Machine Learning Modeling of Hardness–Depth Distribution of Nitrided Steels
TL;DR
A comprehensive database was constructed, comprising over twenty thousand hardness values from steels with various compositions and nitriding parameters, and two predictive models were developed incorporating physically based parameters such as nitrogen diffusion distance and nitriding layer thickness.
Abstract
Nitriding is an essential thermochemical surface treatment that enhances fatigue, wear, and corrosion resistance of steels. The resulting hardness and hardened layer thickness depend strongly on alloy composition and process conditions, yet accurate prediction across multi-component systems remains challenging. In this study, a comprehensive database was constructed, comprising over twenty thousand hardness values from steels with various compositions and nitriding parameters. Several machine learning algorithms were compared, and deep neural networks (DNN) were selected for their superior performance. Two predictive models were developed: one using alloy composition, nitriding temperature, time, and depth as input features, and another incorporating physically based parameters such as nitrogen diffusion distance and nitrided layer thickness. The latter model achieved higher accuracy and broader applicability, successfully reproducing experimental hardness distributions under diverse alloy and process conditions. SHAP (SHapley Additive exPlanations) analysis clarified the contributions of individual alloying elements, indicating that nitride-forming elements such as Cr, V, Al, Ti, and Mo strengthen the surface, while high nitriding temperatures reduce hardness. The model also captured carbon’s suppressive effect on Cr-nitride formation and the influence of nitriding time on layer growth. This hybrid, data-driven and physics-informed framework provides an explainable and versatile tool for designing nitrided steels with optimized surface properties.