AI Predicts Chemical Reaction Yields Using Vision and Cross-Attention

A new AI architecture, called Vision Cross-Attention, has been proposed to enhance the prediction of chemical reaction yields. This model integrates two distinct data modalities: traditional tabular physical-organic descriptors and 2D visual representations of molecular structures. By fusing these inputs, the system aims to overcome the limitations of 1D quantum descriptors that lack explicit spatial information. The study demonstrates that a standard computer vision backbone, when applied to simple 2D skeletal structures, can independently surpass purely quantum-based baseline models in predictive accuracy. When both modalities are synergized through an optimal cross-attention framework, the model achieves superior accuracy, with a test RMSE of 5.27%. Mechanistic analysis revealed that the model actively queries spatial information, effectively offloading macroscopic steric identification to the visual pathway and prioritizing critical steric bottlenecks like aryl halides. The architecture also employs residual skip connections to protect non-spatial electronic parameters during data fusion, providing a scalable and interpretable blueprint for integrating deep visual learning into physical chemistry.