Coarse-Graining and Forecasting Atomic Material Simulations with Descriptors

Atomic simulations offer unparalleled insight into materials and molecules.
However, most data is thrown away, as they cannot be easily stored or analysed.
This is an important and general problem as interesting simulations are usually hard to predict.

In Physical Review Letters, Thomas Swinburne, TSN-CINaM, shows that representations of atomic structure used in modern energy models also offer excellent data compression.
Trajectories can be easily stored, analysed and even forecasted, offering many perspectives for coarse-grained modelling.
For example, the yielding transition in metals is manifest as a sharp change in the descriptor manifold dimension, generalising the classical metallurgical concept of the yield surface to a high-dimensional yield manifold.
Future work will apply the approach to study currently intractable multi-scale simulation problems in materials.

Thomas D. Swinburne – Phys. Rev. Lett. 131, 236101 – Published 8 December 2023