Machine-Learned Energy Functionals for Multiconfigurational Wave Functions.

Area of Science:

• Computational chemistry
• Quantum mechanics
• Machine learning in science

Background:

• Accurate calculation of molecular energies is crucial for understanding chemical reactions.
• Multiconfigurational wave functions are necessary for describing systems with strong electron correlation.
• Existing methods often struggle with balancing accuracy and computational cost.

Purpose of the Study:

• To introduce and validate multiconfiguration data-driven functional methods (MC-DDFMs).
• To assess the ability of machine-learned functionals to correct wave function energies.
• To evaluate the transferability and active-space independence of these new methods.

Main Methods:

• Development of MC-DDFMs using machine-learned functionals.
• Featurization of wave functions using density and on-top density.
• Training and testing on carbene singlet-triplet energy splittings.

Main Results:

• MC-DDFMs achieve near-benchmark performance on unseen systems.
• The methods demonstrate robustness and active-space independence.
• Corrections to the classical energy show better transferability than total energy corrections.

Conclusions:

• Density and on-top density contain essential information for correcting multiconfigurational wave function energies.
• MC-DDFMs show significant promise for developing functionals in multiconfiguration paired-differential function theory (MC-PDFT).
• This approach offers a path towards more accurate and transferable energy calculations in quantum chemistry.