Coupled Physics Informed Neural Network (coPINN) as Standalone Multi-Physics Solver for Additive Friction Stir Deposition

Abstract

Additive Friction Stir Deposition (AFSD) is a metal additive manufacturing process gaining increased attention due to its solid-state processing nature and the desirable properties it produces. Due to the multi-physics nature of the process, with complex coupled partial differential equations, computational modeling entails inaccuracies and slowness in utilizing numerical methods. This work proposes a coupled Physics Informed Neural Network (co-PINN) as a fast and accurate mesh-free solver for predicting the thermo-mechanical processing history of the AFSD process. Unlike a typical notion of data-driven neural networks, the co-PINN is an NN-based standalone solver for PDE-based models. The mesh-free nature of the method allows for better estimations of the gradient-based derived physical quantities at any given algorithmic step, such as the effective strain rate in the case study, eventually leading to a better solution. Motivated by a one-dimensional case of heat and mass transfer in a rod, the proposed co-PINN accurately predicts the material flow and temperature distribution for a steady-state coupled fluid dynamics and heat transfer model of the AFSD process in the Eulerian reference frame. The co-PINN-based solver has a low residual error and takes a fraction of the time (as low as 35 minutes) than the existing volume-based solver at an appropriate mesh level. Experimental validation of the co-PINN shows an excellent match with in situ thermal data. Further, an inverse calibration of the proposed coPINN model with the experimental data accurately identifies the back heat transfer coefficient.