2021 Annual Meeting

(342ak) Predicting Uncertainty in Supervised Machine Learning Predictions of Chemical Kinetics

Authors

Komp, E. - Presenter, University of Washington
Computing reaction rate constants atomistically becomes prohibitively expensive as the size of the system grows.1,2 For this reason, some recent work has focused on training supervised machine learning (ML) algorithms to predict rate constants and activation energies without requiring large ab initio calculations.3–6 These approaches have shown to be successful in predicting target kinetics properties, however they do not provide a measure of error with each new prediction; only training and validation losses are available. In our previous work5, a DNN was trained to predict the product of reaction rate constants and partition functions with a test mean absolute error of 1.1%. While this was encouraging, individual predictions had errors as high as 33.5%. It was thus clear that anticipating individual prediction errors was necessary to inform design choices. Here we present our recent work which employs modified generative adversarial networks (GANs)7 to estimate the uncertainty of each predicted reaction rate constant product. The model was trained on the database created for the original work.5 We show that the consideration of individual prediction error can drastically impact rate constant predictions on a target system.

1. Ju, L. P., Han, K. L. & Zhang, J. Z. H. Global dynamics and transition state theories: Comparative study of reaction rate constants for gas-phase chemical reactions. Journal of Computational Chemistry vol. 30 305–316 (2009).

2. Sumathi, R. & Green, W. H. A priori rate constants for kinetic modeling. Theoretical Chemistry Accounts vol. 108 187–213 (2002).

3. Grambow, C. A., Pattanaik, L. & Green, W. H. Deep Learning of Activation Energies. J. Phys. Chem. Lett. 11, 2992–2997 (2020).

4. Nandi, A., M. Bowman, J. & Houston, P. A Machine Learning Approach for Rate Constants. II. Clustering, Training, and Predictions for the O(3P) + HCl → OH + Cl Reaction. J. Phys. Chem. A 124, 5746–5755 (2020).

5. Komp, E. & Valleau, S. Machine Learning Quantum Reaction Rate Constants. J. Phys. Chem. A 124, 8607–8613 (2020).

6. Ulissi, Z. W., Medford, A. J., Bligaard, T. & Nørskov, J. K. To address surface reaction network complexity using scaling relations machine learning and DFT calculations. Nat. Commun. 8, 14621 (2017).

7. Lee, M. & Seok, J. Estimation with Uncertainty via Conditional Generative Adversarial Networks. arXiv:2007.00334v1 (2020).