Recent developments in Machine Learning (ML) have demonstrated the power of deep neural networks when applied to a variety of subjects, including chemistry. It is now standard practice to train ML potential energy surfaces to “chemical accuracy” (0.5 kcal/mol) of target quantum mechanical methods. Additionally, ML techniques have proven effective at learning a variety of other quantum properties including charge distributions, non-adiabatic couplings, and band gaps.<br/><br/>Giving physical meaning to predicted values presents a much more difficult task for ML approaches. We seek to overcome this challenge by training a deep neural network to predict parameters for effective Hamiltonian models. Effective Hamiltonians and tight binding models have a long history of success in chemistry and include widely used methods such as Hückel theory, AM1, DFTB and many others. By combining these two methodologies, we retain the speed and physical intuition associated with simple physical models while gaining the predictive accuracy of modern ML approaches. These dynamic effective Hamiltonians will be tested against various phenomena and observable properties to illustrate accuracy and generality.

Machines learning physics: Deep tensor neural networks for dynamically optimized effective Hamiltonians

256th National Meeting (2018)

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