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

This paper benchmarks short-range machine learning potentials for charged metal/electrolyte interfaces, revealing that while models trained on single charge states yield consistent results, those trained on mixed-charge datasets produce inconsistent predictions, thereby highlighting the limitations of local architectures and offering practical guidance for dataset construction in electrochemical simulations.

The Gist

Imagine you are trying to simulate a bustling party at the edge of a swimming pool. The "pool" is a metal surface (like gold), and the "party guests" are water molecules and dissolved salt ions (sodium).

In the real world, the metal surface can be electrically charged, like a magnet. This charge changes how the water molecules stand up or lie down and how the salt ions crowd around. To understand this, scientists usually use a super-accurate but incredibly slow method called DFT (Density Functional Theory). It's like trying to film every single molecule's movement with a high-speed camera, but it's so slow you can only watch a split second of the party before the film runs out.

To speed things up, scientists have developed Machine Learning Potentials (MLIPs). Think of these as "AI interns." You show them a few hours of the high-speed footage, and they learn the rules of the party. Then, they can simulate the party for days or weeks in the blink of an eye.

The Problem:
The paper investigates a specific glitch in these AI interns. In a real electrochemical cell, the "charge" of the metal isn't just a local thing; it's a global property. It depends on how many salt ions are in the entire room.

However, most AI interns are designed to be local. They only look at their immediate neighbors (like a person at a party only talking to the people standing within arm's reach). They don't know how many people are in the room overall.

The Experiment:
The researchers tested several different types of AI interns (called DP, ACE, MACE, etc.) on a gold/water/salt simulation. They asked two main questions:

1. Can one AI learn to handle any charge? (e.g., Can the same AI learn the party rules for a neutral metal, a positively charged metal, and a negatively charged metal all at once?)
2. Does it matter if we train the AI on just one specific charge?

The Findings (The "Aha!" Moments):

• The "Mixed Bag" Failure: When they trained the AI on a "mixed" dataset (containing examples of neutral, positive, and negative surfaces), the AI got confused. Because it only looks at its immediate neighbors, it couldn't tell the difference between a neutral room and a charged room if the extra salt ions were just outside its "arm's reach."

  • The Analogy: Imagine a security guard who only looks at the people standing right next to him. If he sees a group of people, he can't tell if the whole building is empty or full of people just by looking at that one group. So, he tries to guess the average behavior, which leads to mistakes. The AI predicted the water molecules were standing up or lying down in the wrong way because it couldn't "see" the global charge.

• The "Specialist" Success: When they trained the AI on only one specific charge (e.g., only negative surfaces), the AI became a specialist. It learned the exact rules for that specific party perfectly. It predicted the water and ion positions accurately.

  • The Analogy: If you hire a guard who only watches the "Negative Charge Party," he learns the specific dance moves for that night perfectly. He doesn't get confused by other types of parties.

• The "Long-Range" Hero: Some newer, smarter AI models (like MACE) have a slightly longer "arm's reach" (a larger receptive field). They could see a bit further than the others. While they were still not perfect when trained on mixed data, they were much more robust and made fewer mistakes than the short-sighted models.

• The "Big Data" Model: They also tested a model trained on a massive, pre-made dataset called OC25 (which has millions of different surface types). This model did a decent job, likely because its huge size and long "reach" helped it generalize better, but it still struggled slightly compared to the "Specialist" models trained on the specific problem.

The Takeaway:

If you want to simulate a charged metal surface in water:

1. Don't try to build one "Super AI" that handles every possible charge. It's too hard for these short-range models to understand the "big picture" (global charge) just by looking at their neighbors.
2. Train a "Specialist AI" for the specific charge you care about. If you are studying a negatively charged battery electrode, train your AI only on negative surfaces. It will give you accurate, reliable results for that specific scenario.
3. Beware of the "Local Blind Spot." If you force a local AI to guess the rules of a global game, it will hallucinate weird behaviors (like water molecules standing up when they should be lying down).

In Summary:
This paper is a warning label for scientists. It says: "These fast AI tools are great, but they have a blind spot regarding global electric charges. To get good results, don't ask them to learn everything at once. Instead, give them a specific job and train them specifically for that job."

Technical Summary

1. Problem Statement

Atomistic simulations of electrochemical interfaces (e.g., metal electrodes in electrolytes) are critical for understanding processes like electrocatalysis. However, they face two major bottlenecks:

• Computational Cost: Density Functional Theory-based Molecular Dynamics (DFT-MD) is accurate but limited to small system sizes (<1,000 atoms) and short timescales (<100 ps). Adequately sampling the Electric Double Layer (EDL) and equilibrating solvated ions often requires nanosecond timescales, which is beyond the reach of standard DFT-MD.
• The Global Charge Problem: Machine Learning Interatomic Potentials (MLIPs) offer a faster alternative by learning from DFT data. However, standard MLIPs rely on the locality assumption, where an atom's energy depends only on its local environment within a cutoff radius.
  • In periodic DFT simulations of charged interfaces, the net surface charge is a global property determined by the total number of counterions in the simulation cell.
  • Short-range MLIPs struggle to "see" these counterions if they fall outside the local cutoff, making it difficult to learn the correct electrostatic response and structural behavior across different surface charge states.

The paper aims to benchmark whether short-range MLIPs can accurately simulate charged metal/electrolyte interfaces, specifically investigating the trade-offs between training on single charge states vs. mixed charge states, and the impact of model architecture (local vs. message-passing).

2. Methodology

Models Benchmarked

The study evaluates five distinct MLIP architectures, varying in body order (many-body correlations) and receptive field (local vs. semilocal):

1. Deep Potential (DP): A local model using 3-body descriptors.
2. DP-MP: A message-passing variant of DP, extending the receptive field.
3. GRACE-1L (ACE): A local model based on Atomic Cluster Expansion (ACE) with high body order ($\nu=4$).
4. MACE: A message-passing equivariant GNN with high body order ($\nu=3$ per layer, effective body order 13).
5. eSEN-OC25: A pre-trained model on the Open Catalyst 2025 dataset (containing ~8M solid/liquid interface configurations with varying charges).

Systems and Datasets

• System: Au(111) slab in contact with water and varying numbers of solvated Sodium ions (Na$^+$), representing different surface charge states (0 to 4 ions).
• Training Datasets:
  • Specific: Trained only on a single charge state (e.g., neutral Au/water OR Au/water/3Na$^+$).
  • Mixed: Trained on a dataset containing all charge states (0–4 Na$^+$) simultaneously.
• Reference: DFT-MD (RPBE functional) used for training labels and validation.
• Simulation: NVT ensemble at 300 K, running up to 1–2 ns trajectories to analyze water orientation and ion distribution.

Evaluation Metrics

• Static Accuracy: Root-mean-squared error (RMSE) for energy and forces.
• Dynamic Stability: Ability to run stable MD trajectories without energy drift.
• Interfacial Properties:
  • Water density profiles.
  • Water dipole orientation (angle of water bisector relative to surface normal).
  • Spatial distribution of Na$^+$ ions.
  • Total dipole moment fluctuations ($P_z$).

3. Key Contributions & Results

A. Accuracy vs. Computational Cost

• Data Efficiency: MACE demonstrated superior data efficiency. A MACE model trained on only 50 structures outperformed a DP model trained on 1,000 structures in terms of energy RMSE.
• Receptive Field: Message-passing models (DP-MP, MACE) generally achieved lower errors than local models (DP, GRACE-1L) when trained on large datasets, benefiting from larger receptive fields (~10 Å vs. 6 Å).
• Speed: Local models (DP) were ~20x faster than MACE. However, DP-MP and GRACE-1L offered a middle ground, being ~8–9x faster than MACE while maintaining reasonable accuracy.

B. The "Mixed Dataset" Failure Mode

The most significant finding concerns training on mixed charge states:

• Inconsistency: Models trained on mixed datasets (0–4 ions) produced inconsistent and inaccurate predictions for interfacial water orientation and ion distributions compared to models trained on specific (single) charge states.
• Water Orientation: Mixed-data models exhibited an anomalously strong preference for "H-down" water orientation at neutral surfaces and weaker orientation at charged surfaces.
• Cause: The surface charge is a global property. Local descriptors (cutoff ~6 Å) cannot "see" the counterions if they are beyond the cutoff. Consequently, the model learns an average behavior that fails to distinguish specific charge states.
• Message-Passing Mitigation: Models with larger receptive fields (MACE, DP-MP, ~10 Å) were more robust to mixed training than local models but still exhibited errors. They could not fully parameterize the global charge even when the cutoff covered most of the interface.

C. Success of Specific Training

• Reliability: Models trained on a single specific charge state (e.g., only neutral or only 3 Na$^+$) produced consistent and accurate equilibrium properties (water structure, ion distribution) across all architectures.
• Implication: For a specific electrochemical condition, training a dedicated model is currently the most reliable approach using short-range MLIPs.

D. Ion Distribution and Uncertainty

• Ion Positioning: Specific models agreed well on ion positions (peaks at ~3.4 Å and ~4.6 Å). Mixed models showed significant divergence depending on the architecture.
• Uncertainty Estimation: The study utilized a committee model approach to estimate uncertainty. Mixed-dataset models showed high uncertainty (large standard deviation) in ion density profiles, serving as a diagnostic tool to detect when a short-range model is failing to capture global charge effects.

E. Pre-trained Models (eSEN-OC25)

• The eSEN model, trained on the massive OC25 dataset (spanning wide charge densities), performed well for neutral interfaces but showed deviations for charged interfaces (favoring ions in the first water layer, similar to mixed-data failures).
• This suggests that even large foundational models struggle with the specific global charge dependency of electrochemical interfaces unless explicitly designed to handle it.

4. Significance and Outlook

• Practical Guidance: The paper provides a clear workflow for researchers: Do not rely on a single "universal" short-range MLIP trained on mixed charges for electrochemical interfaces. Instead, train specific models for the target surface charge to ensure accuracy.
• Limitations of Locality: It highlights a fundamental limitation of current short-range MLIPs: they cannot inherently encode global electrostatic constraints (like total surface charge) without explicit long-range terms or global information.
• Future Directions:
  • Constant-Potential MLIPs: The authors suggest that future models should directly access global charge information (e.g., via constant-potential MD where the Fermi level is predicted) rather than relying solely on local descriptors.
  • Hybrid Approaches: Combining short-range MLIPs for local chemistry with explicit long-range electrostatics (Ewald summation) or continuum models for the bulk electrolyte.
  • Data Efficiency: The high data efficiency of MACE allows for the use of more expensive, higher-accuracy QM methods (beyond DFT) for training, potentially improving the quality of the underlying physics.

In summary, while short-range MLIPs are powerful tools for accelerating electrochemical simulations, their application to charged interfaces requires careful dataset construction. The study establishes that specific training yields reliable results, whereas mixed training introduces significant errors due to the inability of local descriptors to capture global charge states.