Abstract
Physics-informed neural networks (PINNs) are a new tool for solving boundaryvalue problems by defining loss functions of neural networks based on governingequations, boundary conditions, and initial conditions. Recent investigationshave shown that when designing loss functions for many engineering problems,using first-order derivatives and combining equations from both strong and weakforms can lead to much better accuracy, especially when there are heterogeneityand variable jumps in the domain. This new approach is called the mixedformulation for PINNs, which takes ideas from the mixed finite element method.In this method, the PDE is reformulated as a system of equations where theprimary unknowns are the fluxes or gradients of the solution, and the secondaryunknowns are the solution itself. In this work, we propose applying the mixedformulation to solve multi-physical problems, specifically a stationarythermo-mechanically coupled system of equations. Additionally, we discuss bothsequential and fully coupled unsupervised training and compare their accuracyand computational cost. To improve the accuracy of the network, we incorporatehard boundary constraints to ensure valid predictions. We then investigate howdifferent optimizers and architectures affect accuracy and efficiency. Finally,we introduce a simple approach for parametric learning that is similar totransfer learning. This approach combines data and physics to address thelimitations of PINNs regarding computational cost and improves the network'sability to predict the response of the system for unseen cases. The outcomes ofthis work will be useful for many other engineering applications where deeplearning is employed on multiple coupled systems of equations for fast andreliable computations.