Hybrid Tensor Network and Neural Network Quantum States for Quantum Chemistry

Area of Science:

• Quantum many-body physics
• Computational chemistry
• Machine learning in quantum mechanics

Background:

• Neural network quantum states (NQS) are powerful for quantum many-body problems but face challenges in molecular systems.
• Existing NQS methods require improvements for accuracy and efficiency in electronic structure calculations.

Purpose of the Study:

• To introduce innovations overcoming limitations of NQS for molecular systems.
• To enhance the applicability and accuracy of NQS in quantum chemistry.

Main Methods:

• Developed a bounded-degree graph recurrent neural network (BDG-RNN) ansatz, hybridizing tensor and neural network states for molecular suitability.
• Introduced neural network correlators (NNCs), including cos-RBM and Ising-RBM, to improve wave function expressivity and accuracy.
• Implemented a semistochastic algorithm for efficient local energy evaluation, reducing computational cost.

Main Results:

• Achieved chemical accuracy in challenging molecular systems like H50, [Fe2S2(SCH3)4]2-, and H18.
• Demonstrated the effectiveness of the BDG-RNN ansatz and NNCs in enhancing NQS performance.
• Validated the computational efficiency and accuracy of the semistochastic energy evaluation.

Conclusions:

• The developed NQS framework with BDG-RNN and NNCs significantly advances quantum chemistry simulations.
• These innovations provide a more accurate and efficient approach for tackling complex molecular electronic structure problems.
• The open-source package PyNQS facilitates further research and application of these NQS methodologies.