A machine learning-based framework for predicting the solute segregation energy in bicrystal and nanocrystalline grain boundaries

RK Jha and R Kumar and RS Varsheni and S Mandal, JOURNAL OF MATERIALS SCIENCE, 60, 23502-23525 (2025).

DOI: 10.1007/s10853-025-11797-3

In this study, several machine learning approaches, namely linear regression models, decision trees-based models, and a deep learning- based model, are utilized to predict the segregation energy of Nb solute in both bicrystal and nanocrystalline Ni grain boundaries (GBs). In this regard, per-atom metrics (e.g., local stress, Voronoi volume, coordination number, etc.) are extracted from undecorated pure Ni GBs. The results show that XGBoost consistently achieves the highest prediction accuracy, and the performance gap relative to other models (e.g., Random Forest, Extra Trees, Artificial Neural Networks) becomes more pronounced in nanocrystalline specimen with greater structural diversity. To further investigate the influence of solute-solute interactions on GB solute segregation spectra in Ni-Nb alloys, segregation energies are calculated for each GB site in Ni-0.2Nb, and Ni-2Nb. The simulation results indicate that the segregation energy distribution in pure Ni is concentrated within a narrow range, while the alloy systems (i.e., Ni-0.2Nb, and Ni-2Nb) exhibit broader and more dispersed distribution. Furthermore, the machine learning models are employed to predict the solute segregation energy in nine distinct solutes (Co, Fe, Ag, Nb, Zr, Cu, Cr, Mn, and Al) in bicrystal Ni GBs. The analysis reveals that the XGBoost model provides excellent performance across a wide range of solutes with different crystal structures, such as BCC (Fe), HCP (Zr), and FCC (Cu), in predicting the segregation energy in Ni bicrystal GBs.

Return to Publications page