NeurIPS 2026 workshop · Paris
Self-supervision, transfer learning, sampling, and tokenization for scientific data and physical systems.
Representations map observations into structured embeddings whose geometric structure reveals semantic content. In vision and language, deep learning models trained on large unlabeled datasets have produced representations that transfer broadly across tasks.
As AI progressively permeates scientific research, large-scale datasets together with high-throughput simulation and experimental pipelines, are making it possible to learn general-purpose scientific representations at unprecedented scale.
The scientific setting, however, challenges representation learning in fundamental ways. Data are heterogeneous and structured, and often come as unlabeled streams from a dynamical system. Moreover, the learned embeddings must respect conservation laws, geometry, and causal structure. Ultimately, a useful scientific representation should expose structures that scientists can act upon.
This workshop provides a tightly scoped forum for representation learning in physical systems. For this first edition, we welcome discussions and contributions on topics around Self-Supervision, Transfer Learning, Sampling, and Tokenization, which we believe constitute the most interesting open problems in this space.
Scientific data are often abundant but unlabeled. Self-supervised learning offers a natural route to extracting structure from such data, but standard pretext tasks and augmentations can violate scientific meaning. What self-supervised objectives preserve physical constraints such as conservation laws? What semantic structures do SSL objectives discover when constrained by physical priors?
learning with no labelsScientific models must reliably extrapolate into physically meaningful regimes beyond their training distribution. In the physical sciences, however, verifying a model's out-of-distribution prediction often requires massive computational effort or expensive wet-lab synthesis. This theme investigates the opportunities and limits of scientific transfer to ensure that learned embeddings remain falsifiable and actionable rather than just empirically successful.
generalizationUnlike internet-scale text and image corpora, many scientific domains have access to simulators and experimental loops that can generate new data. Molecular dynamics, DFT, CFD, PDE solvers, and high-throughput experimental platforms make it possible to shape the training distribution itself. This makes adaptive simulation and closed-loop data generation a central opportunity for building better scientific representations.
adaptive data generationThe success of foundation models relies heavily on discrete tokenization, but mapping physical sciences into discrete vocabularies remains a fundamental bottleneck. Physical data are inherently continuous, multi-scale, and often non-Euclidean. Naive grid-based patching or quantization can destroy geometric priors, symmetries, and sub-grid dynamics. This theme focuses on novel tokenization methods for physical systems.
physical modalities
Technical University of Munich
Deep learning, partial differential equations, and fluid simulation
Bioptimus
Machine learning for biology, genomics, and precision medicine
University of Cambridge
Statistical ML, uncertainty quantification, and engineering
ISTA
Causal representation learning, causal discovery, and AI for science
Flatiron Institute
Atomistic systems and physically constrained machine learningWorkshop contributions
Explore the work presented at the workshop, with abstracts, presentation schedules, and materials shared by the authors.
Browse accepted papersItalian Institute of Technology
Researcher bridging physics and AI, with a focus on scientific machine learning for dynamical systems and atomistic simulations.
Italian Institute of Technology · UCL
Professor and researcher working on the foundations of machine learning and their applications to scientific modeling.
Télécom Paris · IP Paris
Professor specializing in statistical machine learning for non-Euclidean data, trustworthy AI, and sustainable AI.
Columbia University
Professor working at the intersection of climate science, hydrology, and physics-aware AI for Earth system modeling.
Columbia University
Research scientist developing physics-informed machine learning parameterizations for sub-grid processes in climate models.
University of Stuttgart
Professor researching structured and scientific data through geometric, physics-aware, probabilistic, and generative machine learning.
