Machine-learning local resistive environments of dislocations in complex concentrated alloys from data generated by molecular dynamics simulations

Abstract

Complex concentrated alloys (CCAs) differ from pure metals and conventional dilute alloys in that the multiple constituent elements are prone to develop special local atomic environments (LAEs). Due to the complexity and spatial variability of the LAEs, the resistance that they offer to travelling dislocations cannot be determined a priori by conventional strengthening theories. In this work, molecular dynamics (MD) simulations of a prototypic CCA of NiCoCr were used to generate data for dislocation features that may potentially affect dislocation resistance. Extensive analysis of these features via their Pearson correlation coefficients with dislocation velocity and ablation studies using light gradient-boosting machine learning (ML) models show that (i) the local planar fault energy (PFE), (ii) local gradient of the PFE, and (iii) dislocation core width, while all prime factors for dislocation resistance, do not have strongly linear correlation with the dislocation velocity. However, reasonably high prediction accuracy (>80 %) is achieved when all three factors are included in the ML model. Furthermore, lattice distortion, a much-discussed strengthening factor for CCAs in the literature, is also not strongly linearly correlated and its effect can be well represented by the PFE. These results indicate that CCA strength is governed not by individual dislocation-resistance factors, but a synergistic combination of these factors that goes beyond any a priori assumption. This work highlights the complexity in the nature of CCA strength, and the suitability and success of machine learning as an a posteriori approach for understanding it.