Speaker
Description
Transition metal nitrides (TMNs) exhibit exceptional mechanical properties influenced significantly by crystallographic defects. Modeling these defects with ab initio accuracy is challenging due to large system sizes and often lower symmetries resulting from local relaxations around a defect. In this contribution, we discuss the development of machine learning interatomic potentials (MLIPs) for cubic rocksalt TaN, targeting the impact of vacancies and deformation on potential accuracy. As a representative Group 5 system, TaN is an ideal model for investigating defect-property correlations, as its cubic phase is generally stabilized by high concentrations of vacancies on the nitrogen and tantalum sublattices.
In our work, a training and validation dataset was generated by finite-temperature ab initio molecular dynamics (AIMD) calculations, including vacancy-rich and medium-to-severely deformed supercells. While rather low precision—such as the lowest cutoff energy settings for the plane-wave basis set and only Gamma-point sampling of the reciprocal space—ensures efficient convergence, it introduces a systematic 25 GPa external pressure offset in both the Atomic Cluster Expansion (ACE, as implemented in the pacemaker package) and Moment Tensor Potentials (MTPs, as implemented in the MLIP-3 package) training frameworks. Regarding energy accuracy, both formalisms achieve a low RMSE (< 9 meV/atom), with ACE exhibiting superior performance (< 2 meV/atom). However, MTP-type force fields offer a distinct advantage for mechanical modeling; their training framework allows for explicit stress weighting in the loss function, capturing elastic responses often overlooked in pure energy fitting.
Finally, large-scale molecular dynamics (MD) simulations reveal a vacancy-induced phase transition from the rocksalt structure to a hexagonal-like system. These results underscore the necessity of balancing energy precision with stress-informed weighting to accurately predict the stability and mechanical behavior of defective TMNs.