Hybrid Neural Network Models for the Optimization of Induction Hardening Processes

Abstract

We describe a simple hybrid methodology to simulate an induction heating process that combines observational (black-box) and physics-based (white-box) methodologies. This method uses a neural network to predict the process' physical characteristics, which were previously unknown. A primary emphasis is placed on monitoring temperature variations within a subsurface layer of a bolt sample. The hybrid model incorporates an ordinary differential equation for the heating rate, leading to improved data accuracy compared to a standalone black-box model. This innovative approach not only improves predictive precision but also simplifies interpretability, ultimately serving as a pivotal instrument for the effective management and advancement of induction heating operations.