Predicting grain boundary segregation in magnesium alloys: An atomistically informed machine learning approach

Abstract

Grain boundary (GB) segregation substantially influences the mechanical properties and performance of magnesium (Mg). Atomic-scale modeling, typically using ab-initio or semi-empirical approaches, has mainly focused on GB segregation at highly symmetric GBs in Mg alloys, often failing to capture the diversity of local atomic environments and segregation energies, resulting in inaccurate structure-property predictions. This study employs atomistic simulations and machine learning models to systematically investigate the segregation behavior of common solute elements in polycrystalline Mg at both 0 K and finite temperatures. The machine learning models accurately predict segregation thermodynamics by incorporating energetic and structural descriptors. We found that segregation energy and vibrational free energy follow skew-normal distributions, with hydrostatic stress, an indicator of excess free volume, emerging as an important factor influencing segregation tendency. The local atomic environment’s flexibility, quantified by flexibility volume, is also crucial in predicting GB segregation. Comparing the grain boundary solute concentrations calculated via the Langmuir-McLean isotherm with experimental data, we identified a pronounced segregation tendency for Nd, highlighting its potential for GB engineering in Mg alloys. This work demonstrates the powerful synergy of atomistic simulations and machine learning, paving the way for designing advanced lightweight Mg alloys with tailored properties.