The Representations for the Physical Sciences (R/PS) workshop will take place on December 13, 2026 in Paris, held in conjunction with NeurIPS 2026. R/PS provides a tightly scoped forum bringing together machine learning researchers, computational scientists, and physicists to explore how representation learning can unlock breakthroughs across physical systems.
13 DecWorkshop day, 2026
ParisSorbonne University
5Invited speakers
4Core themes
About the workshop
Physical data are inherently continuous, multi-scale, non-Euclidean, and strictly bound by physical laws. While standard machine learning architectures have driven massive successes in vision and language, standard pretext tasks, patching, and loss functions often fail to preserve underlying physical symmetries, conservation principles, and sub-grid dynamics.
For its first edition, the R/PS workshop addresses the foundational representation learning bottleneck in physical systems. We focus on four key interconnected themes: self-supervision under physical constraints, out-of-distribution transfer to actionable regimes, adaptive simulation sampling, and discrete tokenization of continuous physical modalities.
Core themes
Self-Supervision
Learning with no labels
Extracting physical structures from abundant unlabeled data while ensuring SSL pretext objectives preserve conservation laws and physical constraints.
Transfer Learning
Falsifiable generalization
Ensuring model predictions extrapolate reliably into physically meaningful OOD regimes where wet-lab or high-fidelity validation is expensive.
Sampling
Adaptive data generation
Leveraging simulators (MD, DFT, CFD, PDE solvers) and experimental loops to shape the training distribution via adaptive sampling.
Tokenization
Physical modalities
Designing novel tokenization schemes for continuous, multi-scale, and geometric data without destroying sub-grid dynamics or symmetries.
Topics of interest
Self-supervised objectives preserving physical constraints & conservation laws
Out-of-distribution extrapolation & falsifiable embedding evaluation
Adaptive simulation & closed-loop data generation with PDE/DFT solvers
Geometric, continuous, and multi-scale tokenization for physical data
Research Notes on open problems, benchmarks, paradigms, and datasets
Physics-informed foundation models and geometric representations
Invited speakers
Nils Thuerey
Technical University of Munich
Researches deep learning, partial differential equations, and fluid dynamics simulations.
Jean-Philippe Vert
Bioptimus
Specializes in machine learning for biology, genomics, and precision medicine.
Mark Girolami
University of Cambridge
Focuses on statistical ML, uncertainty quantification, and engineering applications.
Francesco Locatello
ISTA
Works on causal representation learning, causal discovery, and AI for scientific discovery.
Jigyasa Nigam
Flatiron Institute
Researches atomistic systems and physically constrained machine learning.
Program & Format
The workshop is a single-day in-person event featuring invited keynotes, oral presentations of accepted short papers, interactive poster sessions, and community discussions.
KeynotesInvited talks by leading researchers at the intersection of physics, AI, and scientific computing.
Contributed Talks & PostersOral and poster presentations selected from peer-reviewed short paper submissions.
Discussion & NetworkingStructured time for interdisciplinary exchange, open discussions, and collaboration.
Important dates
July 29, 2026 (AoE)Call for papers opens
August 29, 2026 (AoE)Paper submission deadline
September 29, 2026 (AoE)Author notification
December 13, 2026Workshop day
Submission information
Short Papers
Up to 4 pages (excluding references and appendices) for mature or ongoing research.
Research Notes
Accessible explanations of open problems, paradigms, datasets, or algorithms. Clarity & usefulness emphasized.
Submissions must consist of a single anonymized PDF following the NeurIPS LaTeX format. Accepted papers are non-archival and will be presented via poster or oral presentation.
Venue & Location
The workshop takes place at Room 108, Campus Pierre et Marie Curie, Sorbonne University (4 Pl. Jussieu, 75005 Paris, France) as part of the NeurIPS 2026 Paris events.
Organizers
Pietro NovelliItalian Institute of Technology
Massimiliano PontilIIT / UCL
Florence d'Alché-BucTélécom Paris / IP Paris
Pierre GentineColumbia University
Kara LambColumbia University
Mathias NiepertUniversity of Stuttgart
