Benchmark of 23 ML interatomic potentials finds lightweight models are the practical choice
Machine learning interatomic potentials (MLIPs) have become one of the fastest-growing tools in computational materials science. They promise the accuracy of density functional theory (DFT) at a fraction of the cost, in principle. But the field has seen an arms race in model size, with parameter counts ballooning from hundreds of thousands to hundreds of millions. How much real-world benefit does that extra complexity deliver?
A team at the Chinese Academy of Sciences, Hanwen Kang, Tenglong Lu, and corresponding authors Sheng Meng and Miao Liu from the Institute of Physics in Beijing, has provided the first systematic answer. Their benchmark of 23 mainstream MLIPs, posted July 8 on arXiv, reveals a stark accuracy-efficiency trade-off: large state-of-the-art models improve accuracy by only 3–5 meV/atom over lightweight designs while sacrificing one to three orders of magnitude in computational throughput.
The 23 models
The benchmark covered 18 model families in multiple size variants, spanning the full spectrum of current MLIP architectures:
- Lightweight (0.5–5 million parameters): MatterSim v1 1M, Nequix MP PFT, M3GNet, MACE-small, GPTFF, CHGNet, SevenNet, ORB, GRACE, MatRIS, TACE, EquFlash, and NequIP-OAM-S
- Medium (5–10 million): MACE-MPA-0, Allegro-OAM-L, NequIP-OAM-M
- Large (10–730 million): Equiformer V3 (DNS-OAM), DPA4.0 Pro, DPA4-MP trajectory, PET-OAM-XL (PET-730M), eSEN 30M, NequIP-OAM-XL
All models were tested on a standardized benchmark: computing the phonon thermal conductivity of a fixed 192-atom LiCoO₂ cell, a widely studied cathode material for lithium-ion batteries, with DFT (GGA-PBE) as the reference.
The results
The accuracy difference between the best heavyweight models and the best lightweight ones was barely distinguishable at the scale relevant to real experimental work. Lightweight models such as MatterSim v1 1M and Nequix MP PFT achieved accuracy within 3–5 meV/atom of the largest models, a gap the authors describe as “smaller than room-temperature thermal noise” and “below the zero-point vibrational energy of a typical bond.”
The cost difference, by contrast, was enormous. Lightweight models ran hundreds to thousands of times faster than DFT. The largest models, those with hundreds of millions of parameters, sometimes ran less than two times faster than DFT itself. Memory constraints were equally severe: heavyweight models like DPA4.0 Pro, Equiformer V3, and PET-730M could simulate only 500 to 1,000 atoms on an 80 GB GPU, while lightweight models like Nequix could handle approximately 200,000 atoms under the same memory budget.
“The heavyweight models are limited to systems of 500 to 1,000 atoms,” the authors note. “That is not a material. It is a small nanoparticle.”
The Pareto frontier
The study’s central finding is that lightweight MLIPs occupy the Pareto frontier for practical molecular dynamics: no model in the benchmark simultaneously achieved both higher accuracy and higher throughput than the best lightweight designs.
The authors’ recommendations are explicit: for the vast majority of routine molecular dynamics work, studies of diffusion, phase transitions, mechanical properties, and reaction pathways in systems of tens of thousands of atoms, lightweight models are the right tool. Large models add marginal accuracy at prohibitive cost.
This is not to say large models have no role. The authors note that for “high-precision electronic-structure-level prediction of static properties,” single-point energy calculations where throughput is irrelevant, large models may still be useful. And for benchmarking and methodology development, understanding the full accuracy ceiling of different architectures remains valuable.
The benchmark also revealed a cross-platform overhead problem: even the fastest models ran only 1.11–1.4 times faster in the specialized TorchSim framework compared to the generic Atomic Simulation Environment (ASE) pipeline, suggesting that software infrastructure, not just model architecture, constrains real-world MLIP performance.
Broader context
The MLIP field has grown explosively, driven by the promise of bridging the gap between first-principles accuracy and classical-method speed. This benchmark suggests that for many applications, the field may already have reached that goal, but with the lightweight models, not the headline-grabbing large ones.
The paper’s findings underscore a recurring pattern in AI applied to science: bigger is not always better, and the best model for the job depends on the job itself.
Sources:
1. Kang, H., Lu, T., Meng, S. & Liu, M. “Are Machine Learning Interatomic Potentials Truly Practical? A Benchmark of 23 Mainstream Models.” arXiv:2607.07647 (2026).
2. Institute of Physics, Chinese Academy of Sciences, Beijing, China.