Building models for the plasticity, thermodynamics and kinetics of metals is challenging as subtle aspects of atomic cohesion must be faithfully reproduced, and predictions often require averaging over large, complex configuration ensembles. I will discuss how the energy landscapes of atomic systems can be rapidly explored at scale and "coarse-grained" with UQ when the dynamics are thermally activated[1,2] and how descriptor techniques, typically used to regress energies for modern cohesive models, can be used to capture a much wider range of properties such as defect entropics[3] or dislocation properties[4].
These techniques can connect large-scale atomic data to a medium dimensional metric space, offering many scalable, low-overhead approaches to
maximise diversity or mitigate cost at scale. In particular, I show how classical timeseries tools applied to descriptor trajectories can give uncertainty and
extrapolation-aware predictions of nanoparticle aging and dislocation network yielding, complex systems whose dynamics
are largely untreatable with current acceleration techniques[5].
[1] TDS and D Perez, NPJ Comp. Mat 2020
[2] TDS and DJ Wales JCTC 2020
[3] C Lapointe et al. PRMat 2022
[4] P Griorev et al., Acta Materialia 2023
[5] TDS, submitted
Copyright. All Rights Reserved.