Abstract
Foundation models have proven very popular in recent years, with new record-breaking models
coming out every month. But how accurate are they, and can they replace "bespoke" potentials? In
this talk I will present the state of the art in both categories.
First, I will show that using the GRACE2L-OAM foundation model we can achieve very high accuracy
across a broad range of metallurgically-relevant benchmarks "out of the box" with no further
modification or fine-tuning. Overall these models are able to produce results that are almost as
accurate as a collection of custom BP-NNPs for aluminium developed over the course of several years
by hand.
I will then show how the ACE model can be used to create a class of very powerful next-generation
custom potentials. Some techniques, such as ladder learning, and training on individual elements first
will be discussed. A performance evaluation showing how while foundation models can produce very
impressive results, for the greatest accuracy and speed one should still use 'bespoke models'.
This talk will include the challenges in computing material properties: elastic constants, lattice
constants, surface energies, generalized stacking fault energies and dilute solute interactions. I will also
cover how to select and evaluate foundation models and reasons why any of the most popular datasets
(at time of writing) are insufficient for aluminium metallurgy. Computational considerations, such as
how to design effective 'unit tests' for interatomic potentials, and computational performance issues
regarding the ASE python interface vs directly computing with LAMMPS.
While the focus will be on aluminium alloys, the author will also mention the use of foundation
potentials in the context of designing new batteries and magnetic materials.