Uncertainty Qualification for Deep Learning-Based Elementary Reaction Property Prediction.

Area of Science:

• Computational Chemistry
• Chemical Kinetics
• Machine Learning

Background:

• Deep learning (DL) significantly advances the prediction of thermodynamic and kinetic properties for elementary reactions.
• However, the quantification of prediction uncertainty in these DL models remains underexplored, limiting confidence in their practical application.

Purpose of the Study:

• To integrate graph convolutional neural networks (GCNN) with uncertainty quantification techniques.
• To evaluate the performance of different uncertainty prediction methods (deep ensemble, Monte Carlo dropout, evidential learning) for chemical reaction properties.
• To demonstrate the utility of uncertainty quantification in refining kinetic models.

Main Methods:

• Implemented GCNN combined with deep ensemble, Monte Carlo (MC)-dropout, and evidential learning for uncertainty prediction.
• Utilized Monte Carlo Tree Search (MCTS) for extracting explainable reaction substructures.
• Performed uncertainty-guided calibration of a DL-constructed kinetic model.

Main Results:

• The deep ensemble model demonstrated superior accuracy and reliable uncertainty estimation across various datasets.
• The deep ensemble model effectively distinguished between epistemic and aleatoric uncertainties.
• Uncertainty-guided calibration of the kinetic model improved pathway identification by 25% compared to standard calibration.

Conclusions:

• Deep ensemble methods offer a robust approach for uncertainty quantification in DL-based prediction of chemical reaction properties.
• Explainable AI techniques, like MCTS, can provide chemical insights into DL predictions and their uncertainties.
• Uncertainty quantification is crucial for enhancing the reliability and practical utility of DL-generated kinetic models.