Uncertainty Qualification\nfor Deep Learning-Based\nElementary Reaction Property Prediction

The prediction of the thermodynamic\nand kinetic properties\nof elementary\nreactions has shown rapid improvement due to the implementation of\ndeep learning (DL) methods. While various studies have reported the\nsuccess in predicting reaction properties, the quantification of prediction\nuncertainty has seldom been investigated, thus compromising the confidence\nin using these predicted properties in practical applications. Here,\nwe integrated graph convolutional neural networks (GCNN) with three\nuncertainty prediction techniques, including deep ensemble, Monte\nCarlo (MC)-dropout, and evidential learning, to provide insights into\nthe uncertainty quantification and utility. The deep ensemble model\noutperforms others in accuracy and shows the highest reliability in\nestimating prediction uncertainty across all elementary reaction property\ndata sets. We also verified that the deep ensemble model showed a\nsatisfactory capability in recognizing epistemic and aleatoric uncertainties.\nAdditionally, we adopted a Monte Carlo Tree Search method for extracting\nthe explainable reaction substructures, providing a chemical explanation\nfor DL predicted properties and corresponding uncertainties. Finally,\nto demonstrate the utility of uncertainty qualification in practical\napplications, we performed an uncertainty-guided calibration of the\nDL-constructed kinetic model, which achieved a 25% higher hit ratio\nin identifying dominant reaction pathways compared to that of the\ncalibration without uncertainty guidance.