[Paper Review] Benchmarking short-range machine learning potentials for atomistic simulations of metal/electrolyte interfaces

Lucas B. T. de Kam, Jiaxin Zhu+3 more|arXiv (Cornell University)|Feb 26, 2026

Machine Learning in Materials Science0 citations

TL;DR

This paper benchmarks short-range MLIPs (DP, DP-MP, ACE/GRACE, MACE, and eSEN-OC25) for charged Au/water interfaces to assess their ability to reproduce interfacial water orientation and ion distributions across various surface charges.

ABSTRACT

Atomistic simulations of electrochemical interfaces remain challenging due to the long time scales required to adequately sample the structure of the electric double layer. The emergence of efficient, short-range machine learning interatomic potentials (MLIPs) offers a promising alternative to computationally expensive density functional theory-based molecular dynamics (DFT-MD) simulations in this regard. However, in standard periodic DFT calculations of metal surfaces, the surface charge is implicitly set by the number of counterions in the simulation cell, making it a global property that is difficult to represent with strictly local MLIPs. Here, we benchmark common MLIP architectures (DP, ACE, MACE) for charged Au/water interfaces containing solvated sodium ions. We find that MLIPs trained on datasets spanning multiple surface charge states yield inconsistent predictions of interfacial water orientation and ion distributions, although message-passing models with a larger receptive field exhibit greater robustness to training on mixed-charge datasets. In contrast, models trained on a single charge state produce consistent equilibrium interfacial properties. Finally, we assess the performance of the eSEN model trained on the recently released Open Catalyst 2025 dataset, which includes solid/liquid interfaces that span a wide range of surface charge densities. Overall, our results characterize the limitations of short-range MLIPs for simulations of electrochemical interfaces and provide practical guidance for constructing training datasets for simulations of charged metal/electrolyte interfaces.

Motivation & Objective

• Assess how well short-range MLIPs reproduce interfacial water orientation and ion distributions at metal/electrolyte interfaces under different surface charges.
• Evaluate whether a single MLIP can generalize across multiple surface charge states.
• Compare local vs. semilocal and equivariant MLIPs in terms of accuracy, robustness, and transferability.
• Provide practical guidelines for constructing training datasets for simulations of charged metal/electrolyte interfaces.

Proposed method

• Benchmark five MLIPs (DP, DP-MP, GRACE-1L/ACE, MACE, eSEN-OC25) on Au/water interfaces with 0–4 solvated Na+ ions.
• Train on datasets spanning mixed-charge and specific-charge states, plus OC25-based model, using DFT-labeled training data from VASP RPBE-D3 calculations.
• Assess models via interfacial water orientation and ion distributions, using RMSE for energies/forces and MD-level property evaluations.
• Analyze receptive field, body order, and equivariant vs. non-equivariant architectures to understand locality vs. long-range needs.
• Perform MD simulations in NVT with explicit ions and water, and compare trajectory-derived properties (density profiles, water dipole orientation) to reference data.

Experimental results

Research questions

• RQ1 Can short-range MLIPs accurately describe interfacial structure and ion distributions at charged metal/electrolyte interfaces?
• RQ2 Do MLIPs trained on mixed-charge datasets generalize to specific charge states and vice versa?
• RQ3 How do model architecture choices (local vs. message-passing vs. equivariant) affect robustness and transferability across surface charges?
• RQ4 Is a single MLIP viable to describe multiple surface charges, or are charge-specific trainings required?
• RQ5 How does an OC25-trained model perform relative to models trained on explicit DFT data for charged interfaces?

Key findings

• MLIPs trained on mixed-charge datasets show inconsistent predictions of interfacial water orientation and ion distributions across surface charges.
• Message-passing models with larger receptive fields demonstrate greater robustness to mixed-charge training data.
• Models trained on a single charge state yield consistent equilibrium interfacial properties.
• The OC25-based eSEN model provides a reference point from a large, diverse OC25 dataset spanning various surface charges.
• Overall, short-range MLIPs have limitations for charged electrochemical interfaces, and training set design critically affects accuracy and transferability.