Machine learning models for the prediction of energy, forces, and stresses for Platinum

Materials properties such as defect diffusion and/or dissociation, mechanical fracture and void nucleation, under extreme temperatures and pressures, are all governed by the interactions between individual and/or groups of atoms. Computational tools have been instrumental in understanding the atomistic properties of materials at these length scales. Over the past few decades, these tools have been dominated by two levels of theory: quantum mechanics (QM) based methods and semi-empirical/classical methods. The former are time-intensive, but accurate and versatile, while the latter methods are fast but are significantly limited in veracity, versatility and transferability. Machine learning (ML) algorithms, in tandem with quantum mechanical methods such as density functional theory, have the potential to bridge the gap between these two chasms due to their (i) low cost, (ii) accuracy, (iii) transferability, and (iv) ability to be iteratively improved. In this work, we prescribe a new paradigm in which potential energy, atomic forces, and stresses are rapidly predicted by independent machine learning models, all while retaining the accuracy of quantum mechanics. This platform has been used to study thermal, vibrational, and diffusive properties of bulk Platinum, highlighting the framework's ability to reliably predict materials properties under dynamic conditions. We then compare our ML framework to both QM, where applicable, and several Embedded Atom Method (EAM) potentials. We conclude this work by reflecting upon the current state of ML in materials science for atomistic simulations.