Opening lecture of the 5th International Conference on Thermal Process Modeling and Simulation

Abstract

Physics-informed neural networks (PINNs) have become increasingly popular in several engineering domains due to their efficiency in addressing real-world problems characterized by noisy data and occasionally incomplete physics information. In the realm of PINNs, automatic differentiation plays a pivotal role in evaluating differential operators without succumbing to discretization errors. Furthermore, a multi-task learning approach is defined to concurrently accommodate observed data fitting and adherence to the fundamental governing principles of physics.

In this context, we present the application of PINNs to heat treatment operations, with a specific focus on handling realistic conditions that are challenging for traditional computational methods. The estimation of unknown thermal boundary conditions on surfaces is facilitated by predicting them through sparse temperature measurements. We illustrate the applicability of PINNs through several industrial applications related to quenching, underscoring the effectiveness of neural networks in tackling complex heat transfer problems encountered in industrial settings.