Paper ID: 2409.13644

Non-overlapping, Schwarz-type Domain Decomposition Method for Physics and Equality Constrained Artificial Neural Networks

Qifeng Hu, Shamsulhaq Basir, Inanc Senocak

We introduce a non-overlapping, Schwarz-type domain decomposition method employing a generalized interface condition, tailored for physics-informed machine learning of partial differential equations (PDEs) in both forward and inverse scenarios. Our method utilizes physics and equality constrained artificial neural networks (PECANN) in each subdomain. Diverging from the original PECANN method, which uses initial and boundary conditions to constrain the PDEs alone, our method jointly employs both the boundary conditions and PDEs to constrain a specially formulated generalized interface loss function for each subdomain. This modification enhances the learning of subdomain-specific interface parameters, while delaying information exchange between neighboring subdomains, and thereby significantly reduces communication overhead. By utilizing an augmented Lagrangian method with a conditionally adaptive update strategy, the constrained optimization problem in each subdomain is transformed into a dual unconstrained problem. This approach enables neural network training without the need for ad-hoc tuning of model parameters. We demonstrate the generalization ability and robust parallel performance of our method across a range of forward and inverse problems, with solid parallel scaling performance up to 32 processes using the Message Passing Interface model. A key strength of our approach is its capability to solve both Laplace's and Helmholtz equations with multi-scale solutions within a unified framework, highlighting its broad applicability and efficiency.

Submitted: Sep 20, 2024