Data-driven and interpretable machine learning for performance-determining interactions governing C-C coupling and...

Area of Science:

• Electrochemistry
• Materials Science
• Machine Learning

Background:

• Efficient electrochemical conversion of carbon dioxide (CO2) into multi-carbon products (C2+) is challenging due to complex catalyst and environmental interactions.
• Copper-based electrocatalysts show promise but require optimized conditions for high C2+ selectivity.

Purpose of the Study:

• To develop an interpretable machine learning framework to analyze trends in C2+ selectivity for Cu-based electrocatalysts.
• To identify key descriptors influencing CO2 electroreduction (CO2RR) performance.

Main Methods:

• A literature-curated dataset of 380 experimental entries was compiled, including catalyst morphology, particle size, applied potential, and electrolyte composition.
• Multiple regression models were benchmarked, with LightGBM achieving the highest performance (R2 ≈ 0.78) for predicting C2+ selectivity.
• SHAP analysis was used for feature attribution and interaction analysis.

Main Results:

• Applied potential, particle size, and catalyst morphology were identified as dominant descriptors for C2+ selectivity.
• Electrolyte and membrane composition showed secondary but significant contributions.
• Machine learning models suggested optimal C2+ selectivity near moderately negative potentials and with catalyst dimensions under 100 nm.

Conclusions:

• This work provides an interpretable framework for analyzing fragmented CO2RR literature data.
• Identified descriptor-environment relationships can guide future catalyst design and operating condition optimization for enhanced C2+ production.
• The findings offer data-supported trends for designing highly selective electrocatalysts.