Technical Papers Archive

Scaling Neural Network Quantum States to 8,192 GPUs: Treating Transition-Metal Clusters in Hilbert Spaces Beyond 10³¹ Dimensions


Authors: Huan Ma (University of Science and Technology of China); Bowen Kan (Institute of Computing Technology (ICT), Chinese Academy of Sciences (CAS)); Lizhong Fu (University of Science and Technology of China); Yingqi Tian (Institute of Computing Technology (ICT), Chinese Academy of Sciences (CAS)); Jielan Li (Shanghai AI Laboratory); Peiyun Pan, ChenBo Tao, Yongqing Chen, and Anquan Wang (National High Performance Computer Engineering Technology Research Center); and Honghui Shang and Jinlong Yang (University of Science and Technology of China)

Abstract: Computing ground-state energies of strongly correlated transition-metal clusters remains a central challenge due to exponentially growing Hilbert spaces. We present a scalable neural network quantum state (NNQS) framework based on a Transformer-encoder wavefunction with neural-network backflow, targeting large-scale heterogeneous supercomputers. An asynchronous CPU–GPU pipeline, mixed-precision optimization, semi-stochastic coupling for local energy evaluation, and a scalable distributed second-order optimizer collectively achieve 98.9% weak-scaling efficiency on 4,096 and 88.94% strong-scaling efficiency on 8,192 GPU accelerators. We compute ground-state energies for three metalloenzyme active sites, including the P-cluster of nitrogenase (146 spin orbitals, 114 electrons)—doubling the previous NNQS system-size frontier into Hilbert spaces beyond 10³¹ dimensions. Our results establish NNQS as a viable framework for strongly correlated quantum chemistry at unprecedented scales.


Presentation: file

Back to Technical Papers Archive Listing