New Framework Enhances Multi-Phase-Field Simulations Using Neural Networks
A new framework utilizing Physics-Informed Neural Networks (PINNs) has been developed for simulating multi-phase-field dynamics, as detailed in the paper titled "PINNs-MPF: A Physics-Informed Neural Network Framework for Multi-Phase-Field Simulation of Interface Dynamics" by Seifallah Elfetni and Reza Darvishi Kamachali. This innovative approach addresses the complexities of microstructure evolution in materials science.
The authors explain that the framework combines optimization techniques from existing PINNs literature with methods inspired by the Multi-Phase-Field (MPF) approach. The simulation process is structured around a multi-variable time-series problem, treating space, time, and phases as separate discrete subdomains. This allows for a more organized and efficient simulation of interface dynamics.
A significant feature of the framework is the implementation of a multi-networking concept, which divides the simulation domain into multiple batches. Each batch is associated with an independent neural network trained to predict solutions, while a Master Neural Network facilitates interactions among these networks and manages the transfer of learning.
The researchers conducted systematic simulations that benchmark various critical aspects of MPF simulations, including different geometries and types of interface dynamics. They focused on interfacial regions through an automatic and dynamic meshing process, which simplifies the tuning of hyper-parameters and is crucial for addressing MPF problems using machine learning.
The results indicate that the proposed PINNs-MPF framework successfully reproduces benchmark tests with high fidelity, achieving Mean Squared Error loss values ranging from 10-4 to 10-6 compared to ground truth solutions. This advancement could significantly impact the field of materials science by enhancing the accuracy and efficiency of simulations related to microstructure evolution.
The full paper can be accessed at arXiv with the identifier arXiv:2407.02230.