(51a) Deep Learning Quantum Reaction Rate Constants

In this talk we will discuss our recent effort [1] to overcome the cost of ab initio reaction rate constant calculations by using supervised machine learning. A deep neural network, DNN, was trained on our in-house generated ~1.5 million example dataset of quantum reaction rate constants. Rate constants multiplied by reactant partition functions were computed for single, double, symmetric and asymmetric one-dimensional potentials over a broad range of reactant masses and temperatures. The network was trained for hundreds of epochs and able to predict the logarithm of the rate product with a relative mean absolute error of 1.1%. Systems beyond the test set were also studied, these included the diffusion of hydrogen on Ni(100), the Menshutkin reaction of pyridine with CH₃Br in the gas phase, the reaction of formalcyanohydrin with HS^− in water and the F + HCl reaction. The DNN predictions were in good agreement with the exact rates.

While this was encouraging, when exploring the test set predictions, some had errors as high as 33.5%. It was thus clear that anticipating individual prediction errors was necessary to inform design choices. We will discuss our recent effort to estimate that error using modified generative adversarial networks (GANs)[2-3].

[1] E. Komp and S. Valleau, “Machine Learning Quantum Reaction Rate Constants,” J. Phys. Chem. A, 124:8607-8613, 2020.

[2] I. Goodfellow et al., “Generative adversarial networks,” arXiv:1406.2661, 2014.

[3] M. Lee and J. Seok, “Estimation with Uncertainty via Conditional Generative Adversarial Networks,” arXiv:2007.00334 2020.