Bridging the Gap between Transformer-Based Neural Networks and Tensor Networks for Quantum Chemistry.

Area of Science:

• Quantum Chemistry
• Computational Physics
• Materials Science

Background:

• Neural network quantum state (NNQS) methods show promise for ab initio quantum chemistry.
• Calculating molecular systems with large active spaces using NNQS remains a challenge.
• Existing methods struggle with accuracy and convergence in strongly correlated regimes.

Purpose of the Study:

• To develop a novel approach enhancing NNQS accuracy and convergence for large active spaces.
• To integrate tensor network states with transformer-based NNQS (QiankunNet).
• To improve quantum state representation for complex molecular systems.

Main Methods:

• Developed QiankunNet, a transformer-based NNQS architecture.
• Transformed tensor network states into active space configuration interaction wave functions.
• Investigated sweep-based direct conversion (Conv.) and entanglement-driven genetic algorithm (EDGA) for configuration transformation.

Main Results:

• QiankunNet achieved superior accuracy compared to pretraining DMRG and coupled cluster methods.
• The Conv. method demonstrated superior efficiency in configuration transformation.
• Validated on H2O with a large active space (10e, 24o) in the cc-pVDZ basis set.

Conclusions:

• The novel QiankunNet approach offers enhanced accuracy and convergence for large active spaces.
• An efficient routine between DMRG and QiankunNet was established.
• This work presents a promising direction for advanced quantum state representation in computational chemistry.