Energetically Anchored Machine-Learning Interatomic Potential Embeddings for Reliable Hydrogen Evolution...

Area of Science:

• Computational materials science
• Catalysis
• Machine learning for materials discovery

Background:

• Accurate prediction of hydrogen adsorption free energy (ΔG_H) is critical for advancing electrocatalyst design for the alkaline hydrogen evolution reaction (HER).
• Current machine-learning (ML) approaches face challenges due to inconsistent energy definitions, varied density functional theory (DFT) protocols, and poor out-of-distribution (OOD) generalization.
• Developing robust ML workflows is essential for efficient electrocatalyst screening.

Purpose of the Study:

• To present an energetically anchored ML framework for predicting DFT-defined hydrogen adsorption energetics on diverse catalyst surfaces.
• To integrate MLIP-derived energetic descriptors with CHGNet latent embeddings for enhanced predictive accuracy.
• To improve the consistency and generalization of ML models for HER electrocatalyst discovery.

Main Methods:

• Developed an ML framework combining MLIP energetic descriptors and CHGNet latent embeddings.
• Employed protocol-consistent MLIP energy evaluations on reference geometries for thermodynamically aligned descriptors.
• Utilized gradient-boosting regression to combine energetic and structural descriptors.
• Curated datasets from Catalysis Hub and AQCat25, including single-, bi-, and trimetallic adsorption systems.

Main Results:

• The embedding-augmented model achieved high accuracy (R² = 0.976, MAE = 0.054 eV, RMSE = 0.105 eV) at the site level.
• The energetically anchored descriptor was the dominant factor, with structural descriptors providing refinement, especially near the thermoneutral regime.
• The framework maintained meaningful adsorption-energy ranking even with MLIP-relaxed geometries.
• Predictive robustness was sensitive to the balance between compositional diversity and protocol consistency.

Conclusions:

• Energetically anchored MLIP embeddings offer an effective strategy for scalable post-DFT adsorption-energy refinement.
• The framework enables data-efficient electrocatalyst screening for HER.
• Clarified practical limitations of MLIP-driven workflows in heterogeneous catalysis, emphasizing protocol consistency for OOD performance.