Towards a transferable fermionic neural wavefunction for molecules.

Area of Science:

• Quantum Chemistry
• Computational Physics
• Machine Learning

Background:

• Deep neural networks combined with variational Monte Carlo methods offer accurate solutions to the electronic Schrödinger equation.
• Current methods require computationally expensive, system-specific wavefunction optimization from scratch, hindering widespread adoption.

Purpose of the Study:

• To develop a novel neural network ansatz for efficient and transferable wavefunction modeling.
• To reduce the computational cost associated with solving the electronic Schrödinger equation.

Main Methods:

• Proposed a neural network ansatz that maps uncorrelated Hartree-Fock orbitals to correlated neural network orbitals.
• Demonstrated transferability by pre-training a wavefunction model on smaller molecular fragments and applying it to larger compounds.
• Validated the approach with experimental evidence.

Main Results:

• The proposed ansatz effectively learns a single wavefunction across diverse compounds and geometries.
• Successful transfer of a pre-trained model from smaller fragments to larger molecules was achieved.
• Experimental evidence supports the potential for a generalized wavefunction model.

Conclusions:

• The developed neural network ansatz significantly reduces computational expense for high-accuracy ab-initio energy calculations.
• Pre-training across various compounds and geometries can lead to a foundational model for efficient electronic structure studies.
• This approach paves the way for more accessible and rapid computational chemistry simulations.