A Software Package for Generating Robust and Accurate Potentials using the Moment Tensor Potential Framework

TL;DR

PRAPs addresses the challenge of building robust and accurate interatomic potentials for crystal structure prediction by providing an automated workflow to train two moment tensor potentials (robust and accurate) via active learning and by integrating with MLIP, AFLOW, and VASP. The framework includes data filtration, convergence controls, convex-hull analysis, and utilities (mliputils) to manage .cfg data and POSCAR conversions. It demonstrates practical performance and compares against universal interatomic potentials, showing improved hull predictions near stability boundaries. The package lowers manual workload and enables high-throughput CSP workflows with reliable extrapolation behavior.

Abstract

We present the Plan for Robust and Accurate Potentials (PRAPs), a software package for training and using moment tensor potentials (MTPs) in concert with the Machine Learned Interatomic Potentials (MLIP) software package. PRAPs provides an automated workflow to train MTPs using active learning procedures, and a variety of utilities to ease and improve workflows when utilizing the MLIP software. PRAPs was originally developed in the context of crystal structure prediction, in which one calculates convex hulls and predicts low energy metastable and thermodynamically stable structures, but the potentials PRAPs develops are not limited to such applications. PRAPs produces two potentials, one capable of rough estimates of the energies, forces and stresses of almost any chemical structure in the specified compositional space -- the Robust Potential -- and a second potential intended to provide more accurate descriptions of ground state and metastable structures -- the Accurate Potential. We also present a Python library, mliputils, designed to assist users in working with the chemical structural files used by the MLIP package.

Figures (5)

• Figure 1: An illustration of the compositional spaces used for training in a two-component system. The robust potential (RP) is trained on all structures. The accurate potential (AP) is trained on structures below the red line: those within a certain energy from the minimal energy structure of that composition. The solid line is the convex hull, and the red dots represent the structures that lie on it.
• Figure 2: Workflow in the Plan for Robust and Accurate Potentials (PRAPs) package, which automates the generation of a moment tensor potential (MTP), given any quantum mechanical training set. Step 1: Five MTPs are trained on a set of configurations and the best (the pre-Robust Potential, pre-RP) is chosen. Step 2: The pre-RP is employed to initialize the training of the Robust Potential (RP) via an active learning scheme (ALS, inset). Step 3: The RP-relaxed lowest energy configurations are chosen to train an Accurate Potential (AP) via active learning. Step 4: Convex hulls may be generated using the AP and AFLOW.
• Figure 3: Examples of convex hull plots generated by PRAPs at Level 10. (a) The AFLOW-derived convex hull. (b) Relaxation of the AFLOW set with the Robust Potential yields the RR hull. (c) Prediction of the enthalpies of the RR structures with the Accurate Potential results in the AP-RR set, (d) while relaxation of the RR structures with the Accurate Potential yields the AR-RR set. Structures are colored (see color bar) according to the distances from the hull. Black dots are on the hull and purple dots are 1 meV/atom from the hull. This figure is adapted from material created by Josiah Roberts and provided in the Supporting Information of Ref. Zurek:2023n (https://doi.org/10.1038/s41524-024-01321-7) licensed under a Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/).
• Figure 4: Examples of 2D convex hull plots generated by PRAPs for the (a) CHf, (b) CMo, and (c) HfMo systems. The AFLOW-derived convex hulls are on the left. Subsequent relaxation of those structures by the Robust Potential gives the middle plots. A second relaxation by the Accurate Potential yields the plots on the right.
• Figure 5: Examples of convex hull plots generated by PRAPs compared with universal interatomic potentials (UIPs). Structures displayed are within 60 meV/atom of each hull and are colored (see color bars) according to distances from the hull. (a) The AFLOW-derived convex hull. The middle row shows hulls derived by relaxing structures with: (b) the robust potential, and (c) subsequent relaxation with the accurate potential. The bottom row shows convex hulls derived, from the same data, from UIPs without any action by an MTP, showing (d) MACE and (e) MatterSim. Black dots are on the hull and purple dots are 1 meV/atom from the hull. The data has not undergone further DFT relaxations. Portions of this figure are adapted from material created by Josiah Roberts and provided in the Supporting Information of Ref. Zurek:2023n (https://doi.org/10.1038/s41524-024-01321-7) licensed under a Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/).