Predicting electronic screening for fast Koopmans spectral functional calculations.

Area of Science:

• Computational chemistry
• Quantum mechanics
• Materials science

Background:

• Koopmans spectral functionals offer high accuracy for predicting electronic spectral properties.
• These functionals require computationally expensive screening parameters, limiting their application.
• Current methods necessitate calculating these parameters for each new system.

Purpose of the Study:

• To develop a computationally efficient machine-learning model for predicting Koopmans screening parameters.
• To enable wider application of accurate Koopmans spectral functionals.
• To reduce the computational cost associated with predicting spectral properties.

Main Methods:

• Developed a machine-learning model to predict orbital-dependent screening parameters.
• Input for the model: orbital densities from standard density-functional theory (DFT) calculations.
• Validated the model using two prototypical systems.

Main Results:

• The machine-learning model accurately predicts screening parameters with minimal training.
• Orbital energies obtained using predicted parameters deviate by less than 20 meV on average compared to linear-response calculations.
• Significant reduction in computational run times achieved.

Conclusions:

• The developed machine-learning approach drastically cuts down computation time for Koopmans spectral functionals with minimal accuracy loss.
• This method opens doors for applying these functionals to complex problems, including temperature-dependent spectral properties.
• Combines frozen-orbital approximations and machine learning to efficiently measure energy-occupancy curvature.