Combined scaling relations and machine learning for predicting C-F bond activation energies in tetrafluoromethane

Given the potential environmental and health risks posed by per- and polyfluoroalkyl substances, we use CF 4 as a model compound to investigate the trends in C–F bond activation. Using density functional theory calculations and machine learning, we examined three distinct catalyst systems: transition metal (TM) surfaces, single-atom alloys (SAAs), and SiC-supported single-atom catalysts (SACs). On TM surfaces, both activation energy (Ea) versus the reaction energy (ΔE) and Ea versus the fluorine binding energy (BE(F)) exhibit well-defined linear relationships, which are strongly governed by the coordination environment. On SAAs, where a single TM atom is embedded in a host lattice with considerable site heterogeneity, the Ea~ΔE correlation follows a volcano-shaped trend. In contrast, on SACs, which feature lower coordination numbers, both the Ea~ΔE and Ea ~ BE(F) correlations adopt volcano-type curves. For SAAs, a convolutional neural network (CNN) model was developed to predict BE(F) directly from the density of states. Combining the CNN-predicted BE(F) with the SAA-specific Ea ~ BE(F) scaling relation enables efficient estimation of C-F activation energies within this catalyst class. This descriptor-based machine-learning strategy provides a transferable workflow that can be extended to other catalyst classes using system-specific training data and scaling relations. Overall, this work provides a comprehensive framework for understanding and predicting C–F bond activation in CF 4 over different catalysts, paving the way for the rational design of efficient catalysts for perfluorocarbon decomposition.