One-Way Coupling via Patchwise Adaptive Normalization in Transformer Surrogates
PAdaNorm improves prediction accuracy for one-way coupled fields in physical PDEs by using patchwise conditioning. It partitions exogenous fields into spatial patches, maps each patch locally via an MLP to create a conditioning vector, and uses these vectors to modulate the transformer's representations. This method achieves comparable or better accuracy than baselines while significantly reducing computation.