[Paper Review] The Open Catalyst 2025 (OC25) Dataset and Models for Solid-Liquid Interfaces
OC25 provides the largest diverse solid-liquid interface dataset with 7.8 million DFT calculations across 1.5 million unique systems, enabling baseline MLIP models with state-of-the-art accuracy for energy and forces in solvated interfaces.
Catalysis at solid-liquid interfaces plays a central role in the advancement of energy storage and sustainable chemical production technologies. By enabling accurate, long-time scale simulations, machine learning (ML) models have the potential to accelerate the discovery of (electro)catalysts. While prior Open Catalyst datasets (OC20 and OC22) have advanced the field by providing large-scale density functional theory (DFT) data of adsorbates on surfaces at solid-gas interfaces, they do not capture the critical role of solvent and electrolyte effects at solid-liquid interfaces. To bridge this gap, we introduce the Open Catalyst 2025 (OC25) dataset, consisting of 7,801,261 calculations across 1,511,270 unique explicit solvent environments. OC25 constitutes the largest and most diverse solid-liquid interface dataset that is currently available and provides configurational and elemental diversity: spanning 88 elements, commonly used solvents/ions, varying solvent layers, and off-equilibrium sampling. State-of-the-art models trained on the OC25 dataset exhibit energy, force, and solvation energy errors as low as 0.1 eV, 0.015 eV/Å, and 0.04 eV, respectively; significantly lower than than the recently released Universal Models for Atoms (UMA-OC20). Additionally, we discuss the impact of the quality of DFT-calculated forces on model training and performance. The dataset and accompanying baseline models are made openly available for the community. We anticipate the dataset to facilitate large length-scale and long-timescale simulations of catalytic transformations at solid-liquid interfaces, advancing molecular-level insights into functional interfaces and enabling the discovery of next-generation energy storage and conversion technologies.
Motivation & Objective
- Bridge the gap in MLIPs for solid-liquid and electrified interfaces by providing a large, diverse, explicitly solvated dataset.
- Evaluate baseline graph neural network models on OC25 to establish performance benchmarks for energies, forces, and solvation properties.
- Investigate the impact of DFT force convergence and drift corrections on MLIP training and evaluation.
- Provide openly available dataset, models, and code to accelerate long-time-scale simulations of interfacial catalysis.
Proposed method
- Generate configurations from adsorbate+surface in vacuum, then construct solvated interfaces with explicit solvents and ions.
- Perform short-time MD with high temperature and relaxations, followed by DFT single-point or short AIMD calculations using RPBE+D3 in VASP.
- Filter training data using a force drift threshold (1 eV/Å) to ensure consistent force labels.
- Train baseline MLIP models (UMA and eSEN variants) including both energy-conserving and direct-force formulations, plus fine-tuned UMA baselines.
- Evaluate models on OC25 splits including standard and out-of-distribution (OOD) sets for solvents, ions, and both, reporting MAEs for energy and forces.
- Assess the influence of force convergence settings on model performance and report parity plots and error analyses.
Experimental results
Research questions
- RQ1Can OC25 train MLIPs that accurately predict energies and forces for solid-liquid interfaces with explicit solvent and ion effects?
- RQ2How do solid-liquid interface properties (solvation energy, adsorption energetics) behave across diverse solvents, ions, and surface chemistries in OC25?
- RQ3What is the impact of DFT force convergence and drift corrections on MLIP training and evaluation at solid-liquid interfaces?
- RQ4How well do out-of-distribution solvents and ions generalize in OC25-trained models?
- RQ5What are the best-performing model configurations (size, energy-conserving vs direct-force) for OC25 tasks?
Key findings
- OC25 contains 7,801,261 single-point DFT calculations over 1,511,270 unique systems and 88 elements, with explicit solvent environments.
- Energy, force, and solvation energy errors for state-of-the-art models reach as low as 0.1 eV, 0.015 eV/Å, and 0.04 eV respectively, outperforming UMA-OC20 baselines.
- Energy-conserving models generally outperform direct-force models in both energy and force predictions across splits.
- Solvation energy errors are typically lower than test energy errors, suggesting potential error cancellation for relative properties.
- Models trained on OC25 show resilience to moderate label noise; forcing convergence thresholds (drift) significantly affect force accuracy, guiding data curation decisions.
- Out-of-distribution evaluations indicate larger errors for unseen solvents and ions, highlighting opportunities to improve generalization.
Read this paper
Open the full text in the Nubint reader and ask about any passage as you read.
Read this paper on NubintThis review was generated by AI from the original paper. Please check the paper before citing it.