Uncertainty-calibrated deep learning for rapid identification of reaction mechanisms

Analyzing the reaction mechanism is the cornerstone across all chemistry disciplines. However, the existing reaction mechanism data only have limited coverage of the enormous chemical space and are usually collected from various sources, thus, limiting advanced machine learning tools to facilitate classical computational chemistry approaches for obtaining reaction mechanisms. Herein we present an uncertainty-calibrated learning-to-ranking (UC-LTR) model for constructing mechanistic reaction networks. UC-LTR model can make efficient and accurate prediction with quantitative uncertainty learned from inconsistent and sparse data sources. This allows a coherent utilization of existing data, improving the confidence of using predicted results for the following experimental or quantum chemistry refinement. We demonstrated this approach can provide efficient and reliable identification of reaction networks for high-temperature gas-phase reactions. This work showcase a practical deep learning strategy in fields of chemistry where the availability of high-quality big data is limited