CG JEPA: polymer representations and physical prediction

Approach

• Built periodic molecular graphs and temporal-learning pipelines for polymer trajectories, with local molecular descriptors and chain-level representations.
• Compared JEPA-style predictors with persistence, linear autoregression, supervised graph models, and random-encoder controls using common physical observables.
• Examined token fields, local packing information, chain pooling, and relative-displacement representations through controlled ablations.
• Separated current-state representation quality from fixed-horizon future prediction, retaining technical-seed comparisons, time-block uncertainty estimates, and negative results.

What the exploration showed

• Local packing information contributed to physical prediction in the tested pipelines; deleting an input field did not necessarily remove the same information from graph connections and other channels.
• Explicit relative-displacement representations improved current chain-geometry readouts and retained useful signals for future physical observables. These improvements came from a geometric interface and current-state physical supervision, rather than an established temporal JEPA gain.
• The tested temporal-learning variants did not surpass strong linear physical baselines in the completed phase. The frozen public coarse-grained polystyrene benchmark also retained a negative result for the tested JEPA predictor at the 800 ps horizon.

This phase sharpened the distinction between information in a molecular input, information retained by an encoder, and information used by a temporal predictor. The encoder comparison motivated the subsequent MACE JEPA line.

Scope & limitations

• The evaluated targets summarize polymer geometry and local environment. Predicting these observables does not demonstrate generation of complete coarse-grained coordinate trajectories.
• The results concern the recorded polymer configurations, data sources, and finite training protocols; they do not establish experimental-polymer generalization or a universal ranking of learning methods.
• Technical seeds and correlated chain windows are not independent molecular systems. Current-state supervised gains are kept separate from temporal self-supervised gains.