Courant: a State-Adaptive Perceiver-Based Neural Surrogate with Local Support and Interpretable Field Decomposition
Researchers have introduced "Courant," a novel neural surrogate model based on the Perceiver architecture. This model features latent features that adapt to specific physical spaces, mimicking the adaptive refinement seen in traditional numerical solvers. Courant is trained end-to-end using simulation data and achieves competitive accuracy with a standard L2 prediction loss, offering interpretable latent representations that decompose simulated fields. AI
IMPACT Introduces a new neural architecture for scientific machine learning that offers interpretable field decomposition.