A Workflow for Utilizing OpenFOAM Data Structure in Physics-Informed Deep Learning Training

Abstract: This study presents a novel methodology for integrating physics-informed loss functions into deep learning models using OpenFOAM's comprehensive data structures. Leveraging the robust and flexible capabilities of OpenFOAM's data structure for handling complex geometries and boundary conditions, it is demonstrated how to construct detailed loss functions that accurately embed physics constraints and potentially enhance the training and performance of neural networks in handling industrial-level complicated geometry for computational fluid dynamics (CFD) simulations. The present work primarily focuses on the 1D Burger equation to showcase the detailed procedure of constructing initial loss, boundary loss, and residual loss. While the computational geometry employed here is relatively simple, the procedure is sufficiently general to illustrate its applicability to more complex computational domains. The results show the trained operator former (OFormer) neural network can successfully predict the simulation results subject to the OpenFOAM'data structure composed loss. This framework potentially opens new avenues for using deep learning to tackle complex industrial simulation challenges, promising significant advancements in the accuracy and practicality of CFD applications.