Conference Agenda

Laser–plasma acceleration experiments require non-invasive, single-shot diagnostics of rapidly evolving plasma targets, including gas jets and solid targets such as wires. Propagation-based X-ray phase-contrast imaging (in-line holographic geometry) provides penetration and phase sensitivity, but detectors record only intensities. Quantitative interpretation therefore relies on accurate wave-optics forward models and robust phase-retrieval algorithms matched to the imaging regime and material response.

We present a unified PyTorch framework for near-field X-ray wave propagation and phase retrieval, enabling reconstruction of the object’s exit phase (and absorption) from measured intensities using automatic differentiation. The forward model represents the object by a complex transmission and propagates the exit wave to the detector using differentiable operators (Fresnel diffraction and angular spectrum variants). The framework standardizes interfaces for synthetic data generation, forward-model selection, and reconstruction, enabling reproducible benchmarking across classical, gradient-based, and self-supervised physics-informed approaches under a consistent forward model and loss definition.

As a representative use case, we demonstrate the workflow on hydrogen gas jets and wire targets using compound refractive lens (CRL) geometries, including alternative lens configurations. We compare a classical single-distance baseline to a differentiable gradient-descent refinement initialized from that baseline that minimizes an intensity data fidelity term with physics-based regularization, remaining informative in challenging regimes (e.g. defocus sign ambiguities) where classical reconstructions degrade. Recovered exit-phase maps can be converted to projected areal density under standard refractive-index assumptions, providing a pathway toward quantitative single-shot plasma target diagnostics and a common testbed for physics-informed machine learning in experimental diffraction imaging.

Remote-sensing super-resolution promises enhanced spatial detail, but its reliability is fundamentally limited by the ambiguity of the underlying ill-posed inverse problem. We address this challenge by introducing computable performance bounds that characterize the achievable reconstruction accuracy independently of any specific reconstruction method and solely from the problem’s forward operator and the data (Gottschling et al., 2025). These bounds are derived from the null space of the forward operator, i.e., from the image content that is lost under the downsampling process.

We demonstrate the practical value of these bounds through a comprehensive evaluation across multiple quality measures, land-cover types, degradation settings, and super-resolution factors. We consider spatial and spectral metrics to analyze how theoretical performance limits depend on the evaluation criterion, scene characteristics, and degradation model, and how existing super-resolution methods perform relative to the derived lower and upper bounds. A particular contribution of this work is a systematic analysis of the relationship between zoom level and hallucinations. We show how the bounds evolve over a wider range of magnification factors and examine how increasing zoom affects both the frequency and spatial extent of hallucinated content.

We further compare the proposed bounds with uncertainty estimates from pretrained neural super-resolution models (Donike et al., 2025), allowing us to assess whether these uncertainty maps capture the task's fundamental ambiguity or primarily reflect model-specific effects. In addition, we study out-of-distribution samples and imagery from extreme events to better understand model behavior under distribution shift. In this setting, we investigate whether kernel sizes can help relate such cases to known in-distribution land-cover classes and thereby support a more principled interpretation of hallucination risk.

Overall, the evaluation highlights that, as reconstruction methods approach the theoretical performance range, the dominant factors shaping achievable accuracy shift from model choice to scene content, zoom level, and the intrinsic ambiguity of the inverse problem. This positions the proposed framework as a rigorous basis for analyzing reliability, uncertainty, and hallucination risk in super-resolution, with clear relevance for safe and trustworthy GeoAI.

As most perturbations affect the whole body on a molecular scale, different histological techniques have been developed with the intention to capture either whole-body, cellular, molecular changes, or a combination of [1]. However, as combining multiple histological techniques into a single protocol is difficult and animal experiments are limited by their time duration and ethics, biomedical in silico experimentation (BioXPT) by integration of different data modalities with artificial intelligence has gained significant interest. Specifically focused on the mouse brain as a first step, the goal of the BioXPT-Brain project is to create a foundation model that integrates brain morphology in 3D with molecular gene expression. By utilizing our large and unique multimodal dataset obtained via tissue clearing and using contrastive learning to learn the relation between anatomical structures (such as vasculature) and gene expression, BioXPT-Brain can be used to predict molecular information from 3D imaging. Additionally, it can be used to study perturbations such as aging by exploring the microenvironment of structurally abnormal and leaky vessels [2].

[1] Ertürk, A. Deep 3D histology powered by tissue clearing, omics and AI. Nat. Methods 21, 1153–1165 (2024).

[2] Todorov, M. I. et al. Reversible hypervascularization drives cognitive decline and blood-brain barrier damage during aging. Currently in revision (2026).

Single-cell genomics has transformed our ability to measure biological processes on the cellular level. Yet, many biological processes, such as disease and aging, manifest at the system level. Sample representation methods have recently emerged to bridge these scales, enabling understanding of system-level processes from single-cell genomics data. However, existing methods can only be trained to represent a single biological signal, ignoring the complexity of inter-individual variation. Moreover, no current method has the ability to overcome batch effects between datasets to learn patient variation across studies, a critical challenge for single-cell studies that often only include a few donors.

Here, we present SampleCLR, the first batch-aware sample representation method that learns patient variation across single-cell studies. SampleCLR uses a contrastive multiple-instance learning framework, leveraging the assumption that randomly sampled cell subsets from the same sample represent the same underlying object, used to construct positive pairs in representation space. To mitigate dataset-specific batch effects, SampleCLR leverages intra-dataset hard negative sampling, prioritizing samples within the same dataset as negatives, thereby encouraging learning representations that are invariant to technical variation. Our model is trained via self-supervision, with an optional multi-task supervision module that allows the representations to be jointly optimized for multiple clinical covariates of interest. We have evaluated SampleCLR using published benchmarking frameworks and case studies on atlas-level datasets, demonstrating that SampleCLR excels at preserving biological information and removing batch effects from sample representations. Specifically, we show that SampleCLR can identify disease trajectories from blood samples of COVID-19 patients and use the learned representations to predict disease severity of samples from unseen batches, prototyping clinical applicability of reference atlases. Our analysis spans blood, gut, and pancreas atlas-scale datasets to identify age-, sex-, and BMI-specific effects at the sample level. We demonstrate that SampleCLR embeddings of patient samples are interpretable through cell attention, enabling researchers to link sample-level findings to cell-type proportions and gene expression. Our work enables sample-level analysis of reference atlases and paves the way to translational applications of single-cell datasets.

Large-scale medical biobanks provide imaging data complemented by extensive tabular information, such as clinical measurements or demographics. However, this abundance of tabular attributes does not reflect real-world datasets, where only a subset of attributes may be available. With limited tabular information, existing vision–tabular models often fail to outperform unimodal image-only models when both modalities are present, but tabular data contains many missing attributes (Fig. 1). In addition, prior work does not address the scenario where tabular data is entirely missing at inference time. This highlights the need for methods that remain robust across all levels of tabular data availability.

To address this problem, we propose RoVTL (Robust Vision–Tabular Learning), a framework designed to operate across the full range of tabular data availability, from no tabular attributes being available to full tabular information. RoVTL consists of two stages. First, during contrastive pretraining, we simulate tabular missingness as a data augmentation strategy to improve robustness. Second, during downstream fine-tuning, we introduce the Tabular More vs. Fewer (TabMoFe) loss. TabMoFe compares two nested subsets of the tabular data, where one subset is contained within the other, and ensures that the model maintains or improves performance as more tabular attributes are added. Combined with a gated cross-attention fusion module, this training strategy enables consistent performance across different levels of tabular completeness. An overview of the method is shown in Fig. 2.

We evaluate RoVTL on cardiac MRI data from the UK Biobank and on the car advertisement dataset DVM. Our method demonstrates improved robustness to missing tabular data compared to existing approaches on both datasets, for both classification and regression tasks (Fig. 3). In addition, RoVTL generalizes to an external cardiac MRI dataset for multimodal disease classification, highlighting its potential for vision-tabular foundation models.

Artificial intelligence is rapidly transforming scientific discovery across domains such as climate science, health, and physics. However, current AI systems still face two fundamental challenges: limited reliability and high energy consumption.
In this talk, we discuss recent advances toward a mathematical foundation of reliable AI, focusing on generalization, robustness, and explainability in modern learning systems. These aspects are crucial for deploying AI in high-stakes scientific settings, where trustworthiness and interpretability are essential.
At the same time, we address the growing need for sustainable AI. Current large-scale models rely on energy-intensive digital computation, motivating the exploration of alternative paradigms. In this context, we highlight neuromorphic computing and spiking neural networks as promising directions toward more energy-efficient AI.
Bringing these perspectives together, we outline how next-generation AI systems can enable more reliable and scalable approaches to scientific discovery.