Universal force fields by machine learning: eleven models tested on 7,599 multi-component materials

Researchers compare eleven pre-trained interatomic potentials to DFT calculations and show that their errors explode when the atoms are compressed.

Machine-learned universal force fields promise to simulate matter at a fraction of the cost of quantum computations (DFTs). But do they hold up against the complex compositions that materials chemists are exploring? A team led by Hongwei Du and Siqi Shi built a test bed of 7,599 multi-component configurations, inspired by high-entropy alloys, element substitutions, and anion mixtures, and compared it with eleven pre-trained models.

Compression, a weak point of the models

Energies, forces, stresses, but also elastic, vibrational and adsorption properties were compared to DFT. Errors increase when atomic environments move away from the training data, without this explaining everything. Above all, they rise sharply when two atoms are closer than usual: for the same distance, the errors on the compression side are 1.81 to 1.95 times higher than those on the stretch side.

The difficulty in modeling an element is correlated with the sensitivity of its electronic structure to atomic displacements. Finally, for the same number of parameters, representations in spherical harmonics of higher order reduce the errors for 38 to 40 of the 43 elements studied. The work provides benchmarks for choosing a model and orienting the collection of training data.