PINN-IT: A Python Framework for Automatically Creating Neural Network Based Solvers

Industry 4.0 states that manufacturing practices can achieve transformational advances via the use of artificial intelligence (AI). A key driver of Industry 4.0 will be AI-based solvers. A solver is a program that simulates physical phenomena, including heat, fluid flow, mechanical stress, and electromagnetics. Solvers are critical for ensuring that manufacturing processes are accurate, efficient, and safe. However, traditional math-based solvers can be slow, easily taking hours to perform a single simulation. To achieve Industry 4.0, we need solvers that can perform near real-time simulations. In this research, we created PINN-IT, a Python script that creates 2D and 3D solvers using AI rather than traditional mathematical methods. PINN-IT trains AIs to understand partial differential equations (PDEs) using hybrid learning, which combines supervised learning via a collocation dataset with unsupervised learning via a custom loss function that minimizes the residual of the PDE. As examples, PINN-IT solved the Turing pattern reaction-diffusion model, an advection diffusion model, a transient heating model, and a phase field model 64 times, 2041 times, 28740 times, and 6 times faster than a traditional solver, respectively. PINN-IT will help achieve Industry 4.0 by enabling near real-time simulations.