Can machine learning help identify which interfacial descriptors control CO2-reduction microenvironments?

The electrochemical problem

CO2 electroreduction on copper-based catalysts is controlled not only by the catalyst surface but also by the local electric-double-layer environment. Ionic liquids and imidazolium salts can change hydrogen bonding, cation organization, and interfacial fields, all of which can alter reaction energetics.

The challenge is that catalyst composition, tandem active sites, ionic-liquid identity, and interfacial structure are coupled. The paper explores whether multiscale modeling plus machine learning can isolate the descriptors that matter most for pathways toward ethylene and acetamide.

What the researchers did

The study considers CuAg tandem motifs in which Ag assists CO formation and Cu participates in downstream C–C chemistry, while C–N coupling can produce acetamide. Ab initio molecular dynamics is used to examine the interfacial environment created by imidazolium salts.

Regression machine learning with extreme gradient boosting and SISSO is then used to relate candidate descriptors to a limiting-potential target, Umax(η), for the reaction pathways.

What they found

The analysis identifies coupling energy and the average deviation in ground-state band gaps of the constituent elements as important features for both ethylene and acetamide synthesis. Two imidazolium salts highlighted in the study, B2195 and B2320, reduce the reported maximum limiting potential to −0.84 and −1.00 V, respectively.

The paper also maps favorable C–C coupling routes and uses molecular simulation to connect ionic-liquid structure with the electrochemical double-layer environment.

Why this matters

Many electrochemical ML studies treat composition as the main descriptor. This work is interesting because it points toward a more interfacial view: the target is influenced by coupling energetics and by the local environment in which those elementary steps occur.

For CO2 reduction, that matters because selectivity emerges from a network of competing proton, carbon, and nitrogen coupling pathways whose relative energetics are sensitive to the interface.

One important caution

The ML target here is a modeled limiting-potential quantity embedded in a multiscale computational workflow. Descriptor importance within that model should not be confused with a direct experimental proof that those features alone control product selectivity in all CuAg systems.

The strength of the paper is the integration of physical modeling and data-driven descriptor analysis. Generalization to other catalysts and electrolytes will still require targeted experimental tests.