A Two-Stage End-to-End Deep Learning Approach for Predicting Astrochemical Reactions

Understanding the astronomical evolution of celestial regions necessitates reconstructing evolutionary pathways within dynamic physical environments, which heavily relies on precise and comprehensive astrochemical reaction networks. Traditional methods rely on expert knowledge and incur substantial time and cost. In this study, we introduce a novel 2-stage end-to-end deep learning approach for predicting astrochemical reaction products, marking the first application of these techniques in this field. Our method comprises 2 primary phases: a generative phase leveraging a graph encoder and transformer architecture for the generation of potential reaction products, and a contrastive learning-based phase for re-ranking the potential products. We rigorously evaluated the performance of our approach using the ChemiVerse dataset. Experimental results show notable accuracy rates of 82.4% (Top-1), 91.4% (Top-3), 93.0% (Top-5), and 93.7% (Top-10). This study demonstrates the feasibility and effectiveness of using advanced deep learning techniques for end-to-end astrochemical reaction prediction.