Graph Convolutional Neural Networks for NMR Chemical Shift Prediction

This technical report discusses the development of Graph Convolutional Neural Networks (GCNNs) for the accurate prediction of ¹⁹F-NMR chemical shifts. It outlines how molecular structures are represented as graphs, with atoms as nodes and bonds as edges, allowing the network to learn relationships between molecular structure and NMR shifts. The report details the training of a model on data from over 14,000 labeled organo-fluorine compounds, achieving a median absolute error of approximately 2.2 ppm. It also notes the challenges faced with under-represented environments and proposes future work to enhance prediction accuracy through hybrid strategies and architectural refinements. The report emphasizes the importance of extending chemical coverage and integrating the predictive tool into automated workflows for NMR-based prediction and structural verification. Overall, the current framework demonstrates high accuracy comparable to other nuclei studied, with ongoing development aimed at improving its capabilities.