Toward Doubly Local Double Hybrid Functionals Using Neural-Network Local Mixing Functions.

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Materials Science

Background:

• Traditional double hybrid functionals use fixed mixing parameters.
• Local mixing functions (LMFs) offer position-dependent parameterization.
• Neural networks can be used to train complex functions like LMFs.

Purpose of the Study:

• To optimize and evaluate novel "doubly local double hybrid" (DLDH) functionals.
• To investigate the performance of DLDHs with position-dependent exchange and correlation mixing.
• To assess the necessity of empirical dispersion corrections for DLDHs.

Main Methods:

• Development and training of neural network-based local mixing functions (LMFs) for exchange and correlation.
• Evaluation of DLDH functionals using W4-17 atomization energies and BH76 barrier heights.
• Comparison with constant-admixture local double hybrids and local hybrids.
• Application to the argon-benzene dissociation curve.

Main Results:

• DLDH functionals achieve high performance, matching or exceeding existing methods.
• Optimized DLDHs, particularly the DL2DH variant, render empirical dispersion corrections unnecessary.
• Large correlation LMF values in noncovalent interaction regions explain the improved performance.
• Position-dependent mixing offers potential for improved treatment of strong correlation effects.

Conclusions:

• DLDHs represent a promising advancement in density functional theory.
• Position-dependent mixing in DLDHs significantly enhances accuracy and can eliminate the need for dispersion corrections.
• Further development of DLDHs could lead to more accurate and efficient quantum chemical calculations.