A central goal of computational chemistry is to accurately and efficiently predict molecular properties. Quantum mechanical (QM) methods offer high levels of accuracy but are limited in their applicability due to high computational costs. Classical force fields, on the other hand, have much lower computational costs but at the expense of transferability and accuracy. Machine learning algorithms have been developed to bridge this gap. The goal of the ANAKIN-ME (ANI) method is to develop a universal neural-network model to predict the energies of molecules with the accuracy of QM and the time scale of classical methods. One model developed with this method, ANI-1ccx, implements transfer learning to increase its accuracy, approaching CCSD(T)/CBS accuracy. This model has been shown to closely match coupled cluster results when predicting molecular energies. A previous ANI model, ANI-1x, has been shown to predict vibrational frequencies closely matching DFT with the functional and basis set it has been trained to, an important property for thermodynamic calculations. To test the applicability of ANI-1ccx, we used numerous sets of chemical reactions to assess how well this model predicts reaction enthalpies. ANI-1ccx outperforms DFT and shows similar accuracy to MP2/cc-pVTZ compared to experimental results and is ~106 times faster.

Thermochemical studies with machine learning

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.