In brief

QuestionCan ML accelerate catalyst screening while still revealing a chemically interpretable reason for ligand effects?
ResultAn FNN trained on 81 DFT structures reached R2 = 0.94 and RMSE = 0.14 V on the independent test set; DFT rechecks of the top 17 candidates averaged 0.05 V prediction error.
Why it mattersThe workflow uses explainable ML to move from fast screening back to orbital-level interpretation, showing that the optimal axial ligand is metal-dependent rather than universal.
Machine learning-accelerated screening of axial functional group-decorated single-atom catalysts for efficient photocatalytic and electrocatalytic CO2 reduction to HCOOH
Liangshi Wang et al. · Journal of Energy Chemistry · Published 23 September 2026
DOI: 10.1016/j.jechem.2026.09.028

The electrochemical problem

Axial ligands can strongly modify the electronic structure of a single-atom catalyst, but combining many metal centers with many possible ligands quickly creates a large design space.

For CO2 reduction to formic acid, the challenge is not only to find a low limiting potential but also to understand why one ligand helps one metal and fails for another.

What the researchers did

The authors generated a DFT dataset of 81 axial-ligand-decorated TM-X@C3N3 structures and trained four machine-learning models to predict the limiting potential for HCOOH formation.

The best feedforward neural network was used to screen a much larger candidate set. The authors then re-evaluated the top 17 candidates with DFT, performed constant-potential implicit-solvent calculations, and used SHAP plus electronic-structure analysis to identify influential descriptors and orbital interactions.

What they found

The FNN achieved R2 = 0.94 and RMSE = 0.14 V on the independent test set. For the 17 top candidates selected by ML, subsequent DFT validation gave an average prediction error of about 0.05 V.

Constant-potential calculations highlighted Nb-F@C3N3, Mo-O@C3N3, and Mo-S@C3N3 as promising electrocatalysts. SHAP and orbital analysis indicated that axial ligands redistribute metal d-orbital interactions with the *HCOO intermediate, and the favorable ligand sequence changes with the metal center.

Why this matters

The paper is most useful when it connects three levels of reasoning: DFT supplies chemically grounded labels, ML expands the search, and explainability sends the result back to electronic structure.

That loop avoids a common weakness of catalyst ML studies in which a model produces a ranking but little guidance for how to redesign the catalyst family.

One important caution

The independent test statistics are strong for this dataset, but the training set contains only 81 DFT structures. The selected catalysts are validated computationally rather than experimentally.

The reported candidates should therefore be read as experimentally testable hypotheses within the defined TM-X@C3N3 family, not as established high-performance CO2RR catalysts.

Primary source

Liangshi Wang et al. Machine learning-accelerated screening of axial functional group-decorated single-atom catalysts for efficient photocatalytic and electrocatalytic CO2 reduction to HCOOH. Journal of Energy Chemistry (2026). DOI: 10.1016/j.jechem.2026.09.028.

Research-note facts and quantitative values were checked against the publisher record linked above. The “Why this matters” and “One important caution” sections are EchemAI editorial interpretation.