Answers are grounded in this paper and cite their sources.
Electrochemical flow-through reactors, including reactive electrochemical membranes, can destroy organic micropollutants in a single pass. Predicting how much they remove across heterogeneous literature data, however, remains difficult. We present a grey-box machine-learning model for parent-compound removal, built on 222 single-pass records from 40 studies and 29 pollutants. The model expresses removal through an apparent first-order rate constant and uses a CatBoost model to correct the remaining systematic prediction error. A controlled six-route ladder isolates the contribution of each component. Cross-validated R 2 rose from 0.766 for a random forest on raw features to 0.818. The gain is statistically robust (one-sided paired Wilcoxon signed-rank test, p = 3 × 10 −4 ) and came mainly from re-casting removal as a reaction rate and from the error-correction step; the added Nernst–Planck transport descriptors changed accuracy by less than the run-to-run spread of the cross-validation, which is ±0.019 in R 2 . The final model predicted removal within 10 percentage points for 76.6% of held-out records, and predicted removal increased with residence time and approached a plateau, although this shape is partly imposed by the embedded rate law. Under leave-one-study-out evaluation, 82% of the error on an unseen study was a constant study-level offset. Estimating this offset from one measurement raised R 2 from −0.104 to 0.474, whereas the same calibration provided little benefit for the black-box baseline. Within-study ranking was not recovered consistently. All fitting and calibration used this corpus alone, and on six independent studies outside it the frozen model performed less well again, so the model interpolates within the sampled operating space rather than predicting new reactor families.
No takes yet. Share an insight, caveat, or question.
He et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: