Transfer learning for scalability of neural-network quantum states

Area of Science:

• Quantum physics
• Machine learning
• Computational science

Background:

• Neural-network quantum states (NQS) are powerful tools for simulating many-body quantum systems.
• Scalability challenges limit the application of NQS to larger and more complex systems.
• Transfer learning in machine learning offers methods to reuse trained models for new tasks.

Purpose of the Study:

• To investigate the potential of transfer learning to enhance the scalability of NQS.
• To develop and present novel, physics-inspired transfer learning protocols for NQS.
• To evaluate the efficiency and effectiveness of these protocols in quantum system simulations.

Main Methods:

• Devised physics-inspired transfer learning protocols for NQS.
• Implemented protocols using restricted Boltzmann machines on graphics processing units (GPUs).
• Evaluated protocols on the transverse field Ising and Heisenberg XXZ models in 1D and 2D, up to 128 spins.

Main Results:

• GPU implementation alone provided speedups over CPU-based methods.
• Certain transfer learning protocols significantly improved efficiency and accuracy compared to random initialization.
• Demonstrated effectiveness across different quantum models and phases with increasing system sizes.

Conclusions:

• Transfer learning is a viable strategy to improve the scalability of NQS.
• Physics-inspired transfer learning protocols offer a more effective and efficient approach for quantum simulations.
• The developed methods pave the way for studying larger quantum systems.