Physics-Informed Deep-Learning for Scientific Computing

Physics-Informed Neural Networks (PINN) are neural networks that encode the problem governing equations, such as Partial Differential Equations (PDE), as a part of the neural network training. PINNs have emerged as an essential tool to solve various challenging problems, such as computing linear and non-linear PDEs, completing data assimilation and uncertainty quantification tasks. In this work, we focus on evaluating the PINN potential to replace or accelerate traditional approaches for solving linear systems. We solve the Poisson equation, one of the most critical and computational-intensive tasks in scientific computing, with different source terms. We test and evaluate PINN performance under different configurations (depth, activation functions, input data set distribution, and transfer learning impact). We show how to integrate PINN with traditional scientific computing approaches, such as multigrid and Gauss-Seidel methods. While the accuracy and computational performance is still a limiting factor for the direct use of PINN for solving, hybrid strategies are a viable option for the development of a new class of linear solvers combining emerging deep-learning and traditional scientific computing approaches.