Work

Physicochemical Machine Learning for Predicting Faradaic Efficiency in Electrochemical Nitrogen Reduction: Critical Analysis and Optimization

Year: 2026
Type: dataset
Abstract: The electrochemical nitrogen reduction reaction (NRR) offers a potentially sustainable route to ammonia synthesis under ambient conditions, but its practical implementation remains severely constrained by the competing hydrogen evolution reaction (HER) and the resulting low Faradaic efficiency (FE). Machine learning (ML) has emerged as a promising strategy for accelerating NRR catalyst discovery; however, the reliability and generalizability of models trained on heterogeneous experimental literature remain poorly established. Here, we critically reassess the predictive performance of literature-derived ML models for NRR and introduce a physicochemically informed framework based on continuous descriptors designed to improve transferability across catalyst chemistries. We initially extracted 692 experimentally reported NRR measurements from 61 publications and, following systematic quality control and manual verification, curated a final dataset of 243 measurements for model development. Through a systematic replication of a previously reported ML workflow, we show that the reported predictive performance progressively deteriorates from R2 = 0.91 under the original validation protocol to R2 = 0.66 under conventional cross-validation and to R2 = −0.23 when evaluated on independently extracted external data, revealing substantial information leakage and severe overestimation of predictive performance in conventional literature-based workflows. We subsequently develop an ML framework using 14 continuous physicochemical descriptors, including electronegativity (χ), cationic radius (Rcat, pm), electrochemical potential, pH, electrolyte concentration, and derived interaction terms. Using three-way validation across 20 independent seeds, the optimized ExtraTrees model achieves a mean test R2 = 0.651 ± 0.055 and MAE of 13.29 ± 0.88%. Unlike categorical representations, the continuous descriptor space encodes transferable physicochemical relationships and reduces dependence on catalyst identity, enabling more reliable extrapolation beyond the chemical space represented in the training data. These results demonstrate that rigorous validation and chemically meaningful representations are essential for obtaining reliable performance estimates in ML-driven electrocatalyst discovery and establish a robust framework for data-driven NRR catalyst screening and prospective experimental validation.
Language: English
Subfield: Catalysis
Open Access status: green
OpenAlex ID: W7204171302
Cites:
Cited by:
Related to: 10