Deterministic and Probabilistic Descriptor Models for Complex Atomic Simulations

I will discuss how descriptors (high-dimensional, many-body order parameters) can be used to coarse-grain atomic simulations of materials. These are typically used in a deterministic setting, i.e. to predict energy and forces given atomic positions. Capturing model uncertainty in this regime requires proper treatment of model misspecification which is entirely neglected when training with the expected loss, but the true generalization error can be approximately minimised[1]. Deterministic models can also capture important microstructural features such as dislocation densities, correlation functions, or vibrational entropies, offering many efficiencies when deploying at scale. Descriptor timeseries extracted from simulation trajectories can be used to make probabilistic forecasts of simulation futures; in this setting, ideas from active learning are very useful to quantify and qualify prediction uncertainty[2]. The opportunities provided by descriptors for data-driven analysis of extreme defect dynamics will be illustrated through application to nanoparticles and plastic deformation of metals.<br/>[1] TD Swinburne and D Perez, arXiv 2402.01810 (2024)<br/>[2] TD Swinburne, Phys. Rev. Lett. 236101 (2023)