[Paper Review] Performance of universal machine learning potentials in global optimization

Edan T. Marcial, L. N. Chaudhary+2 more|arXiv (Cornell University)|Feb 26, 2026

Machine Learning in Materials Science0 citations

TL;DR

This study benchmarks nine universal machine learning potentials (uMLPs) in unconstrained global structure searches across diverse inorganic systems to evaluate their ability to locate DFT-ground-state structures.

ABSTRACT

Rapid development of universal machine learning potentials (uMLPs) and expansion of training data sets are reshaping the state of the art in atomistic simulation, highlighting the need for concurrent systematic benchmarking of their capabilities. Global optimization is among the most demanding uMLP applications because unconstrained exploration includes probing motifs not present in reference sets. We examined the latest generation of uMLPs in unconstrained evolutionary searches to assess whether these models can consistently predict complex crystal structure ground states across diverse inorganic systems. Our findings demonstrate that the considered M3GNet, MACE, SevenNet, EquiformerV2, MatterSim, GRACE, eSEN, Orb-v3, and PET-MAD models span a wide performance range, from near ab initio to essentially non-predictive, in their ability to resolve competing phases within low-energy basins. Additional tests on hcp-Zn, MB$_4$ (M = Cr, Mn, and Fe), and LiB$_{y}$ ($y\approx 0.9$) ground states reveal that several uMLPs capture fine energy differences arising from subtle electronic structure features.

Motivation & Objective

• Assess how well current uMLPs locate ground-state crystal structures in unconstrained global searches across diverse chemistries.
• Evaluate robustness of uMLPs in exploring low-energy basins beyond training data.
• Compare surrogate-potential-driven results to reference DFT across multiple functionals.
• Identify strengths and failure modes of different uMLP architectures in structure prediction tasks.

Proposed method

• Use out-of-the-box pretrained uMLPs (M3GNet, MACE, SevenNet, EquiformerV2, MatterSim, GRACE, eSEN, Orb-v3, PET-MAD) without fine-tuning.
• Perform zero-temperature evolutionary searches with MAISE to generate populations of candidate structures.
• Relax candidates with each uMLP and re-optimise with DFT (PBE/PBEsol/r2SCAN) for ground-truth comparison.
• Compute structure and energy proximity metrics by merging minima pools across uMLPs and re-relaxing with DFT.
• Assess ranking fidelity using a ranking RMSE that subtracts average energy shifts between uMLP and DFT pools.
• Analyze how well uMLPs reproduce known phases and identify any spurious minima.

Figure 1: Stability of the tI10-Na 2 CN 2 , mS28-MgB 3 C 3 , and oI28-MgB 3 C 3 phases identified in this work relative to reported mS10-Na 2 CN 2 and hP14-MgB 3 C 3 evaluated with common DFT functionals.

Experimental results

Research questions

• RQ1 Can current uMLPs reliably resolve competing low-energy crystal structures in unconstrained global searches across diverse inorganic chemistries?
• RQ2 How do surrogate PES landscapes produced by different uMLPs compare to DFT in terms of energetic ordering and structural proximity of minima?
• RQ3 Which uMLPs best reproduce ground-state motifs in challenging cases (e.g., Zn c/a anomaly, MB4 compounds, off-stoichiometric LiBx phases)?
• RQ4 What are common failure modes of uMLPs in global optimization (e.g., spurious minima, misranking, vdW or stacking issues) and how might they be mitigated?

Key findings

• uMLPs span a wide performance range from near ab initio to essentially non-predictive in resolving competing phases within low-energy basins.
• Several uMLPs (notably eSEN and some mid-to-large architectures) can capture fine energy differences arising from subtle electronic structure features in ground-state predictions.
• Some compounds (e.g., Zn c/a anomaly, MB4 derivatives, LiBx off-stoichiometries) reveal specific weaknesses such as misranking, overestimated interlayer spacings, or spurious low-energy minima.
• Surrogate-relaxation energy landscapes show modest but systematic deviations from DFT upon full relaxation, enabling potential hybrid workflows (NN relaxation with single-shot DFT energies).
• Merged pools across nine uMLPs allow consistent assessment of proximity to DFT minima and ranking fidelity, with ranking RMSE values typically in the range of a few to a few tens of meV/atom for relevant cases.
• uMLPs generally perform better than prior system-specific Behler–Parrinello NN potentials for the studied M-Sn binaries in terms of ranking accuracy.

Figure 2: Performance metrics on merged pools assessed relative to the reference DFT method, PBEsol for PT and PBE for the rest, and averaged over 11 compounds, excluding AgClO 4 . The schematics at the top clarify the definitions of the proximity and ranking metrics introduced in the text. The repr