Theory‐Guided Machine Learning to Predict the Performance of Noble Metal Catalysts in the Water‐Gas Shift Reaction

• Published in: ChemCatChem Vol. 14; no. 16
• Main Authors: Chattoraj, Joyjit, Hamadicharef, Brahim, Kong, Jian Feng, Pargi, Mohan Kashyap, Zeng, Yingzhi, Poh, Chee Kok, Chen, Luwei, Gao, Fei, Tan, Teck Leong
• Format: Journal Article
• Language: English
• Published: Weinheim Wiley Subscription Services, Inc 19-08-2022
• Subjects: Artificial neural networks; Carbon monoxide; Catalysts; Chemical reactions; Data structures; extreme gradient boosting; Machine learning; Model accuracy; Neural networks; noble metal catalysts; Noble metals; Optimization; Principles; Shift reaction; Thermodynamic equilibrium; thermodynamic loss function; water-gas shift reaction
• DOI: 10.1002/cctc.202200355
• EISSN: 1867-3899
• ISSN: 1867-3880

Machine learning (ML) has widespread applications in catalyst discovery and reaction optimization. We present a theory‐guided machine learning framework to evaluate the carbon monoxide (CO) conversion performance of noble metal catalysts in water‐gas shift (WGS) reaction. Our study is based on an open source WGS dataset, which we modify significantly to be consistent with the chemical reaction principles. We apply state‐of‐the‐art ML models including artificial neural networks, extreme gradient boosting to predict CO conversion percentage. These models show superior regression performance than the previously reported results in the literature. We further generalize the existing data structure by including physical, chemical and surface chemistry properties as fingerprint features that rationalize the importance of all the input features for CO conversion. We noticed that purely data‐driven ML models frequently violate the thermodynamic equilibrium principle and predict unphysical CO conversion percentage. We address these two problems by developing a custom loss function and an additional activation function in our neural networks architecture. Our proposed theory‐guided ML model displays high accuracy (R2 score is 0.95 and root mean square error is 6.87) and physically robust predictions. The model also opens up promising possibilities to improve CO conversion percentage, which were previously unexplored in experiments.
Theory‐guided machine learning for water‐gas shift reaction optimization. We showed that pure data‐driven models can violate the thermodynamic equilibrium principle as well as can predict non‐physical carbon monoxide conversion percentage. We resolved these two problems by developing a theory‐guided machine learning model with a unique thermodynamic loss function and one additional activation function. Our model outperforms the known models in the literature. It also shows promising reaction conditions to achieve high carbon monoxide conversion.