How far can we take an atomistic simulation?

From atomic environments to dynamics

Describe the atomic environment

Use local descriptors to encode atomic configurations, with the symmetries required for modelling atomic interactions.

Conceptual workflow for a machine-learning interatomic potential. The atomic arrangement and model blocks illustrate the method; no simulation is running.

The scientific question

Atomistic simulation repeatedly asks for energies and forces. Accurate reference calculations can make long trajectories prohibitively expensive. My PhD focused on developing learned potentials that make more of that simulation possible.

Choosing what the model sees

Our 2018 work develops automatic selection of atomic fingerprints and reference configurations. It studies how to retain useful structural information while reducing the size of the representation and training problem.

Designing the potential

The later divide-and-conquer paper combines specialised linear potentials using smooth, configuration-dependent weights. Across its benchmarks, the combined models improve on single linear models with little or no extra computational cost.

Knowing the scope of a prediction

Our extrapolation work examines how high-dimensional representations complicate the distinction between interpolation and extrapolation. A useful potential needs evaluation in the environments and regimes where it will be used; speed alone does not establish reliability.