Physics-Informed Neural Networks Framework for High-Throughput Multi-Phase-Field Microstructure Simulation

• Applicant: Professor Dr. Reza Darvishi Kamachali
• Subject Area: Computer-Aided Design of Materials and Simulation of Materials Behaviour from Atomic to Microscopic Scale ; Thermodynamics and Kinetics as well as Properties of Phases and Microstructure of Materials
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 552892608

Today's computational techniques have revolutionized the way we conduct materials research. On the mesoscale, the generalized multi-phase-filed (MPF) modelling and simulation have grown to a versatile tool with huge capacity to capture intricate microstructural evolution, with high accuracy and efficiency, in multi-phase, multi-component materials. While MPF provides a tremendous flexibility is formulating the free energy functionals adopted to specific problems, this comes with the drawback that any additional feature to the model changes the underlying free energy functional and the corresponding equations of motion, huge re-programming and re-optimization efforts are involved in each stage of development, often down to the deep levels of redefining key parameters and functions, even modifying the memory/storage structures in the software. As a result of this drawback, MPF simulation packages often contain a large body of codes and develop at a much slower pace (thus being much less popular) than their atomistic compartments, such as ab-initio and molecular dynamics simulation toolkits in the community. In this project, we propose to produce a physics-informed neural network (PINNs) framework and software that allow easier adoption of the MPF framework from an existing problem to a new one. Through significant preliminary work, we show that such solution software is feasible. Within the current proposal, we aim to develop the new PINNs-MPF capable of performing advanced microstructure simulations. We draw a delicate plan to develop an attentive mechanism combined with several major machine learning algorithms to efficiently trace the diffuse interface in the MPF context, while massively parallel software structure will be designed to carry the simulations in space and time. Benchmark applications in grain growth, precipitation and secondary-phase particle drag are designated. When successful, the novel software provides a complementary solution at the interface of machine learning and microstructure modelling.

• DFG Programme: Research Grants