Igor A. Abrikosov

F. Bock, B. O. Mukhamedov, F. Tasnádi

Department of Physics, Chemistry and Biology (IFM), Linkoping University, SE-581 83, Sweden

Combining ab-initio and machine learning methods in simulations of properties of alloys for hard coatings and biomedical applications

First-principles calculations of materials parameters in the framework of the density functional theory (DFT) are recognized as a state-of-the-art tool within condensed matter theory and materials science. However, the computational demands associated with calculating properties of materials relevant for practical applications using DFT-based methods are often prohibitively high, especially when studying multiple systems at the conditions of their operation, e.g. as cutting tools, that is at very high temperature and stresses. Machine learning approaches have been introduced into the field of materials modeling to address the challenge. Considering a prototype hard-coating alloy, B1 Ti_0.5Al_0.5N, we present an active learning (AL) workflow to model its properties at temperatures up to 1500 K. With a minimal prior knowledge about the alloy system for the AL workflow, we train interatomic potential described through moment tensors, the so-called moment tensor potential (MTP) [1,2] and benchmark the simulation results against alternative techniques, including high-accuracy ab initio molecular dynamics simulations [3,4]. Using an automated high-throughput workflow, we built a database of industrially relevant hard-coating materials, Hard-coating Alloys Database (HADB), containing data on binary and ternary nitrides [5,6]. We explore ways for machine learning to support and complement the designed databases. Employing MTPs, we investigate temperature dependence of elastic moduli in Ti-based alloys relevant for biomedical applications and predict attractive properties for alloys with compositions in a vicinity of mechanical instability.