Activity Descriptors of Mo2C-based Catalysts for C-OH Bond Activation

Authors:

Abstract
Deriving structure-activity relationships is crucial for designing efficient catalysts. To aid in this quest, data-driven methods such as machine learning (ML) are emerging to unravel the complicated multi-dimensional structure–activity relationships in heterogeneous catalysis. In this work, we incorporate ML tools with accurate density functional theory (DFT) energetics, electronic and geometric features to establish the structure-activity relationships in the context of designing efficient orthorhombic molybdenum carbide (Mo2C)-based catalysts for hydrodeoxygenation (HDO) reaction. Our previous studies show that C-OH activation is the rate-determining step in the HDO reaction over Mo2C (101) surface. In this work, DFT is used to obtain accurate barriers (Ea), reaction energy (∆E) for the C-OH activation in HDO over the most stable and meta-stable facets of Mo2C and transition metal doped Mo2C. Then we constructed the feature database of all structures of Mo2C based catalysts and employed scikit-learn’s ML models to obtain the best primary features correlating with the activity (Ea). The results show that Kernal Ridge Regression (KRR) gives the best ML model for predicting Ea of C-OH bond breaking (on unseen data) with a mean test RMSE of 0.25 eV. SHapley Additive exPlanations (SHAP) analyses for all the ML models was combined in a systematic manner to reveal that ∆E, d-band center, and d MC are the most important features governing the activity of Mo2C-based catalysts for C-OH bond activation. Finally, SISSO was used to validate the SHAP results, and to obtain a 2-dimensional generic descriptor which is also physically interpretable to some extent. This SISSO model yielded test RMSE score of 0.14 eV, which is well below the DFT level error margin (±0.20 eV).