular, I would be interested in discussing the following aspects:

*. POTENTIAL AND LIMITATIONS OF VISION-LANGUAGE MODELS
-----------------------------------------------------------------------------------------------------------
** What are the real capabilities of these models beyond impressive demonstrations?
** To what extent can they support reasoning, interpretation, and decision-making from remote sensing imagery?
** What are the current limitations in terms of reliability, explainability, and generalization?


* Challenges and limitations in rapid disaster response
----------------------------------------------------------------------------------------
** Can VLMs provide actionable and trustworthy information during emergencies such as floods, wildfires, or earthquakes?
** How do these models perform under conditions of limited, noisy, or heterogeneous data?
** What are the main bottlenecks preventing operational deployment in time-critical scenarios?


*Challenges and limitations for sustainable agriculture and greenhouse gas mitigation
------------------------------------------------------------------------------------------------------------------------
**How can VLMs contribute to monitoring sustainable agricultural practices and environmental impacts?
**Can these models support better understanding of crop stress, land management, and emissions-related processes?
**What are the risks, biases, or technical limitations that still need to be addressed?

Description

This World Café explores key issues in AI engineering for safety-critical applications. The focus is on integrating artificial intelligence methods with engineering processes to ensure that AI-based systems can be deployed responsibly and in compliance with standards across a range of fields, from automated mobility and critical infrastructure to biological, medical, and life science applications. Using their own examples, participants can discuss challenges such as the safety, robustness, controllability, and certifiability of learning systems. In addition, we discuss methods for evaluating, validating, and testing AI throughout its entire lifecycle.
The goal is to develop a shared understanding of the technical, regulatory, and operational requirements that safe and secure AI systems must meet, as well as the new engineering approaches needed to achieve them.

Description

Artificial intelligence is moving beyond models that only predict, classify, or generate content. The next wave will be Physical Intelligence: AI systems that reason with physical constraints, interact with scientific instruments and simulations, validate their own outputs, and iteratively improve through executable feedback from the real world. This topic proposes an open discussion of how Physical Intelligence can serve as a unifying paradigm for AI in Science. The starting point is the convergence of three complementary directions: GENIUS, which develops agentic workflows for generating and repairing executable simulation protocols; GENIUS-X, which extends this vision toward explainable, executable, and experimentally aware AI for materials science; and SABER-X, which brings self-validating AI principles into multimodal imaging, synchrotron/experimental workflows, and biofilm-related scientific discovery. Together, these projects suggest a shift from static AI models toward closed-loop scientific agents: systems that combine knowledge graphs, foundation models, physical constraints, uncertainty estimation, simulation engines, experimental data, and automated validation. Instead of asking only whether AI can make accurate predictions, Physical Intelligence asks whether AI can participate in the scientific method itself: proposing hypotheses, generating protocols, executing or validating them, detecting failure modes, and refining its own knowledge. This discussion will explore what is technically required to build such systems, what “trust” and “validation” mean in this context, and how Helmholtz infrastructures could become living environments for physically grounded, self-correcting AI.

Description

Generative and agentic AI systems are rapidly changing how educational resources are created, revised, and delivered. Educators can now use AI to generate explanations, exercises, visualizations, and interactive demonstrations, as well as to review and improve existing teaching materials. At the same time, these tools raise important pedagogical questions: How can AI meaningfully support course design and teaching workflows? How do we ensure quality, accuracy, and pedagogical value in AI-assisted materials? And how should we guide learners in using AI tools responsibly as part of their learning process, including considerations around assessment, academic integrity, and personalization?

This discussion table invites educators, researchers, and practitioners to explore how AI can become a productive partner in teaching. We will reflect on how educational materials have traditionally been created, how AI is already changing individual workflows, and where the greatest opportunities lie: generating course content and visualizations, designing exercises, improving student engagement and creating more inclusive and adaptive learning experiences. We discuss how to evaluate whether AI-assisted workflows genuinely enhance learning outcomes, rather than merely making learning material development faster. We will also touch on evolving practices to make assessment “AI-proof” and how to evaluate whether AI-assisted workflows genuinely enhance learning outcomes.

The discussion will be structured in three moderated rounds. After a short framing introduction, participants will engage in focused discussions on (1) current practices and experiences with AI in teaching, (2) quality, assessment, and learning impact, and (3) inclusion, personalization, and opportunities for collaboration, including how AI can enable interactive learning experiences, simulations, and visualizations using emerging tools and approaches.. The session will be actively facilitated with guiding questions and light time management to ensure a dynamic exchange and that a wide range of perspectives can be heard. The session will conclude with a synthesis of key insights and potential next steps for collaboration within the community.

Beyond sharing experiences with tools and workflows, the session aims to connect participants interested in education across research centers and disciplines. By exchanging practices and ideas, we hope to identify opportunities for collaboration, shared resources, and community-driven approaches to AI-assisted teaching within the Helmholtz AI network and beyond.

Participants are welcome regardless of their prior experience with AI in teaching; both experimentation and skepticism are valuable perspectives for the discussion.

Description

Artificial Intelligence (AI) is increasingly central to biomedical research, from medical imaging and genomics analysis, to drug design, to clinical decision support, and much more. While this technology harbours extensive potential for progressing biomedical research, it is essential for AI researchers to consider how societal and cultural biases permeate medical practices and technology development, therefore introducing biases in data collection and analysis, and models development and evaluation. This is especially important to consider when applying novel technologies such as AI to make predictions (medical or other) about minoritized groups that are historically underrepresented in medical studies, such as women, BIPOC (1) individuals, LGBTQIA+ (2) individuals, and others. At the HAICON 2026 World Café we want to open a conversation on where AI fits into the progress of biomedical research, talking about the current state of AI applications in minority health, its future impacts, and how to best leverage these emerging tools for a more inclusive biomedical research.

An interesting read for you before our meeting (not compulsory, ~1h read): https://data-feminism.mitpress.mit.edu/pub/frfa9szd/release/6

(1) Black, Indigenous and People Of Colour
(2) Lesbian, Gay, Bisexual, Trans, Queer, Intersex, Asexual and Others

Description

Generative AI is unparalleled at writing syntax, but it fundamentally optimizes for compiling code, not necessarily factual truth or domain accuracy. When tasked with complex and domain-specific problems, an unguided AI can often confidently hallucinate rules or silently discard established principles just to produce working code within its context window. This can lead to "invisible failures", outputs that look highly convincing but are fundamentally invalid.

To safely harness AI in scientific analyses without letting it silently overwrite your expertise, we must redefine how we work together. This table explores how adopting a strict "Domain Architect vs. Syntax Engine" workflow acts as the ultimate guardrail against AI hallucinations. Join us to discuss how you can strictly dictate the logic and requirements of your code, restrict the AI entirely to syntax generation, and share practical strategies for catching silent errors in your own field.

Description

To effectively use a model, the instructions how to perform inference and set up the required compute environment needs to be clear. However, models developed by scientists often lack clear documentation. Furthermore, comparing different models on the same problem can be time-consuming, as each model requires its own setup procedures. Therefore, I want to discuss how to best provide (unified) access to different models and to which extend standardisation makes sense. Some of the relevant questions for this World Café are:
-How to unify access to models? Is it realistic to provide a python “hub” package?
-What level of abstraction is needed?
-What do agents need to easily use models?
-What do humans need to easily use models? And do we need to care now that we have agents?
-Can we formalise model metadata (e.g. to make it easier to pick the right model for the problem)?
-What can be a good interface to interference providers?
A starting point for this discussion can be the already existing initiatives that try to address part of these questions:
-Timm/pytorch image models https://github.com/huggingface/pytorch-image-models https://huggingface.co/docs/timm/index (standardised interface to computer vision models)
-Helmholtz Model Zoo https://helmholtz-imaging.de/news-news/news/helmholtz-model-zoo-now-open-to-all-helmholtz-members/ as inference provider
-Kipoi https://kipoi.org/ (genomic ML models)

Description

The rapid growth and increasing complexity of AI/ML demand for an effective management and monitoring of models throughout the entire lifecycle. This approach, known as Machine Learning Operations (MLOps), has become an essential practice for developing, deploying, and maintaining reliable AI/ML systems. In this World Café we are going to start with the following key points:

* In your daily work, what parts of the AI/ML lifecycle do you encounter?
* Which ones do you spend the most time on?
* Which parts do you find cumbersome?
* Which tools/platforms do you know of? Which ones do you use? Which ones would you like to use?
* Do you see use cases for integrating agentic workflows or tooling?

As a result of the discussion, we hope to define a prioritized and up-to-date AI/ML lifecycle loop where for every stage there is a list of the most demanded and widely used tools and services, potentially tailored to the needs of Helmholtz researchers.

Description

Traditional centralized machine learning requires aggregating data in one place, raising serious concerns around privacy, compliance, and scalability. Federated Learning (FL) addresses this by enabling model training directly where the data lives. Only model updates, not raw data, are shared. With regulations like GDPR and the EU AI Act tightening data governance, and frameworks like Flower and NVFlare maturing rapidly, FL is getting easier than before to apply in real practice.

The frontier today is applying FL to large foundation models, from language models to vision transformers, in cross-institutional settings. Adapting these models to specialized tasks in healthcare, environmental monitoring, or urban infrastructure is especially challenging: data is distributed across organizations with different hardware, data formats, modalities, and computational capacity. Existing FL approaches were simply not designed for this scale and heterogeneity.

In this session, we'll discuss:

In your daily work, do you encounter situations where data cannot easily be centralized? If so, how do you handle it?

How do you ensure compliance with GDPR and the EU AI Act when handling sensitive data in your work?

Which use cases in your research or organization could benefit from Federated Learning?

What technical safeguards are necessary: secure aggregation, differential privacy, encryption, access control, audit logs, or trusted execution environments?

What are the main barriers you see for applying Federated Learning in practice?

Do you see a need for federated fine-tuning of foundation models across institutions?

Which challenges seem most critical: communication overhead, non-IID data, infrastructure heterogeneity, privacy risks, evaluation, or governance?

How can a model be evaluated when data, labels, and ground truth are distributed across different sites?

What kind of shared infrastructure, pilot projects, or working group would help Helmholtz researchers explore Federated Learning in practice?

As a result of the discussion, we hope to identify the most relevant Federated Learning use cases, challenges, tools, and collaboration opportunities for Helmholtz researchers

Description

Recent advances in neural operators and physics-informed networks have shown great potential for accelerating engineering simulations while maintaining good accuracy.

This discussion explores the question:

**Can AI become a practical alternative to traditional numerical simulations such as FEM and CFD, or will hybrid AI-physics approaches remain the future?**

We invite participants from AI, engineering, and scientific computing to share perspectives on accuracy, trustworthiness, computational efficiency, real-world applications, and the challenges of deploying AI for complex physical systems.

Description

Publishing in AI research is highly competitive, and the path from submission to acceptance is rarely straightforward. Yet many decisions follow consistent patterns that are seldom communicated openly. As Editor-in-Chief of Wiley's Advanced Computing, I will share an honest perspective on what happens behind the scenes - from first submission to final decision, and the use of AI in publishing.

This table explores:
- AI in publishing - from AI-assisted writing to disclosure policies and risks to scientific integrity (https://www.wiley.com/en-gb/publish/article/ai-guidelines/)
- What makes or breaks a paper - structure, framing, and the unwritten rules that separate acceptance from desk rejection
- How editorial decisions really get made - triage, reviewer selection, novelty thresholds, and what editors wish authors knew

Whether you're an early-career researcher, seasoned author, or just curious about how scientific knowledge gets validated, this table is for you.

Description

“Cities are becoming increasingly intelligent through AI — but the real question is whether this intelligence can actually make urban life healthier, fairer, safer, and more human.”

Description

Two patients receive the same treatment, yet only one improves. Current AI models often predict outcomes but rarely explain how treatment decisions should be made under uncertainty. This discussion asks whether healthcare AI should move beyond prediction toward individualized decision-making by integrating clinical evidence, molecular biology, and Bayesian decision theory.

The session introduces integrated Predicted Individual Treatment Effects (iPITE): a framework combining clinical trials, genomic information, and uncertainty-aware models to estimate patient-level treatment response and optimize treatment strategies.

Core Research Questions

- How should AI transform uncertain patient-level predictions into clinically actionable treatment decisions?
- How can molecular evidence (genomics, transcriptomics, epigenomics) improve individualized treatment recommendations rather than function as auxiliary predictors?
- How should uncertainty from high-dimensional models be propagated into treatment recommendations?
- How do we validate treatment recommendations when predictive accuracy does not imply clinical usefulness?
- What defines decision validity in AI-assisted medicine?

Central Question for Debate
A model that predicts well does not necessarily recommend well. Should healthcare AI be evaluated by predictive accuracy—or by the quality of treatment decisions it enables?