A five-day workshop on learning predictive models of the physical world — from joint-embedding predictive architectures to world models that reason in latent space.
World models and joint-embedding predictive architectures (JEPA) start from a simple premise: intelligence rests on internal models that predict how the world evolves — learned in abstract representation spaces rather than raw pixels or measurements.
Physics has built predictive models of the world from first principles for centuries. This workshop asks what these two traditions can learn from each other. Can self-supervised, joint-embedding methods recover physical structure — symmetries, conservation laws, dynamics — directly from data? Can physical priors make world models more sample-efficient, interpretable, and reliable?
Over five days at the Aspen Center for Physics, roughly seventy researchers from machine learning and the physical sciences work through these questions in focused morning sessions and long, unstructured afternoons built for collaboration.
JEPA and self-supervised representation learning as a route to abstract, predictive features.
Learning state spaces, rollouts, and model-based planning over learned representations.
Conservation laws, equivariance, and geometry as inductive biases for learning.
Large pretrained models for physics, astronomy, and the natural sciences.
What physical structure emerges in learned dynamics, and how to read it.
Energy-based and predictive objectives against generative modeling in latent space.
We invite contributions at the intersection of world modeling, representation learning, and the physical sciences. Submit your abstract through the official workshop form for consideration in the contributed program.
Professor of Mathematics and Atmosphere/Ocean Science at NYU Courant. Her research combines theory, numerical simulation, statistics, and machine learning to understand ocean dynamics, climate predictability, turbulence, heat and carbon uptake, and sea-level rise.
NYU profile ↗
Research Scientist in EIT’s AI & Robotics Institute, working on foundational machine learning for the natural sciences. Previously at Microsoft Research AI for Science, she co-created Aurora, a foundation model for the Earth system, and worked on quantum chemistry and drug discovery.
EIT profile ↗The workshop is hosted at the Aspen Center for Physics, an independent institution that has convened physicists for focused, collaborative research since 1962. Its format — short mornings, open afternoons, no distractions — is the reason the meeting is structured the way it is.
Aspen sits at 2,400 m in the Elk Mountains of Colorado. Late February is deep winter: bring layers. Lodging and travel guidance will be shared with registered participants.
