Source-linked AI summary
NEP89: Universal neuroevolution potential for inorganic and organic materials across 89 elements
Ting Liang, Ke Xu, Eric Lindgren, Zherui Chen, Rui Zhao, Jiahui Liu, Esmée Berger, Benrui Tang, Bohan Zhang, Yanzhou Wang, Keke Song, Penghua Ying, Nan Xu, Haikuan Dong, Shunda Chen, Paul Erhart, Zheyong Fan, Tapio Ala-Nissila, Jianbin Xu
TL;DR
Large-scale atomistic simulations need accurate potentials, but ab initio methods are prohibitively costly and existing universal models remain limited in material coverage or efficiency. NEP89 combines a curated inorganic-and-organic training dataset with a neuroevolution potential across 89 elements. It often matches or outperforms representative foundation models while extending accessible simulation scales by at least three orders of magnitude.
Problem
Ab initio molecular dynamics is prohibitively costly for million-atom simulations, while many-element potentials remain mostly inorganic and computationally expensive.
Source: §I · p. 1Method
NEP89 combines a neuroevolution potential with an iteratively curated training set that unifies inorganic and organic materials across 89 elements.
Source: §I · p. 2, §II · p. 2Results
NEP89 often matches or outperforms representative foundation models across static and dynamic properties and extends accessible simulation scales by at least three orders of magnitude.
Source: §II · p. 4, §III · p. 9Takeaways & Limitations
NEP89 supports large-scale simulations across diverse inorganic and organic systems and can be fine-tuned into specialized models using limited training data.
Source: §III · p. 9NEP89 lacks explicit long-range interactions, which may affect quantitative accuracy in ionic, aqueous, and large biomolecular systems.
Source: §III · p. 9
Abstract
from arXiv · showhide
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 near-empirical-potential 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.
I. INTRODUCTION
The introduction identifies a gap in broadly applicable, efficient potentials for large-scale simulations across inorganic and organic materials. NEP89 is presented as a fast, high-fidelity model across 89 elements, supported by a curated unified training dataset.
I. INTRODUCTION
- AIMD offers the desired fidelity but is prohibitively costly for simulations involving millions of atoms or more.Large-scale MD depends critically on the accuracy of its interatomic potentials.Source: §I · p. 1
- Existing many-element potentials were largely built from inorganic materials, while high computational cost limits their use in large-scale or long-timescale simulations.Earlier models were developed for specific materials or properties, and broader models expanded from single elements to nearly the periodic table [8–17].Source: §I · p. 1
- NEP89 targets high-fidelity simulations of inorganic and organic materials across 89 elements, with speeds several orders of magnitude above previous universal potentials.It is based on the neuroevolution potential approach, which combines accuracy with efficiency.Source: §I · p. 2
- A unified inorganic-and-organic potential requires curated datasets with consistent targets, dispersion interactions, and optimized relative energies across datasets.The training set is built iteratively from subsampled OMAT24 and structures from other public datasets [23], with D3 corrections added to some datasets.Source: §I · p. 2
II. RESULTS
NEP89 combines a curated dataset spanning inorganic and organic materials with a fast, fine-tunable potential. It shows competitive accuracy across static and dynamical benchmarks and supports large-scale simulations and targeted applications.
Iterative training and fine-tuning
- NEP89 training combines diverse datasets through descriptor-space sampling, iterative error-driven refinement, and outlier removal, while adjusting reference energies for consistency.Fine-tuning reuses NEP89 parameter statistics and can adapt the model using trajectories sampled at target conditions and single-point DFT calculations.Source: §II · pp. 2, 3, 7
Benchmarks on static properties
- NEP89 shows comparable accuracy across nine static properties, with strong results for GNoME formation energies, lattice energies, and shear modulus.It ranks second on GNoME formation energies and S66 binding energies, leads on DMC-ICE13 lattice energies and Materials Project shear modulus, and is comparable on other evaluated properties.Source: §II · p. 4
Evaluation of computational performance
- On the tested GPUs, NEP89 simulates up to 8 × 10^6 atoms at 5 × 10^6 atom-steps per second or 1.5 × 10^7 atoms at 7 × 10^6 atom-steps per second.The paper reports at least three orders of magnitude greater accessible spatiotemporal scale than other foundation models, with competitive accuracy and multi-GPU scalability.Source: §II · p. 4
Benchmarks of dynamical properties
- NEP89 matches experimental or AIMD results across amorphous carbon, liquid water, and three solid-state electrolytes, outperforming other foundation models on water and electrolyte diffusion.Water density and oxygen–oxygen radial distribution function are particular strengths; electrolyte diffusion agrees with AIMD and a specialized force field across 300–1100 K.Source: §II · pp. 4–6
Out-of-the-box large-scale MD simulations
- Out of the box, NEP89 models a 1,221,240-atom alloy under compression and predicts its 9.2 GPa yield strength and 5.8% yield strain, close to experimental values.It also predicts ion diffusivities in water consistent with experiment and captures Arrhenius behavior in NaCl dissolution.Source: §II · p. 6, Figure 4 · p. 7
- At 3000 K, NEP89 predicts methane combustion pathways consistent with ANI-1xnr but a reaction rate constant of 2.90 × 10^15 cm3 mol−1 s−1, about six times larger.For protein–ligand dynamics, it captures binding and conformational preferences qualitatively, but its allowed Ramachandran area is lower than CHARMM36 and its free-energy landscape remains limited.Source: §II · pp. 6–7
Fine-tuning applications
- With 100 added configurations, fine-tuning improves MoSi2N4 fracture strength and Young’s modulus toward experiment; around 200 configurations also yield close metallic-glass Tg and improved icosahedral coordination.For MoSi2N4, predictions change from 272 ± 1 to 463 ± 4 GPa in fracture strength and from 34.2 ± 0.4 to 45.7 ± 0.2 GPa in Young’s modulus.Source: §II · pp. 8–9
- Fine-tuning with 104 configurations brings monolayer MoS2 thermal conductivity near customized-NEP and DFT-force-constant BTE results, while 119 structures improve benzene spectral agreement.For benzene, fine-tuning improves phonon-mode frequencies and relative peak intensities, though a broad feature near 100 meV remains red-shifted.Source: §II · pp. 8–9
III. DISCUSSION
NEP89 combines broad and efficient atomistic simulation with competitive static and dynamic predictions and rapid fine-tuning. Its current scope is bounded by missing long-range interactions and incomplete coverage of some complex systems.
Training data
- Descriptor-space subsampling, targeted data curation, and reference-energy adjustments produce a compact dataset balancing diversity and reliability for static and dynamic predictions.The dataset is described as flexible for future extension.Source: §III · p. 9
Results
- NEP89 often matches or outperforms representative foundation models across static and dynamic properties, with agreement with experiments and robustness across applications.Benchmarks include formation and binding energies, phonon frequencies, elastic moduli, liquid-water properties, and lithium-ion transport.Source: §III · p. 9
Large-scale applications
- Empirical-potential-comparable efficiency supports large-scale simulations, with quantitative agreement for yield strength and water-ion diffusivities and expected Arrhenius behavior in rocksalt dissolution.The demonstrations also cover complex alloys, methane combustion, and protein-ligand dynamics; protein structural features and interactions are captured qualitatively.Source: §III · p. 9
Fine-tuning
- Fine-tuning from limited training data improves agreement with quantum-mechanical calculations and experiments across specialized materials and properties.Demonstrations include MoSi2N4, metallic glass, monolayer MoS2, and crystalline benzene.Source: §III · p. 9
Limitations
- Without explicit long-range interactions, quantitative accuracy may be affected in ionic, aqueous, and large biomolecular systems; detailed free-energy landscapes for complex biomolecules remain challenging.The paper also attributes this limitation to incomplete training-data coverage of protein-ligand configurations.Source: §III · pp. 9–10
Scope and extensions
- The current model lacks variable-charge and magnetic descriptors, which may restrict applicability to systems with substantial charge transfer, strong magnetism, or external-field coupling.The authors identify charge-extended NEP as an existing route to adding environment-dependent charges and explicit long-range electrostatics, with magnetic descriptors as a future extension.Source: §III · p. 10
Future work
- The demonstrated applications are not exhaustive, and the authors anticipate further refinement and domain-specific models fine-tuned from NEP89.They frame these as future efforts to characterize capabilities and identify areas for improvement.Source: §III · p. 10
IV. METHODS
The methods describe computational-efficiency benchmarks, multiple atomistic simulation protocols, and fine-tuning workflows for application-specific models.
Static-property benchmarks
- Static-property benchmarks used gpumd-wizard, ASE, calorine, phonopy, and MatCalc, with detailed procedures deferred to Supplementary Sections 5–10.Atomic snapshots were generated with OVITO.Source: §IV · p. 10
Computational performance evaluation
- Computational efficiency was compared over 1,000 MD steps on a 20-element equiatomic alloy, scaling systems from the unit cell until memory limits were reached.The compared implementations used GPUMD for NEP89 and LAMMPS for MACE-MP-0, CHGNet, and M3GNet.Source: §IV · p. 10
MD simulation protocols
- Amorphous carbon, water, and thiophosphate electrolytes were simulated with MD protocols using systems of 512, 699, and 900 atoms, respectively.The water and electrolyte studies used multiple independent runs at each temperature or material-temperature combination.Source: §IV · p. 10
- Compression simulations used a 1,221,240-atom, ten-grain CuMoTaVW system, with yield strength determined by the 0.2% offset method.The system was equilibrated at 300 K before uniaxial compression along x to 10% strain.Source: §IV · pp. 10–11
- Dissolution, combustion, and protein-ligand dynamics were simulated using a 32-unit NaCl nanocrystal in 14,000 water molecules, reactive CH4/O2 mixtures, and a 33,917-atom protein complex.The dissolution study averaged ten runs per temperature; combustion products were averaged over five trajectories.Source: §IV · p. 11
Fine-tuning
- For MoSi2N4 and Pd42.5Cu30Ni7.5P20, fine-tuning used sampled trajectory structures labeled with DFT reference data and ran for 5,000 steps.The workflows then simulated a 44,800-atom MoSi2N4 monolayer and a 32,000-atom metallic-glass system.Source: §IV · pp. 11–12
- MoS2 fine-tuning used 104 sampled configurations, followed by thermal-conductivity calculations on an 11,484-atom model using homogeneous nonequilibrium MD.Ten independent 10 ns simulations were conducted.Source: §IV · p. 12
- Crystalline-benzene simulations used a 57,024-atom supercell, path-integral equilibration, and a 1 ns production run to compute the dynamic structure factor and neutron spectrum.The spectrum was obtained by weighting and integrating S(q, ω), with the TOSCA resolution function applied.Source: §IV · p. 12