Forecasting Catalytic Property-Performance Correlations for CO2 Hydrogenation to Methanol via Surrogate Machine Learning Framework

The hydrogenation of CO2 to methanol has been studied by several researchers owing to its two significant merits, namely, mitigation of CO2 emissions and production of renewable fuel. Consequently, numerous experiments have been reported for carbon-neutral methanol synthesis over copper-based catalysts over the past few years. However, it is very expensive and time-consuming to always observe reaction changes with respect to input parameters (operating conditions and catalytic properties) experimentally. Herein, this study develops an ultrafast machine learning (ML) based framework to predict CO2 conversion and methanol selectivity by comprehensive knowledge extraction (extensive database construction) from existing published literature. Among various ML algorithms, artificial neural networks (ANNs) exhibit best accuracy with R-test(2) values of 0.885 and 0.861 for CO2 conversion and methanol selectivity, respectively. Subsequently, the efficacy and fidelity of developed neural networks are depicted by satisfactory performance (R-2 >= 0.80) for majority of unseen test datasets (literature-based as well as the experimental dataset for Cu/ZnO/Al2O3 catalyst). Finally, this work concludes that this ML-based concept allows to uncover catalytic property-performance correlations hidden in large body of existing experimental research. This offers an open and generic platform for a major breakthrough of ML in heterogeneous catalysis which can help in planning future experiments.

Automated summary

This study develops an ultrafast machine learning (ML) based framework to predict CO2 conversion and methanol selectivity by extracting comprehensive knowledge from existing published literature. Among various ML algorithms, artificial neural networks (ANNs) exhibit the best accuracy. The efficacy and fidelity of the developed neural networks are depicted by satisfactory performance for majority of unseen test datasets. This work concludes that ML-based concept allows to uncover catalytic property-performance correlations hidden in existing experimental research.