Graph atomic cluster expansion for foundational machine learning interatomic potentials

TL;DR

This work introduces GRACE, a Graph Atomic Cluster Expansion-based framework for universal foundation interatomic potentials that span the entire periodic table. By extending ACE to tree-graphs and leveraging complete basis representations with recursive, tensor-decomposed expansions, GRACE achieves high accuracy and efficient evaluation (FP64) across diverse materials datasets. Extensive validation across MatBench Discovery, thermal conductivity, elastic properties, and defect structures demonstrates Pareto-optimal accuracy vs. efficiency and robust transferability, including long-time MD stability in complex systems. The authors further show practical versatility through fine-tuning for Al-Li and hydrogen combustion, and model distillation to compact, fast student models, highlighting GRACE as a flexible, scalable platform for next-generation atomistic simulations with broad applicability and minimal retraining bottlenecks.

Abstract

Foundational machine learning interatomic potentials that can accurately and efficiently model a vast range of materials are critical for accelerating atomistic discovery. We introduce universal potentials based on the graph atomic cluster expansion (GRACE) framework, trained on several of the largest available materials datasets. Through comprehensive benchmarks, we demonstrate that the GRACE models establish a new Pareto front for accuracy versus efficiency among foundational interatomic potentials. We further showcase their exceptional versatility by adapting them to specialized tasks and simpler architectures via fine-tuning and knowledge distillation, achieving high accuracy while preventing catastrophic forgetting. This work establishes GRACE as a robust and adaptable foundation for the next generation of atomistic modeling, enabling high-fidelity simulations across the periodic table.

Figures (19)

• Figure 1: Model performance for stable structure identification (F1 score in MatBench Discovery benchmark) and thermal conductivity prediction ($\kappa_\mathrm{SRME}$) versus computational time per atom. A higher F1 score and lower $\kappa_\mathrm{SRME}$ indicate better performance. The blue dashed line links Pareto optimal models. Computational performance is estimated via ASE (filled symbols) and LAMMPS (open symbols), with GRACE models indicated in red.
• Figure 2: The symmetric relative mean error (SRME) and MAE ($\Delta C$ in GPa) for elastic constants, categorized into three subgroups: longitudinal ($C_{11}, C_{22}, C_{33}$), Poisson's ratio-related ($C_{12}, C_{13}, C_{23}$), and shear ($C_{44}, C_{55}, C_{66}$). See text for more details.
• Figure 3: Error metrics for unary grain boundaries formation energies: $\gamma_\mathrm{GB}$-SRME (left) and mean absolute error $\Delta \gamma_\mathrm{GB}$ (right). GRACE models are highlighted in red.
• Figure 4: Error metrics for unary surface formation energies: $\gamma_\mathrm{surf}$-SRME (left) and mean absolute error $\Delta \gamma_\mathrm{surf}$ (right). GRACE models are highlighted in red.
• Figure 5: Error metrics for SRME of point defects formation energies in unaries: self-interstitials and vacancies. GRACE models are highlighted in red.