New Method Accelerates Physics-Constrained Generative AI

Generative models are increasingly used as scalable alternatives to physical simulations, but they often lack guarantees that their outputs adhere to fundamental physical laws, such as conservation principles or boundary conditions. Constrained sampling techniques can enforce these constraints during inference without requiring model retraining, but they incur significant computational costs, especially for nonlinear constraints, due to repeated projection, correction, and trajectory optimization steps. Standard machine learning frameworks exacerbate this issue because their reliance on dense tensor algebra and limited support for sparse solver composability obscure the inherent structure of physical constraints. This makes efficient batched nonlinear optimization challenging to implement. The proposed solution, SNAP-FM (Sparse Nonlinear Accelerated Projection for Physics-Constrained Generative Modeling), tackles this bottleneck by exploiting the block-sparse Jacobian and KKT systems that arise from sample-wise batching and local PDE couplings. It exposes this structure using ExaModels.jl and solves the resulting sparse nonlinear programs with MadNLP.jl and GPU sparse factorization. Applied to Physics-Constrained Flow Matching (PCFM) on various PDE benchmarks, SNAP-FM significantly accelerates nonlinear constraint projection while maintaining precise constraint satisfaction, demonstrating that sparse GPU nonlinear optimization is a practical foundation for scientific machine learning.