[Paper Review] NEP89: Universal neuroevolution potential for inorganic and organic materials across 89 elements

Ting Liang, Ke Xu+17 more|ArXiv.org|Apr 30, 2025

Machine Learning in Materials Science6 citations

TL;DR

NEP89 is a foundation machine-learned potential that covers 89 elements across inorganic and organic materials, delivering near-first-principles accuracy with empirical-potential-like efficiency and enabling large-scale MD simulations and fine-tuning for user-specific applications.

ABSTRACT

While machine-learned interatomic potentials offer near-quantum-mechanical accuracy for atomistic simulations, many are material-specific or computationally intensive, limiting their broader use. Here we introduce NEP89, a foundation model based on neuroevolution potential architecture, delivering empirical-potential-like speed and high accuracy across 89 elements. A compact yet comprehensive training dataset covering inorganic and organic materials was curated through descriptor-space subsampling and iterative refinement across multiple datasets. NEP89 achieves competitive accuracy compared to representative foundation models while being three to four orders of magnitude more computationally efficient, enabling previously impractical large-scale atomistic simulations of inorganic and organic systems. In addition to its out-of-the-box applicability to diverse scenarios, including million-atom-scale compression of compositionally complex alloys, ion diffusion in solid-state electrolytes and water, rocksalt dissolution, methane combustion, and protein-ligand dynamics, NEP89 also supports fine-tuning for rapid adaptation to user-specific applications, such as mechanical, thermal, structural, and spectral properties of two-dimensional materials, metallic glasses, and organic crystals.

Motivation & Objective

• Aim to develop a universal interatomic potential spanning 89 elements for both inorganic and organic materials.
• Curate and harmonize diverse public datasets to train a unified model.
• Achieve competitive accuracy with existing foundation models while dramatically improving computational efficiency.
• Demonstrate out-of-the-box large-scale MD capabilities and fine-tuning for specialized applications.

Proposed method

• Use the NEP architecture with atom-centered descriptors built from Chebyshev and Legendre polynomials.
• Employ a separable natural evolution strategy to train a neural network with one hidden layer.
• Encode species via radial function expansions with independent coefficient sets and per-species network parameters.
• Iteratively curate and balance a diverse training set from multiple public datasets, adding D3 dispersion corrections where needed.
• Adjust reference energies across datasets to enable a unified single-task training outcome.

Experimental results

Research questions

• RQ1 Can a single interatomic potential model accurately describe both inorganic and organic materials across 89 elements?
• RQ2 How does NEP89 perform in static property benchmarks compared to other foundation models?
• RQ3 Is NEP89 capable of efficient large-scale MD simulations and fine-tuning for specialized tasks?
• RQ4 Can a unified training dataset from multiple public sources yield reliable energies, forces, and stresses for a broad chemical space?
• RQ5 What is the potential for fine-tuning NEP89 to improve accuracy for specific materials or properties?

Key findings

• NEP89 achieves competitive accuracy across multiple static-property benchmarks relative to other foundation models.
• The model delivers 3-4 orders of magnitude speedup and memory efficiency gains over comparable models for 20-element alloys, enabling much larger simulations.
• NEP89 reproduces key dynamical properties such as bonding statistics in amorphous carbon and structural/dynamic features of water and solid-state electrolytes.
• Out-of-the-box large-scale MD with NEP89 demonstrates qualitative agreement with experiments and AIMD in various systems, including multicomponent alloys, methane combustion, and protein-ligand interactions.
• Fine-tuning NEP89 with a small dataset yields markedly improved agreement with experiments for specific materials, as shown in MoSi2N4 example, and can be used to tailor models for target properties.