Fig. 2. Pre-trained and fine-tuned NewtonNet performance on Hessian prediction of the test set.

a The pre-trained model accurately predicts Hessians at R and P minima geometries but fails dramatically at TSs. b The fine-tuned model using the T1x data significantly improves the accuracy at TSs but with notable underestimation of Hessian eigenvalues. c Augmenting the T1x dataset with compressed bond configurations creates more balanced training data and improves the overall performance. More comprehensive comparisons of the pre-trained and fine-tuned ML prediction accuracy for Hessians is provided in Supplementary Figs. S4–S8. R: reactant; TS: transition state; P: product; RMSE: root mean squared error; DFT: density functional theory; ML: machine learning. Source data for this figure are provided with this paper.