Full-field temperature recovery during water quenching processes via physics-informed machine learning

Water quenching is an effective heat treatment process to produce high-quality metallic structures. Accurate and efficient prediction of the full-field temperature inside the part to capture and control the residual stresses and part quality remains a challenging task. This paper proposes a simple and easy-to-use model for full-field temperature recovery during water quenching processes, using physics-informed machine learning (ML). The novelty of the ML framework is that it only needs temperature measurements of sparse locations to efficiently/accurately recover the full spatio-temporal temperature field without invoking sophisticated multiphysics simulations. The ML framework consists of two tightly connected neural network (NN) models: (1) Firstly, a physics-informed neural network (PINN)-based surrogate model is constructed. The surrogate model, which approximates a high-fidelity finite element model, is responsible for quickly outputting the full-field temperature distribution from the parameterized thermal boundary conditions (BCs). (2) Then, another neural network is constructed to project the available experimental data onto the surrogate model and learn the optimal thermal BC from the parametric space, which produces the best full-field temperature prediction in the surrogate model. The proposed ML framework features high efficiency, accuracy, and universality for temperature prediction in quenching processes. These features are carefully demonstrated and the framework is validated using experimental measurements.