Multi-output machine learning for photocatalytic CO2 reduction and inverse design of catalysis
AIP Advances · 01.10.2026 · 16 (10)
DOI: 10.1063/5.0353715 ↗Kurz gefasst
Fachartikel in der Zeitschrift AIP Advances, online erschienen am 29.10.2026. Verfasst von Edward Crosby und 6 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Indian Institute of Technology Madras, Durham University.
Abstract (Originalsprache)
Photocatalytic carbon dioxide reduction offers a promising solar-driven route to carbon-neutral fuel production. However, progress remains largely empirical, with catalyst development often guided by trial and error. The vast combinatorial space of catalyst compositions, morphologies, and reaction conditions makes exhaustive exploration impractical, motivating data-driven approaches. To address fragmented datasets and single-outcome prediction, we present a unified machine learning framework comprising three concepts. First, multi-output models simultaneously predict reaction products and yield rates, capturing the coupled nature of photocatalytic performance. Second, we curate and standardize a dataset of 461 entries from 63 studies, linking three-component catalyst descriptors and experimental conditions to measured products and yields. Third, an inverse design framework integrates gradient-based and evolutionary optimization to generate catalyst compositions and reaction conditions tailored to user-defined targets. On the curated PRIDMO dataset, study-grouped cross-validation, evaluating zero-shot transfer to entirely unseen experiments, yields negative (R2) values [−0.034 random forest (RF); −0.314 neural network (NN)] and reduced F1 scores (0.418 RF; 0.425 NN), indicating limited generalization without experiment-specific data. Standard ungrouped fivefold cross-validation approximates a setting where domain-specific data are available, achieving (R2) values of 0.704 (RF) and 0.679 (NN) for yield prediction, with weighted F1 scores of 0.822 and 0.867 for product classification. Few-shot experiments support this interpretation, showing substantial performance recovery with a small number of domain-specific examples, although this remains preliminary. The inverse design framework generated physically plausible candidates with predicted yields differing from targets by ∼4%–24%, providing a hypothesis-generation and candidate-screening tool requiring experimental validation.
Förderung
- Indian Institute of Technology Madras
- Durham University
Zitieren
APA
Crosby, E., Gilleece, G., Hedayati, M. K., Verumandi, R., Kunnan, A. P., Roy, S. C., & Zeze, D. A. (2026). Multi-output machine learning for photocatalytic CO2 reduction and inverse design of catalysis. AIP Advances, 16(10). https://doi.org/10.1063/5.0353715
Vancouver
Crosby E, Gilleece G, Hedayati MK, Verumandi R, Kunnan AP, Roy SC, et al. Multi-output machine learning for photocatalytic CO2 reduction and inverse design of catalysis. AIP Advances. 2026;16(10). doi:10.1063/5.0353715
BibTeX
@article{Crosby2026,
author = {Crosby, Edward and Gilleece, Grainne and Hedayati, Mehdi Keshavarz and Verumandi, Ragulkrishnan and Kunnan, Aswathi Pullan and Roy, Somnath C. and Zeze, Dagou A.},
title = {Multi-output machine learning for photocatalytic CO2 reduction and inverse design of catalysis},
journal = {AIP Advances},
year = {2026},
volume = {16},
number = {10},
doi = {10.1063/5.0353715}
}
EUVIMED