AI-Powered Quantum Chemistry Modeling by China’s Sunway Supercomputer: A Revolution in Large-Scale Simulation.
The Exponential Challenge of Quantum Simulation
It has long been a challenge to accurately calculate the ground state of interacting quantum matter. The “curse of dimensionality” refers to the fact that the Hilbert space dimension increases exponentially with the system size, making it extremely challenging to simulate quantum many-body systems.
The whole quantum behavior of complex systems cannot be computationally simulated on conventional supercomputers due to their complexity. Historically, thorough molecular investigations were limited to relatively small compounds due to the scaling constraints of current approximation approaches.
Researchers created Neural Quantum States (NQSs) to overcome these inherent challenges. NQS encodes the many-body wavefunction into artificial neural networks to approximate it.
By condensing the quantum state representation in terms of the network parameters instead of storing every possible coefficient, this method seeks to get around the exponential scaling barrier. This approach promises to combine the required quantum accuracy with the scalability of AI, creating a new avenue for research that was previously unattainable with conventional techniques. Numerous many-spin and many-fermion situations have been used to illustrate the efficacy of NNQS.
Scaling AI on the Sunway Supercomputer
Utilizing the synergy between AI and High-Performance Computing (HPC), Chinese scientists recently made a significant scientific and technological achievement. In particular, a new computational framework was proposed and modified for the Sunway supercomputer that uses Artificial Intelligence(AI) to handle quantum many-body issues.
The Oceanlite supercomputer, a member of the Sunway series of supercomputers, was used to carry out the project. A staggering 37 million compute processing element (CPE) cores were used in the simulation, with the potential to scale to 40 million heterogeneous cores. The framework was modified to accommodate the Sunway SW26010-Pro CPU’s special architecture, which consists of local memory and clusters of tiny compute cores that are intended for HPC applications rather than typical AI ones.
A customized NNQS framework tailored to the SW26010-Pro CPU was created by the researchers. They created a hierarchical communication model in which millions of “lightweight” CPEs carried out the local quantum computations, while management cores coordinated processors and nodes. Additionally, a dynamic load-balancing technique was put into place to stop cores from sitting idle because of unequal computing loads.
Across the 37 million cores, this endeavor produced 92% strong scaling and 98% weak scaling. A significant achievement for the supercomputing community, this remarkable efficiency shows a nearly flawless synchronization between the algorithms created and the hardware design of the supercomputer. The repetitive training loops necessary for deep learning, in particular, made this architecture extremely well-suited to the needs of NNQS computations. The difficult tasks of producing enormous quantities of random samples and figuring out the local energy required effective parallelization.
Modeling Molecular Chemistry at Unprecedented Scale
The successful large-scale implementation showed that NNQS may be employed on contemporary supercomputers for quantum physics research. At the scale of actual molecules, the scientists were able to accurately model intricate quantum chemistry. The simulation, which was the largest AI-driven quantum chemistry computation yet carried out on a classical supercomputer, concentrated on molecule systems with 120 spin orbitals.
In order to identify the most likely electron configurations and motions within the molecule, a neural network was trained to approximate the molecule’s wavefunction. Until its predictions matched the actual quantum energy pattern of the molecule, the system iteratively computed the local energy for sampling electron configurations and modified the network. This accomplishment demonstrates the possibility of addressing intricate chemical simulations with current conventional supercomputing capabilities, marking a significant advancement for China’s AI and quantum businesses.
Also Read About Quantum-Secured Optical Interconnects For AI Data Centers
The MinSR Optimization Breakthrough
Notwithstanding its potential, NQS had a significant drawback when it came to training large-scale, contemporary deep network topologies with high quantum state precision. Stable convergence in the complicated quantum landscape usually necessitates the stochastic reconfiguration (SR) technique, which is computationally costly and precludes training very deep networks.
Researchers developed the minimum-step stochastic reconfiguration (MinSR) technique to get around this fundamental optimization challenge. A logical reformulation of conventional SR intended for efficiency is MinSR. MinSR maintains the same high accuracy as traditional SR but decreases the optimization complexity by orders of magnitude. MinSR offers a significant speedup over conventional SR in the region where the number of parameters (Np) is significantly greater than the number of Monte Carlo samples (NS).
By training previously unheard-of deep networks, researchers were able to fully leverage the potential of NQS for this efficiency gain. MinSR made it possible to train NQS designs with over 10^5 parameters and up to 64 layers. Additionally, by utilizing the constrained memory space offered by contemporary GPU or TPU hardware, MinSR enables the effective acquisition of direct pseudo-inverse solutions, assisting in avoiding the use of enormous CPU resources.
Benchmarking Accuracy and Physical Insights
The deep NQS trained by MinSR outperformed all current variational approaches in terms of precision for the non-frustrated 10×10 Heisenberg model, obtaining a relative error level of 10 −7. The deep NQS trained by MinSR produced variational energies that surpassed all previous neural network quantum states and tensor networks, as well as all other numerical results, for the difficult, highly frustrated scenario (J2 /J1 = 0.5).
Even when scaling up to a big 16×16 lattice, this better computing benefit remained. Strong numerical evidence indicating the occurrence of gapless QSLs in both the square and triangular lattice J1-J2 models at their most frustrated places was provided by these incredibly accurate simulations, which also produced novel physical insights.
Future Directions
This work shows how machine learning and sophisticated classical supercomputing architecture can function in strong harmony. The success of deep NQS enabled by optimization techniques such as MinSR indicates a promising avenue for resolving quantum many-body issues that were previously unsolvable.
Future work will focus on developing ab initio quantum chemistry, especially in the strongly interacting region where conventional approaches fail, or expanding this methodology to fermionic systems, such as the well-known Hubbard model. In addition to not being specific to NQS, the MinSR method is a general optimization tool for variational Monte Carlo methods that may improve the expressivity of other traditional approaches, such as tensor networks.
Furthermore, MinSR’s optimization principles could be used for a variety of machine learning applications, like giving reinforcement learning a more precise natural policy gradient. In the end, this development provides a means to use currently available classical resources to greatly speed up innovation in important areas like medication design and materials discovery.
