Convolutional RBM achieves 10x speedup in simulating frustrated lattice systems

Geometric frustrations in materials lead to fascinating and often unexpected physical properties, but simulating these complex systems poses significant computational challenges. Pratik Brahma, Junghoon Han, and Tamzid Razzaque of the University of California, Berkeley, along with Saavan Patel of Berkeley and InfinityQ Technology Inc., and Sayeef Salahuddin of Berkeley and Lawrence Berkeley National Laboratory, are demonstrating a powerful new approach to overcome these limitations. Crucially, the team developed a convolution-restricted Boltzmann machine, a type of neural network that efficiently captures the inherent symmetries of frustratis, and implemented this on purpose-built digital hardware. This advance enables the simulation of systems containing hundreds of interacting spins, accurately reproducing the known phases of matter and characterizing complex spin behavior, while achieving speedups of three to five orders of magnitude compared to traditional computer simulations. This research establishes a path to scalable and reprogrammable hardware that can simulate large-scale physical systems with unprecedented efficiency, opening new avenues of materials discovery.

Frustrated magnetism gives rise to exotic phases such as spin liquids, and accurately simulating these systems requires computational power, especially for large lattices. The researchers harnessed the power of RBM, a type of generative neural network, to learn the probability distribution of low-energy states in the Shastri-Sutherland model, a system known to exhibit spin-liquid behavior. Implementing RBM on FPGAs enables parallel processing and custom hardware design, overcoming the limitations of traditional CPU- or GPU-based simulations and offering potential speedups and energy efficiencies. Efficient data transfer and optimized memory management were essential to achieving high performance. Recognizing the limitations of traditional Monte Carlo methods, the team turned to machine learning to improve sampling efficiency. They pioneered a CRBM formulation that directly captures interactions within the SS lattice and exploits its inherent translational symmetry to create a more efficient probabilistic neural network. This approach is in contrast to fully connected RBM, which is inefficient when representing large systems.

The core innovation lies in implementing a convolutional filter that is amenable to unit cell size, effectively capturing physical interactions, and reducing the number of free parameters that are independent of system size. This design significantly improves both representation and Monte Carlo sampling, allowing faster convergence and more uncorrelated samples. To speed up the simulations, the scientists designed a dedicated digital hardware accelerator for the CRBM architecture, allowing for simultaneous updates of the entire spin lattice. In experiments involving lattices containing up to 324 spins, we successfully recover all known phases of the SS Ising model, including the long-range ordered fractional plateau. This hardware characterizes the behavior of spins at critical points and in the spin liquid phase, achieving speedups of 3 to 5 orders of magnitude compared to GPU-based implementations, while providing excellent scalability, room temperature operation, and reprogrammability. They successfully simulated a lattice containing 324 logical spins and recovered all known phases of the Shastri-Sutherland (SS) Ising model, a geometrically frustrated system exhibiting exotic magnetic behavior. This achievement confirms the ability of CRBM hardware to accurately represent and explore the complex energy landscape of stressed systems and validates the potential of CRBM hardware in materials discovery and fundamental physics research. The core of this progress lies in the ability of CRBM to efficiently capture interactions within the SS lattice, a system characterized by competing nearest-neighbor and next-nearest interactions.

By exploiting the translational symmetry of the lattice, the team designed a convolutional filter that accommodates the size of a unit cell, resulting in a more efficient representation compared to a fully connected network. This design significantly reduces the number of parameters required to represent the system, allowing faster Monte Carlo sampling, regardless of the system size. Experiments demonstrated speedups of 3 to 5 orders of magnitude compared to comparable algorithms running on graphics processing units, achieving processing times ranging from 33 nanoseconds to 120 milliseconds. Additionally, this hardware characterizes the behavior of spins at critical points and within the spin liquid phase, providing insight into new properties of these materials.

Analysis of the spin structure factors of the ground state configuration confirms the connection with experimental observations such as diffuse neutron scattering. The team’s results demonstrated that the quantum annealer’s performance is within an order of magnitude while offering excellent scalability, room temperature operation, and the ability to be reprogrammed for a variety of systems. Focusing on the Shastri-Sutherland Ising lattice, the researchers successfully analyzed the spin configuration, explored the emergent classical spin liquid phase, and determined the phase diagram, achieving results comparable to established simulation methods. The hardware implementation provides significant speedups of 3 to 5 orders of magnitude (33 ns to 120 ms) compared to traditional GPU-based sampling techniques across different magnetic phases. This speedup comes from key architectural features such as optimized bitwise operations, fixed-point weight representation, and, importantly, exploitation of translational symmetry within the system.

The performance is within two orders of magnitude over that achieved by quantum annealers, with advantages in terms of scalability and the ability to be reprogrammed for different simulations. Seamless integration with standard CPU hosts enables acceleration of variational Monte Carlo algorithms where CRBM hardware acts as a variational wave function. The researchers acknowledge that the current implementation is specific to systems exhibiting translational symmetry, limiting direct application to all materials. Future research will focus on extending the architecture to accommodate more complex symmetries and exploring the possibility of simulating a wider range of quantum materials and frustrate lattices, ultimately enabling faster and more efficient large-scale simulations.