Improving the Performance of Long-Range-Corrected Exchange-Correlation Functional with an Embedded Neural Network

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Density Functional Theory (DFT) relies on accurate exchange-correlation functionals.
• Existing functionals like LC-BLYP have limitations with fixed parameters.
• Machine learning offers a novel approach to improve DFT accuracy.

Purpose of the Study:

• To develop a general-purpose machine-learning exchange-correlation functional for DFT.
• To enhance the accuracy of thermochemical and kinetic energy calculations.
• To adapt the range-separation parameter dynamically for diverse systems.

Main Methods:

• Developed a machine-learning functional (LC-BLYP-NN) based on LC-BLYP.
• Embedded a neural network to determine the system-specific range-separation parameter (μ).
• Optimized the neural network using a dataset of 368 accurate energetic properties.

Main Results:

• LC-BLYP-NN demonstrated balanced performance across various energetic properties.
• Significantly improved accuracy for atomization energies and heats of formation compared to fixed-μ LC-BLYP.
• Maintained similar or slightly reduced accuracy for ionization potentials, electron affinities, and reaction barriers.

Conclusions:

• Machine learning techniques show significant potential for improving DFT calculations.
• The LC-BLYP-NN functional offers enhanced accuracy for specific energetic properties.
• Dynamic optimization of parameters via ML is a promising avenue for computational chemistry.