Validation of OC25-trained models for interfacial properties and reactivity

Determine whether machine learning interatomic potentials trained on the Open Catalyst 2025 (OC25) dataset can accurately predict interfacial properties and reactivity at solid–liquid interfaces, given that OC25 configurations use relatively shallow solvent layers and higher ion concentrations than many experimental conditions.

Background

The OC25 dataset introduces large-scale density functional theory data for solid–liquid interfaces with explicit solvents and ions, enabling training of machine learning models to simulate catalytic transformations. However, the dataset employs relatively shallow solvent layers and higher ion concentrations to manage computational costs, which differ from many experimental conditions.

The authors state that while models trained on OC25 may still be able to predict interfacial properties and reactivity accurately, this has not yet been tested, leaving an explicit uncertainty about model validity under these interface conditions. Establishing this would confirm the dataset’s utility for realistic electrocatalytic simulations and guide future model and dataset design.

References

Although the models in this work may still be able to accurately predict interfacial properties and reactivity, these aspects remain to be tested in future studies.

— The Open Catalyst 2025 (OC25) Dataset and Models for Solid-Liquid Interfaces  (2509.17862 - Sahoo et al., 22 Sep 2025) in Section 4: Outlook and future directions

Whether the giant, MPB-derived response survives at such an interface over months of continuous wear is, at present, an open question that bulk computation cannot answer.

— Lead-free piezoelectric perovskites for arterial-pulse e-skin: from configurational complexity to equivariant machine-learning potentials  (2609.00580 - Ismail et al., 1 Sep 2026) in Section 7.3, “Open Challenge: Machine-Learning Potentials for the Wet, Dynamic Skin Interface”