Active learning with moment tensor potentials to predict material properties: Ti0.5Al0.5N at elevated temperature

Abstract

Transition metal nitride alloys possess exceptional properties, making them suitable for cutting applications due to their inherent hardness or as protective coatings due to corrosion resistance. However, the computational demands associated with predicting these properties using ab initio methods can often be prohibitively high at the conditions of their operation at cutting tools, that is, at high temperatures and stresses. Machine learning approaches have been introduced into the field of materials modeling to address the challenge. In this paper, we present an active learning workflow to model the properties of our benchmark alloy system cubic B1 Ti0.5Al0.5N at temperatures up to 1500 K. With a minimal requirement of prior knowledge about the alloy system for our workflow, we train a moment tensor potential (MTP) to accurately model the material’s behavior over the entire temperature range and extract elastic and vibrational properties. The outstanding accuracy of MTPs with relatively little training data demonstrates that the presented approach is highly efficient and requires about two orders of magnitude less computational resources than state-of-the-art ab initio molecular dynamics.