Prediction of Stability of Austenitic Doping Systems Based on First-Principles Calculation and Machine Learning

X.F. Jiao, S.H. Wang, 3, Y.B. Song, R. Xie, K. Liu, and G.L. Ni

DOI: https://doi.org/10.64486/m.66.1.1
Online publication date: May 15, 2026

Abstract: This study presents a novel methodology that combines machine learning with first-principles calculations to efficiently screen austenitic stabilizing elements. Correlation analysis identified the lattice constant as a critical factor influencing the stability of austenitic doping systems. Subsequently, the random forest, support vector regression, and AdaBoost models were evaluated, among which the random forest achieved the highest prediction accuracy (R² = 0.748). Furthermore, SHapley Additive exPlanations was employed to interpret the model, further verifying the potential role of the lattice constant. Based on the above studies, Cr, Ni, and Mn were identified as doping elements. Finally, the first-principles calculation is employed to verify the prediction results of the machine learning. First-principles calculation results revealed that the austenite doped with Cr exhibits the lowest system energy (-29609.59 eV) and solid solution energy (-11.795 eV). Electronic structure analysis (including charge density difference and density of states) reveals the underlying mechanism: the larger the lattice constant of the doping atom, the weaker its interaction with the iron atoms, as specifically manifested by the reduction in electron cloud density and covalent bonding. These findings not only confirm the scientific validity of the proposed integrated machine learning and first-principles approach but also offer important guidance for the rational design of high-performance austenitic materials.

Keywords: austenitic stainless steels; machine learning; doping; first-principles

This article is published online first and will appear in Metalurgija, Vol. 66, Issue 1 (2027).