Active Matter JEPA

Project overview

Built JEPA-style self-supervised learning models for active matter simulation videos. The goal was to learn useful latent representations by predicting future latent states from context windows, without using physical parameter labels during representation learning.

Motivation

Active matter systems contain rich spatiotemporal structure. Instead of training directly on labeled physical parameters, this project studied whether self-supervised latent prediction could learn representations that preserve information useful for downstream physical inference.

Approach

Compared global, spatial, and hierarchical/multiscale representation-learning designs under a controlled evaluation setup. Representations were frozen and evaluated using simple probes, including linear regression and kNN, for physical parameters such as alpha and zeta.

Dataset

Simulation video sequences from active matter dynamics with held-out parameter regimes for downstream evaluation.

Model design

Context-target JEPA-style latent prediction with comparative global and spatial encoders, plus multiscale hierarchy variants.

Evaluation

Frozen-representation probing via linear regression and kNN on downstream physical parameter inference tasks.

Results / takeaways

Spatial and multiscale representations were especially important for preserving local tensor-field structure, while global representations were more aligned with broad velocity-like features.

Links

Report (placeholder URL) · Slides (placeholder URL) · Code (placeholder URL)