Accurate Ring Strain Energy Predictions with Machine Learning and Application in Strain-Promoted Reactions

Ring strain energy (RSE) is crucial for understanding molecular reactivity. However, quantitatively determining RSE through experiments or quantum mechanics is resource-intensive, limiting its application on a large scale. We developed a physics-based workflow and a data-driven graph neural network (GNN) capable of reliably predicting RSE in minutes or milliseconds, respectively. For each molecule, the workflow first identifies low-energy conformers, then computes the RSE using the AIMNet2 machine learning interatomic potentials. We validated the approach both computationally and experimentally. Compared to the ωB97M-D4/Def2-TZVPP method, the workflow achieved an R² value of 0.997 and a mean absolute error (MAE) of 0.896 kcal/mol. Using this workflow, we distinguished reactive from non-reactive molecules in copper-free click chemistry and ring-opening metathesis polymerization, demonstrating the workflow's generalizability to diverse molecules. Furthermore, we compiled "RSE Atlas," a computational database of 16,905 single-ring molecules, providing a rich resource for examining factors influencing RSE. Employing this dataset, we trained a GNN that predicts RSE in milliseconds using only 2D molecular information. Our methods render RSE a readily computable property for on-the-fly applications in experimental and computational work.