Deep learning representations for quantum many-body systems on heterogeneous hardware

The quantum many-body problems are important for condensed matter physics, however solving the problems are challenging because the Hilbert space grows exponentially with the size of the problem. The recently developed deep learning methods provide a promising new route to solve long-standing quantum many-body problems. We report that a deep learning based simulation can achieve solutions with competitive precision for the spin J1-J2 model and fermionic t-J model, on rectangular lattices within periodic boundary conditions. The optimizations of the deep neural networks are performed on the heterogeneous platforms, such as the new generation Sunway supercomputer and the multi graphical-processing-unit clusters. Both high scalability and high performance are achieved within an AI-HPC hybrid framework. The accomplishment of this work opens the door to simulate spin and fermionic lattice models with state-of-the-art lattice size and precision.

Automated summary

Quantum many-body problems are important for condensed matter physics, but solving them is challenging due to the exponential growth of the Hilbert space with problem size. Deep learning methods offer a promising new approach to tackle long-standing quantum many-body problems. We demonstrate that a deep learning-based simulation can achieve competitive precision for the spin J1-J2 model and fermionic t-J model on rectangular lattices with periodic boundary conditions. The deep neural networks are optimized on heterogeneous platforms, such as the new generation Sunway supercomputer and multi-GPU clusters, achieving both high scalability and high performance within an AI-HPC hybrid framework. This work's accomplishment opens the door to simulate spin and fermionic lattice models with state-of-the-art lattice size and precision.