Benchmarking short-range machine learning potentials for atomistic simulations of metal/electrolyte interfaces

TL;DR

It is found 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.

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.

Figures (19)

• Figure 1: Body order of local models and message-passing models. The black node is the central atom. The graphs represent interactions simultaneously considered in the model. Each message-passing layer makes the graph branch into $\nu$ neighbors, where $\nu$ is the correlation order of the layer. The effective body order is the total number of nodes in the graph. Note that this is a schematic graph illustrating the body order, not a graph representing the full atomistic structure, which connects all atoms within the cutoff radius.
• Figure 2: Accuracy and performance of the deep potential (DP), deep potential with message-passing (DP-MP), 1-layer GRACE (GRACE-1L) and MACE using the neutral-surface dataset. (a-b): root-mean-squared error (RMSE) of the energy (a) and force components (b) for models trained on a varying number of structures and evaluated on a test set containing 200 structures. Shaded areas indicate the minimum and maximum results over three different data splits; the markers indicate the mean. MACE was repeated only once. (c) Molecular dynamics speed for an Au/water interface with 384 atoms and a 0.5fs timestep on an NVIDIA A100 GPU.
• Figure 3: Water density profiles in the direction perpendicular to the gold surface from 100ps trajectories driven by models trained on different amounts of training data, $n$, for the neutral-surface dataset. For each model and training set, the mean of three repeats is shown (in color) with models trained on different subsets of the data and a different initial configuration. The min/max deviation over the three repeats is within the plotted line thickness. If trajectories were unstable, they are not included, and instead the number of unstable trajectories is indicated next to the red $\times$ in the density profile plots. Black lines indicate the mean over three 10ps DFT-MD trajectories.
• Figure 4: Water structure at a neutral gold surface for models trained only on the target system ('specific') vs. models trained on a dataset with differently charged surfaces ('mixed'). (a) Energy RMSE on the neutral Au/water test set with 2690 structures. (b) Water density profiles and (c) dipole orientation profiles obtained from 1ns molecular dynamics trajectories. $\theta$ is the angle between the water bisector and the surface normal. Black line shows the mean and min/max spread over three 10ps DFT-MD simulations. (d) Total dipole ($P_z$) histograms from molecular dynamics. The mean of the distribution is indicated. The combined result from three 10ps DFT-MD trajectories is shown in gray and reproduced in all plots.
• Figure 5: Number of Na^+ ions within the receptive field of interfacial water molecules. Each subplot shows the $x,y$ positions of oxygen atoms from frames of MACE trajectories; marker transparency depicts the distance $\Delta z$ from the surface with $\Delta z = 3\angstrom$ corresponding to the least transparent markers and $\Delta z = 7.5\angstrom$ corresponding to the most transparent markers. Marker color indicates the number of Na^+ ions within the receptive field of each oxygen atom. Open markers denote the positions of the Na^+ ions. The receptive field radius and the total number of sodium ions in the simulation cell are indicated in the top-left and top-right corners of each subplot, respectively.