Unveiling Catalytic Design Principles for C 2 Product Formation in CO 2 Electroreduction via Machine Learning

ABSTRACT Electrochemical CO 2 reduction (eCO 2 RR) offers a promising route for converting waste CO 2 into valuable fuels and chemicals, yet rational catalyst design remains challenging due to complex reaction pathways and the vast material space. Here, we develop two interpretable machine learning frameworks that bridge molecular‐scale and lab‐scale insights: a DFT‐ML model trained on reaction energy diagrams data for key intermediates ( * CO, * COOH, * CHO), and an Expt.‐ML model trained on experimental data points spanning four major products (CO, HCOOH, C 2 H 4 , H 2 ). Through systematic feature ablation and nested cross‐validation, we identify that combining just 5–7 physically meaningful descriptors achieves accuracy statistically equivalent to full 22‐feature models. Further, the “coupling” of both models by incorporating DFT‐ML predicted reaction energies as a descriptor into the Expt‐ML model led to improved accuracy in predicting the Faradaic efficiency, although these descriptors were themselves only moderately accurate. This suggests that the inclusion of features that already contain information toward the performance of the catalysts could help with training better models even with datasets that are skewed and smaller datasets. The Expt.‐ML model uncovers a fundamental competition between CO and HCOOH at lower overpotentials, while C 2 H 4 formation is governed by a hierarchy of factors including mass transport related. Notably, copper emerges as uniquely capable for multi‐carbon formation, but our analysis suggests that morphological engineering of pure copper surfaces may be more effective than alloying strategies. From the predictions of both the models, it is evident that CO‐producing catalysts could be further explored for the deeper reduction of CO 2 by the exploration of other conditions, which has not been explored thus far. This work establishes a scalable, data‐driven pathway for accelerating catalyst discovery while maintaining physical interpretability.