Acid-Compatible Catalyst Development and Machine-Learning-Assisted Cell Optimization Toward Cost-Effective CO2 Reduction Reaction: Co-P4VP-Derived Catalyst and Nafion-Based Membrane Electrode Assembly

Abstract

Electrochemical CO2 reduction reaction (CO2RR) has gained significant attention as a possible carbon-neutral way of producing chemical feedstock in conjunction with renewable energy. The enormous efforts have been mainly focused on catalyst research, resulting in a spectrum of efficient catalysts for CO2RR into CO, an attractive feedstock due to the relatively preferable reaction conditions with two-electron and proton transfers. However, most catalysts were shown to promote CO2RR only in a CO2-saturated less-acidic aqueous solution or with an anion exchange membrane to suppress hydrogen evolution reaction, which impedes the implementation of a scalable and cost-effective nafion-based membrane electrode assembly (MEA). Another underlying challenge is that reported catalysts usually require multi-step and often complex synthetic conditions, which push the production costs. Additionally, the subsequent performance evaluation and cell optimization are often laborious because of a vast optimization space, which may delay the feasibility study of catalysts. Such constraints impose practical difficulty and raise the overall cost of CO2RR deployment.Here, we present scalable catalyst design and machine-learning-assisted cell optimization to address the challenges mentioned above. First, we report a straightforward one-pot synthesis of cobalt and organic [poly-4-vinylpyridine (P4VP)] precursors with carbon supports as a catalyst compatible with nafion-based MEA [1]. Electrochemical studies indicated that this catalyst performs CO2RR predominantly over HER across a wide range of bulk pH from 2 to 7 due to lower pH dependence of CO2RR reaction rate. The structural investigation revealed that CoO@Co nanoparticles and pyridine moieties co-exist, forming a local environment that is preferable for CO2RR. Pourbaix-diagram analysis indicates that concomitant reduction of Co2+ and pyridine moieties are likely involved in a reaction step, including proton-decoupled electron transfer. The initial optimization of the Co-P4VP-derived catalyst and MEA reached the remarkable performance of Faradaic efficiency (FE) of 92% at 85 mA/cm2, and a preliminary durability test showed stable FE for 20 h operation. We postulate that these outstanding performances reflect the synergistic effect of CoO and pyridine moieties that were achieved by coordinating P4VP to Co during the appropriate thermal process.To further optimize the MEA and accelerate the feasibility study cycle of the Co-P4VP-derived catalyst, we introduced a machine-learning-assisted, goal-oriented optimization procedure with energy efficiency (EE) and partial current density for CO2RR into CO (jCO2RR) as target features. This method includes data annotation, feature selection, model selection, data sorting, statistical stability test, and experimental validation. We selected the feature based on multiple non-linear correlation analyses, which highlighted the importance of both component and process parameters. While a conventional cross-validation technique did not provide a significant difference in model accuracy, the distribution of predicted data in the reverse engineering process heavily depended on the investigated models. Accordingly, we chose the model that best reflected experimental results and our understanding of the physical limitations of the MEA cell. The predicted data points were sorted with target-oriented criteria, including the production cost of CO, and several were selected based on a statistical stability test. The ensuing experiments successfully validated the effectiveness of this process as we achieved the promising performance values of 57% for EE and 115 mA/cm2 for jCO2RR with nafion-based MEA. Results from data feedback in the algorithm predicts those values are close to optimization within the given parameter space which was reached much more quickly than by conventional unguided use of machine learning.We expect this comprehensive study to serve as an effective example of how to incorporate exploratory catalyst design, analysis-based knowledge extraction, and ML-assisted cell optimization to accelerate material-based electrochemical advancement in developing areas such as CO2RR.[1] N. Fujinuma, A. Ikoma, S.E. Lofland, Adv. Energy Mater. 10 (2020) 2001645.