Efficient moment tensor machine-learning interatomic potential for accurate description of defects in Ni-Al Alloys

Combining the efficiency of semi-empirical potentials with the accuracy of quantum mechanical methods, machine-learning interatomic potentials (MLIPs) have significantly advanced atomistic modeling in computational materials science and chemistry. This necessitates the continual development of MLIP models with improved accuracy and efficiency, which enable long-time scale molecular dynamics simulations to unveil the intricate underlying mechanisms that would otherwise remain elusive. Among various existing MLIP models, the moment tensor potential (MTP) model employs a highly descriptive rotationally-covariant moment tensor to describe the local atomic environment, enabling the use of even linear regression for model fitting. Although the current MTP model has achieved state-of-the-art efficiency for similar accuracy, there is still room for optimizing the contraction process of moment tensors. In this work, we propose an effective genetic algorithm based optimization scheme that can significantly reduce the number of independent moment tensor components and intermediate tensor components. This leads to a speedup of nearly one order of magnitude in efficiency and also improved accuracy compared to the traditional MTP model for intricate basis sets. We have applied our improved MTP model to predicting the energetic and dynamical properties of various point and planar defects in Ni-Al alloys, showing overall good performances and in general outperforming the semi-empirical potentials. This work paves the way for fast and accurate atomistic modeling of complex systems and provides a useful tool for modeling defects in Ni-Al alloys.