Closed-loop, machine learning–driven optimization of reactor yields in reactive carbon electrolyzers

Abhishek Soni, Siwei Ma, Andrew Jewlal, Mehrdad Mokhtari, Ribwar Ahmadi, Daniel Lin, Karry Ocean, Giuseppe V. Crescenzo, Kevan Dettelbach, Nada Nasr, Curtis Berlinguette

Keywords: Reactive carbon capture, Self driving labs, Electrochemistry, Reactor yield, Bayesian optimization, Automation

TL;DR: AdaCarbon, a self-driving lab for optimizing reactor yield of a reactive carbon electrolyzer.

Abstract: Reactive carbon capture combines CO2 capture and conversion in a single system. Reactive carbon electrolyzers receive a liquid eluent from a CO2 capture unit containing a sorbent that has captured CO2. This electrolyzer releases CO2 electrochemically and converts it into a value-added product like CO. The effectiveness of this system depends on high CO2 utilization and high product formation rates. We define their product as “reactor yield.” Here, we used a closed-loop, automated workflow with Bayesian optimization to maximize reactor yield in an electrolyzer operating with alkaline CO2 capture solutions. We explored a six-dimensional parameter space and found that a bicarbonate concentration of 1.5 M and carbonate concentration of 0.75 M achieved the highest reactor yield (44 mA cm⁻2). Interestingly, this optimum occurred at non-maximum values of CO partial current density (54 vs. 87 mA cm⁻2) and CO2 utilization (81% vs. 100%), highlighting the need for joint optimization of both factors.

Submission Track: Benchmarking in AI for Materials Design - Full Paper

Submission Category: All of the above

Institution Location: MIT, Cambridge, Boston, UBC, Vancouver, Canada

AI4Mat RLSF: Yes

Submission Number: 100