Enhancing Thermal Solutions with Smart Sensors using Physics-Informed Neural Networks

In recent studies, the potential of Physics-Informed Neural Networks (PINNs) has been highlighted as an alternative to conventional mesh-dependent partial differential equation (PDE) solvers for solving complex thermal-fluid problems with intricate domains and boundary conditions. However, the high computational cost associated with PINNs currently limits their practical use in engineering applications. To unlock the full potential of PINNs for industry, this work aims to develop an efficient and accurate PINN-based model by incorporating sensors into the network. By strategically placing a small number of sensors and employing active sensor placement techniques, a significant acceleration in convergence and improved accuracy of PINNs in solving flow and thermal problems can be achieved. This study investigates the achievable speedup and provides guidelines for sensor placement. Furthermore, the methodology developed here can be extended to cooling strategies, including natural and forced heat convection, particularly in the near-wall or surface region of heat sources. MATLAB wide suite of deep learning tools provide an easy to adapt environment for this development.