Fig. 2: Development of a fast surrogate objective function.
a, Prediction accuracy as a function of density functional, basis set and solvation model on an experimental database of 174 redox potentials46. b, GNN topology for predicting stability and redox potential. Two separate models are trained. The first predicts spin density and buried volume for an input molecule at each atom. The second predicts the OP and RP for an entire molecule. Input dimensions for node and edge features are n, number of atoms and r, 2 × number of bonds, respectively. c, Learning curve for redox potentials showing prediction accuracy versus number of training molecules. d, Distribution of redox prediction errors for the final trained model. e, Learning curve of the prediction of parameters governing radical stability. Error bars in c and e extend to ±1 s.d. across three replicates.
