Federated Physics-Informed Neural Networks for Distributed Multiphysics Simulations

Abstract

Multiphysics simulations play a key role in modeling multiphysical systems with coupled physical processes in the real-world; but traditional more centralized methods are typically computationally expensive and limited by data-sharing barriers in crossing distributed systems. To resolve these issues, this paper suggests a new Federated Physics-Informed Neural Network (F-PINN) system that combines the concepts of federated learning with physics-informed neural networks to make it possible to perform distributed simulations in a scaled and privacy-preserving way. The suggested method enables joint training of PINNs on a set of decentralised clients without exchanging the raw information, thus providing the guarantee of data locality and confidentiality. Moreover, an effective Multiphysics coupling approach is presented to faithfully represent interactions between different physical domains in the federated context. The model takes into account adaptive optimization methods to improve the stability of convergence and reduce the heterogeneity of clients. The experimental analysis of representative Multiphysics scenarios shows that the suggested F-PINN model is more accurate in terms of a reduced prediction error and better physics consistency, and is much less communication-demanding than the centralized and the baseline federated models. The framework is also highly scaled to larger numbers of distributed clients. Such findings demonstrate the prospects of federated physics-informed learning as an effective and powerful paradigm of next-generation distributed scientific computing and Multiphysics simulations at scale.