Fig. 4: The quality of optimized TSs using NewtonNet.
From: Analytical ab initio hessian from a deep learning potential for transition state optimization
a NewtonNet predicts reactions that match the intended reactions with higher success rates, both with and without full Hessians, compared to DFT. The full Hessian also finds ~10% more TSs which involve chemical reactions as opposed to conformational changes. b The value of the full Hessians over the approximate QN approach is apparent when the quality of the initial TS guesses deteriorates. The QN convergence decays with additional noise to the guess structures, while the full Hessian convergence is more robust to perturbations. c Comparing whether the left most frequency found on the ML PES is also a negative frequency on the DFT PES using the ML geometry. d Reoptimizing the ML transition state structure on the DFT surface demonstrates superior performance for 2-end matches and identifying chemical reactions compared with starting from the original KinBot initial guess. TS: transition state; QN: quasi-Newton; DFT: density functional theory; ML: machine learning; PES: potential energy surface. Source data for this figure are provided with this paper.
