Late Fusion Neural Operator Architecture Improves PDE Extrapolation Across Parameter Space

A new neural operator architecture called Late Fusion Neural Operator has been introduced to address the challenge of predicting system behavior governed by partial differential equations (PDEs) across unseen parameter regimes. This approach specifically tackles distribution shifts that occur when physical parameters vary between training and prediction scenarios. The architecture disentangles learning state dynamics from parameter effects, combining neural operators for latent state representations with sparse regression to incorporate parameter information. This design improves predictive performance both within and beyond the training distribution. The work focuses on robust generalization for scientific and engineering applications where parameter variations induce significant distribution shifts. The method's key innovation lies in how parameters are incorporated into neural operator models, particularly when state and parameter representations become entangled. The research was announced on arXiv with identifier 2604.16721v1 under a cross announcement type.