When Does Physical Representation Alignment Help Crystal Generation?

摘要

Pretrained interatomic potentials offer energy- and force-trained representations for crystal generators whose structural objectives provide no direct energetic supervision. Although representation alignment provides a route for transferring this knowledge, its benefit may depend on both the generator's existing priors and the target materials. We use DiffCSP and DiffCSP++ to formulate a controlled study of these dependencies. Preliminary comparisons suggest a larger gain from masked atom prediction (MAP) for DiffCSP on MP-20 than for DiffCSP++ on the same dataset. Carbon-24 and Perov-5 observations instead include limited gains or degradation. These patterns motivate three questions: whether transferred information overlaps with explicit symmetry priors, whether the teacher supplies informative targets for the materials domain, and how masking changes the information used for prediction. To examine the third question, MAP-DiffCSP predicts frozen MACE descriptors through retained atom states after latent masking. With the backbone fixed, we characterize the population-optimal shared prediction under squared and cosine losses as a context-conditioned summary of masked environments. This characterization separates teacher-target predictability from downstream usefulness and motivates controls for contextual transfer, architectural overlap, and teacher-domain mismatch. The study treats heterogeneous outcomes as a route to understanding when physical alignment contributes information that crystal generation can use.