[Paper Review] MLIP-3: Active learning on atomic environments with Moment Tensor Potentials

Evgeny V. Podryabinkin, Kamil Garifullin+2 more|arXiv (Cornell University)|Apr 25, 2023

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

MLIP-3 introduces an enhanced software package for constructing and actively training moment tensor potentials (MTPs) on atomic neighborhoods within large-scale simulations. It enables efficient, parallelized active learning on local atomic environments, significantly improving data efficiency and model accuracy while maintaining high computational performance for materials modeling applications.

ABSTRACT

Nowadays, academic research relies not only on sharing with the academic community the scientific results obtained by research groups while studying certain phenomena, but also on sharing computer codes developed within the community. In the field of atomistic modeling these were software packages for classical atomistic modeling, later -- quantum-mechanical modeling, and now with the fast growth of the field of machine-learning potentials, the packages implementing such potentials. In this paper we present the MLIP-3 package for constructing moment tensor potentials and performing their active training. This package builds on the MLIP-2 package (Novikov et al. (2020), The MLIP package: moment tensor potentials with MPI and active learning. Machine Learning: Science and Technology, 2(2), 025002.), however with a number of improvements, including active learning on atomic neighborhoods of a possibly large atomistic simulation.

Motivation & Objective

• To develop a next-generation software package for training moment tensor potentials (MTPs) with improved usability and scalability.
• To enable active learning on local atomic neighborhoods extracted from large-scale atomistic configurations, reducing data requirements and computational cost.
• To enhance parallelization capabilities for distributed computing environments to accelerate MTP training and simulation.
• To support high-accuracy, low-cost machine learning potentials for complex materials properties such as diffusion, melting, and thermal conductivity.
• To provide a user-friendly, extensible framework that maintains the accuracy and generalization of MTPs while enabling efficient model refinement through active learning.

Proposed method

• Leverages moment tensor potentials (MTPs), which express the energy as a sum over local atomic neighborhoods using tensor-based basis functions.
• Applies active learning by selecting the most uncertain or informative atomic neighborhoods from a large simulation for first-principles training.
• Uses a linear expansion of MTP basis functions: $ V^{\text{MTP}}(\mathfrak{n}_i) = \sum_{\alpha=1}^{N_{\text{lin}}} \xi_\alpha B_\alpha(\mathfrak{n}_i) $, where parameters $ \xi_\alpha $ are trained via regression.
• Implements parallel computation using MPI for distributed training and evaluation across multiple compute nodes.
• Performs convergence testing on VASP-calculated energies and forces using k-mesh, energy cutoff, and vacuum size to ensure accuracy of reference data.
• Validates model performance by comparing forces on central atoms in isolated clusters versus the full periodic system, confirming consistency across scales.

Experimental results

Research questions

• RQ1 Can active learning on atomic neighborhoods improve the data efficiency and accuracy of moment tensor potential training?
• RQ2 How does the force on a central atom in a locally extracted cluster compare to its force in the full periodic system?
• RQ3 What are the optimal convergence parameters (k-mesh, ENCUT, vacuum size) for first-principles reference data used in MTP training?
• RQ4 To what extent does the new MLIP-3 package improve scalability and usability over MLIP-2 for large-scale materials simulations?
• RQ5 Can MTPs trained via active learning on local environments maintain high accuracy for complex materials properties like diffusion and thermal conductivity?

Key findings

• The force on the central atom in a locally extracted 8 Å radius cluster (165 atoms) closely matches the force in the full 336-atom periodic system, with differences below 0.05 eV/Å.
• Convergence testing confirmed that an energy cutoff of 300 eV, 1×1×1 k-mesh, and 9 Å of vacuum per direction are sufficient for accurate first-principles reference data.
• The optimal k-mesh for the periodic configuration was found to be 3×3×1, yielding a converged energy of -1181.358 eV.
• The MTP model trained on active learning of local neighborhoods achieved high accuracy with minimal reference data, reducing computational cost by focusing on critical atomic environments.
• MLIP-3 demonstrated improved parallelization and usability over MLIP-2, enabling efficient training on large-scale simulations via MPI and active learning workflows.
• The package successfully supports high-accuracy modeling of materials properties such as melting points, thermal conductivity, and diffusion coefficients, as validated in prior studies using MTPs.