Integrating artificial neural networks with a classical kinetic model for phase transformation predictions in steels

Abstract

In this work, we present a hybrid modeling approach that integrates a classical kinetic model with neural networks to predict anisothermal metallurgical transformations. By leveraging neural networks to approximate kinetic parameters, we investigate two model architectures. The first architecture models kinetic parameters as functions of temperature and cooling rate alone, while the second extends this dependency to include phase fractions involved in the transformation. Specifically, the model addresses the decomposition of austenite into four distinct phases, each governed by ordinary differential equations. We train the model on continuous cooling data for a single steel. The results demonstrate that our approach not only accurately replicates the experimental data but also offers robust generalization capabilities within a given chemistry. Finally, we discuss the potential to extend this model to incorporate chemical composition effects and to simulate isothermal transformations.