HITS-SIMPLAIX Joint Colloquium Volker Deringer: Machine-learning-driven advances in modelling and understanding disordered materials

By Volker Deringer, Department of Chemistry, University of Oxford, UK

Machine learning (ML) based interatomic potential models are increasingly popular simulation tools for molecular and materials systems and hold promise for use on exascale supercomputers [1]. ML potentials are fitted to large sets of quantum-mechanical reference data, and therefore developing high-quality datasets and automated training approaches is becoming an increasingly important re-search challenge. In this seminar, I will highlight some recent developments in ML-driven molecular-dynamics simulations of structurally complex inorganic materials, combining methodological aspects and practical applications. In regard to methods, I will discuss the use of cheaply available “synthetic” data in pre-training atomistic ML models [2], which can improve accuracy and robustness of neural-network interatomic potentials compared to direct training on quantum-mechanical data [3]. Regarding applications, I will showcase device-scale simulations of phase-change memory materials (which encode digital “ones” and “zeroes” in data-storage devices) [4]. Finally, I will discuss perspectives for the development of both purpose-specific and generally applicable ML potentials for materials.

[1] C. Chang et al., Nat. Rev. Mater. 8, 309 (2023).
[2] J. L. A. Gardner et al., Digital Discovery 2, 651 (2023).
[3] J. L. A. Gardner et al., arXiv:2307.15714 [physics.comp-ph].
[4] Y. Zhou et al., Nat. Electron., DOI: 10.1038/s41928-023-01030-x (2023).

Short CV:

Please see here: Deringer_Short_CV

REGISTRATION: