Exchange functionals trained on exact exchange for molecular and solid-state systems

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Self-interaction error (SIE) in density functional theory (DFT) causes significant inaccuracies in calculated chemical reaction barrier heights and solid-state band gaps.
• Accurate prediction of these properties is crucial for understanding chemical reactions and designing new materials.

Purpose of the Study:

• To develop a novel neural network-based exchange functional designed to mitigate SIE-related inaccuracies.
• To enhance the predictive accuracy of DFT calculations for chemical and material properties.

Main Methods:

• Developed a neural network-based exchange functional trained on exact exchange data.
• Evaluated the functional's performance against established functionals like PBE, SCAN, M06L, and revM06L.
• Assessed accuracy for barrier heights (BH), band gaps, atomization energies (AE), ionization potentials (IP), and vibrational frequencies (VF).

Main Results:

• The developed functional demonstrates improved accuracy in predicting barrier heights and band gaps compared to PBE, SCAN, M06L, and revM06L.
• The functional also shows reliable performance for other key properties, including atomization energies, ionization potentials, and vibrational frequencies.
• The neural network approach effectively reduces SIE-related errors.

Conclusions:

• The new neural network-based exchange functional offers a promising approach to reduce self-interaction error in DFT.
• This advancement leads to more accurate predictions of critical chemical and material properties.
• The functional provides a reliable tool for computational chemistry and materials science research.