Minimalist machine-learned interatomic potentials can predict complex structural behaviors accurately (2511.17449v1)

Abstract: The past decade has witnessed a spectacular development of machine-learned interatomic potentials (MLIPs), to the extent that they are already the approach of choice for most atomistic simulation studies not requiring an explicit treatment of electrons. Typical MLIP usage guidelines emphasize the need for exhaustive training sets and warn against applying the models to situations not considered in the corresponding training space. This restricts the scope of MLIPs to interpolative calculations, essentially denying the possibility of using them to discover new phenomena in a serendipitous way. While there are reasons to be cautious, here we adopt a more sanguine view and challenge the predictive power of two representative and widely available MLIP approaches. We work with minimalist training sets that rely on little prior knowledge of the investigated materials. We show that the resulting models -- for which we adopt modest/default choices of the defining hyperparameters -- are very successful in predicting non-trivial structural effects (competing polymorphs, energy barriers for structural transformations, occurrence of non-trivial topologies) in a way that is qualitatively and quasi-quantitatively correct. Our results thus suggest an expanded scope of modern MLIP approaches, evidencing that somewhat trivial -- and easy to compute -- models can be an effective tool for the discovery of novel and complex physical phenomena.