Thomas Rocke

Machine-learning interatomic potential (MLIP)-based molecular dynamics enables efficient exploration of mass-transport mechanisms in complex materials together with quantitative determination of diffusion coefficients at finite temperatures. Using an identical training dataset, we develop and assess three MLIP formalisms [moment tensor potentials (MTP), atomic cluster expansion (ACE), and message passing ACE (MACE)] and evaluate their transferability for diffusion simulations in hexagonal transition metal diborides. After benchmarking against 0 K ab initio migration energies and comparison of computational efficiency, the MTP model is employed to investigate boron-vacancy diffusion in α-TiB2, α-TaB2, and ω-WB2, representing group IV–VI transition metal diborides. Two distinct monovacancy diffusion pathways are identified within the puckered boron layer of ω-WB2 and analyzed as a function of temperature. The combined effects of strain and temperature on diffusivity are further examined, revealing a weak sensitivity to external stress under conditions relevant to high-temperature machining applications. Overall, nanosecond-scale finite-temperature simulations demonstrate the reliability and practical applicability of MLIPs for studying diffusion phenomena in complex ceramic materials.

The problem of constructing a dataset for MLIP development which gives the maximum quality in the minimum amount of compute time is complex, and can be approached in a number of ways. We introduce a ‘Bayesian selection’ approach for selecting from a candidate set of structures, and compare the effectiveness of this method against other common approaches in the task of constructing ideal datasets targeting Silicon surface energies. We show that the Bayesian selection method performs much better than Simple Random Sampling at this task (for example, the error on the (100) surface energy is 4.3× lower in the low data regime), and is competitive with a variety of existing selection methods, using ACE (Drautz 2019 Phys. Rev. B 99 014104) and MACE (Batatia et al 2022 Advances in Neural Information Processing Systems pp 11423–36) features.

Behaviour of materials is governed by physical phenomena that occur at an extreme range of length and time scales. Computational modelling requires multiscale approaches. Simulation techniques operating on the atomic scale serve as a foundation for such approaches, providing necessary parameters for upper-scale models. The physical models employed for atomic simulations can vary from electronic structure calculations to empirical force fields. However, construction, manipulation and analysis of atomic systems are independent of the given physical model but dependent on the specific application. matscipy implements such tools for applications in materials science, including fracture, plasticity, tribology and electrochemistry.