Speaker
Description
The development of machine learned interatomic potentials (MLIPs) has revolutionised atomistic simulations over recent years. MLIPs enable simulations with accuracy comparable to quantum mechanical approaches at up to 10 million times smaller computational cost. The ephemeral data derived potentials (EDDPs) are a flexible class of MLIP that can utilize comparatively small neural networks for fast and accurate simulations. Materials discovery techniques, such as random structure searching, can exploit this acceleration in a number of ways. Larger system sizes and chemical spaces can be explored, or more computationally intensive structure optimizers, utilizing molecular dynamics, can be incorporated. Through a range of applications, we demonstrate that more complex optimizers can improve search performance for large systems. Through example simulations in LAMMPS, we further show that EDDPs can be employed beyond structure searching, to explore complex dynamical behaviour and synthesis pathways across a range of conditions and chemical systems.