Simple machine-learned interatomic potentials for complex alloys

Abstract

Developing data-driven machine-learning interatomic potentials for materials containing many elements becomes increasingly challenging due to the vast configuration space that must be sampled by the training data. We study the learning rates and achievable accuracy of machine-learning interatomic potentials for many-element alloys with different combinations of descriptors for the local atomic environments. We show that for a five-element alloy system, potentials using simple low-dimensional descriptors can reach meV/atom-accuracy with modestly sized training datasets, significantly outperforming the high-dimensional SOAP descriptor in data efficiency, accuracy, and speed. In particular, we develop a computationally fast machine-learned and tabulated Gaussian approximation potential (tabGAP) for Mo-Nb-Ta-V-W alloys with a combination of two-body, three-body, and a new simple scalar many-body density descriptor based on the embedded atom method.

Complex Ga2O3 Polymorphs Explored by Accurate and General-Purpose Machine-Learning Interatomic Potentials

“…In our previous study, a ML Gaussian approxima-7 tion potential (GAP) was developed to describe two- ter descriptor [64] and EAM is a descriptor correspond-19 ing to the pairwise-contributed density used in embedded 20 atom method potentials [63] (see Supplementary Infor- The validation of the soapGAP and tabGAP is shown 43 in Fig. 2, where both potential energies and force compo- The standard deviations of the force error distribution are σ tabGAP ≃ 3.33σ soapGAP , similar to that of the energy errors.…”

General-purpose machine-learned potential for 16 elemental metals and their alloys

“…2c). The comparison moreover shows that the UNEP-v1 model trained against 1-component and 2-component structures also performs very well for 3-component [34], 4-component [33], and 13component [32] structures extracted from the datasets in the previous studies [32][33][34]. The testing RMSEs of the UNEP-v1 model for these three datasets are respectively 57 meV/ Å, 196 meV/ Å, and 258 meV/ Å, which are com- parable to those reported as training RMSEs in the original publications [32][33][34].…”

Angular-dependent interatomic potential for large-scale atomistic simulation of W-Mo-Nb ternary alloys

“…The good accuracy of defects description, combined with high computational efficiency, indicates that the proposed ADP can be used to study a variety of materials scientific problems, such as diffusion or mobility of defects in pure metals and CCAs. The validation of pure metal properties showed that among the interatomic potentials of the W-Mo-Nb system, only the tabGAP model [43] can compete with the developed ADP. Overall, ADP provides a more accurate description of molybdenum and niobium, and tabGAP reproduces the properties of tungsten with better accuracy.…”

“…Results for the EAM potential of Nb are taken from published works [15,61]. Also, we included in this validation part a comparison of classical potentials with recently developed advanced ML-models for RAs: SNAP [40], MTP [41] and tabGAP [43].…”

“…Recently, a new generation of machine-learning (ML) potentials have gained a lot of attention. For refractory CCAs, several ML potentials can be highlighted: a spectral neighbor analysis potential (SNAP) for Nb-Mo-Ta-W [40]; a moment tensor potential (MTP) for Mo-Nb-Ta-W [41]; the recent tabulated Gaussian approximation potential (tabGAP) for Mo-Nb-Ta-V-W [42,43]. These ML potentials are significantly more accurate than the set of Zhou's potentials.…”