Statistical AI Enables Precise Screening of Multielement Catalysts

Area of Science:

• Materials Science
• Computational Chemistry
• Catalysis

Background:

• Discrepancies between simplified computational models and complex synthesized materials hinder accurate performance prediction.
• Existing methods often rely on single-structure models, failing to capture the full performance potential of catalytic materials.

Purpose of the Study:

• To introduce a novel statistical descriptor (φ) for more accurate quantification of macroscopic catalytic activity.
• To bridge the gap between theoretical predictions and experimental performance in materials science.
• To accelerate the discovery of efficient catalysts for the oxygen evolution reaction.

Main Methods:

• Development of the Seq-Equiformer model, a graph neural network integrating LSTM with EquiformerV2 for dynamic structural analysis.
• Prediction of overpotentials for 250 million 3d transition metal doped CoOOH structures.
• Calculation of the statistical descriptor φ for element combinations and identification of optimal dopants.
• Application of Bayesian optimization-driven AI experiments for fine-tuning catalyst compositions.

Main Results:

• Identification of six optimal dopant combinations with the highest φ values.
• Discovery of a highly efficient catalyst (Mn_0.07Fe_0.09Ni_0.14Cu_0.01Co_0.69OOH) with a low overpotential (246.5 mV at 100 mA cm^-2) and excellent stability.
• The statistical descriptor achieved 80% accuracy in identifying top catalysts, a 30% improvement over traditional single-structure screening.

Conclusions:

• The statistical descriptor φ provides a more accurate assessment of catalytic activity by considering the distribution of high-activity configurations.
• Integration of statistical modeling, machine learning, and autonomous experimentation significantly accelerates catalyst discovery and enhances prediction accuracy.
• This approach offers a powerful strategy for designing next-generation catalysts for energy applications.