Machine Learning Enhanced DFTB Method for Periodic Systems: Learning from Electronic Density of States.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning Applications

Background:

• Density Functional Tight Binding (DFTB) offers low computational cost for quantum chemical simulations.
• Increasing DFTB accuracy, especially for complex systems like solids with defects, remains a challenge.
• Direct machine learning predictions can be difficult for systems with long-range effects.

Purpose of the Study:

• To enhance the accuracy of the DFTB method using machine learning.
• To optimize DFTB parameters for defective periodic silicon and silicon carbide systems.
• To validate the machine learning approach for predicting material properties, specifically Density of States (DOS).

Main Methods:

• Implemented a machine learning algorithm, specifically backpropagation, to optimize DFTB parameters.
• Trained the model using Density Functional Theory (DFT) results as a reference.
• Tested the generalization capability of the trained model on unseen geometries.

Main Results:

• The machine learning-optimized DFTB model significantly reduced discrepancies between DFTB and DFT calculations for DOS.
• Derived properties such as Mulliken population distribution and projected DOS remained physically sound after optimization.
• The trained model demonstrated good transferability to different system configurations.

Conclusions:

• The developed method provides a computationally efficient approach for accurate simulations of defective materials.
• This machine learning-enhanced DFTB method offers a viable compromise for simulating systems where direct ML predictions are challenging.
• The approach requires moderate training efforts for substantial improvements in simulation accuracy.