A Density Functional Tight Binding Layer for Deep Learning of Chemical Hamiltonians

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Mechanics

Background:

• Current neural networks rely on quantum chemistry solely for training data.
• Integrating quantum chemistry directly into models can improve prediction accuracy.

Purpose of the Study:

• To implement Density-Functional-Tight-Binding (DFTB) theory as a layer within deep learning models.
• To explore the use of splines and feed-forward neural networks as inputs to the DFTB layer.
• To reduce overfitting and improve model performance on larger molecules through regularization.

Main Methods:

• Developed a deep learning framework incorporating a self-consistent-charge DFTB layer.
• Utilized backpropagation for efficient model training.
• Applied regularization techniques to penalize nonmonotonic behavior and Hamiltonian deviations.
• Evaluated models on 15,700 hydrocarbons, comparing root-mean-square errors to the initial DFTB model.

Main Results:

• The spline model reduced energy and dipole moment errors by 60% and 42%, respectively, when trained on molecules up to seven heavy atoms.
• The neural network model achieved greater reductions, with 67% for energy and 59% for dipole moments.
• Training on smaller molecules (up to four heavy atoms) showed reduced, but still significant, error reductions of approximately 53% for energy and 25% for dipole moments.

Conclusions:

• Integrating DFTB theory as a layer in deep learning models significantly enhances molecular property prediction accuracy.
• The neural network input approach demonstrated superior performance compared to the spline input.
• The developed method offers a promising direction for more accurate and efficient computational chemistry predictions.