PINNs-MPF: A Physics-Informed Neural Network Framework for Multi-Phase-Field Simulation of Interface Dynamics

TL;DR

This paper introduces PINNs-MPF, a neural network framework that combines physics-informed learning and multi-network strategies to simulate complex multi-phase interface dynamics with high accuracy.

Contribution

The paper develops a novel PINNs-based framework with multi-networking and dynamic meshing for efficient multi-phase-field microstructure evolution simulations.

Findings

Successfully reproduces benchmark tests with high fidelity.

Achieves low Mean Squared Error loss values between 10^{-4} and 10^{-6}.

Effectively handles complex geometries and interface dynamics.

Abstract

We present an application of Physics-Informed Neural Networks to handle MultiPhase-Field simulations of microstructure evolution. It has been showcased that a combination of optimization techniques extended and adapted from the PINNs literature, and the introduction of specific techniques inspired by the MPF Method background, is required. The numerical resolution is realized through a multi-variable time-series problem by using fully discrete resolution. Within each interval, space, time, and phases are treated separately, constituting discrete subdomains. An extended multi-networking concept is implemented to subdivide the simulation domain into multiple batches, with each batch associated with an independent Neural Network trained to predict the solution. To ensure efficient interaction across different phasesand in the spatio-temporal-phasic subdomain, a Master NN handles efficient interaction among the multiple networks, as well as the transfer of learning in different directions. A set of systematic simulations with increasing complexity was performed, that benchmarks various critical aspects of MPF simulations, including different geometries, types of interface dynamics and the evolution of an interfacial triple junction. A comprehensive approach is adopted to specifically focus the attention on the interfacial regions through an automatic and dynamic meshing process, significantly simplifying the tuning of hyper-parameters and serving as a fundamental key for addressing MPF problems using Machine Learning. The pyramidal training approach is proposed to the PINN community as a dual-impact method: it facilitates the initialization of training and allows an extended transfer of learning. The proposed PINNs-MPF framework successfully reproduces benchmark tests with high fidelity and Mean Squared Error loss values ranging from 10$^{-4}$ to 10$^{-6}$ compared to ground truth solutions.