QuTech team solves lattice gauge theory with constrained neural networks

QuTech researchers, in collaboration with ETH Zurich, have developed a constrained neural network that can solve complex lattice gauge theories. This is an important step toward overcoming limitations in quantum mechanical simulations. The research team tackled a core paradox in quantum computing: the need for more computational power to model the very systems these computers are designed to build. This method focuses computational resources on the features that truly determine the energy levels in the model by designing a neural network that automatically ignores nonphysical variations. “A lot of what computers consider is just another way of describing the same situation,” says QuTech’s Thomas Spriggs. This approach, which preserves a continuous field description and incorporates symmetry constraints directly into the network’s architecture, achieves lower energies than traditional symmetry-based baselines and provides an important basis for validating future quantum simulations, explains lead researcher Elishka Greprova. “If quantum processors are to simulate nature in areas where classical methods are difficult, they need reliable reference calculations and clear validation targets.”

Lattice gauge theory realizes physically constrained neural networks

New applications of artificial intelligence offer potential solutions to the long-standing challenge in quantum computing of accurately modeling complex physical systems. Researchers from QuTech and ETH Zürich have developed a neural network specifically designed to estimate quantum states, bypassing the limitations of current computational power. The research team focused on lattice gauge theory, a complex model in particle physics that discretizes space into a grid. A “field” exists on the connections between grid points. An important aspect of this theory is the local freedom to change descriptive labels without changing measurable outcomes. This freedom is essential for physics, but it creates computational difficulties because algorithms can get bogged down when modeling nonphysical variations. This new approach avoids traditional computational bottlenecks by embedding physical constraints directly into the neural network’s architecture, automatically ignoring irrelevant changes and allowing you to focus on the features that truly impact energy levels.

The method employs a feedback loop in which the network is trained to suggest potential quantum states, sample field configurations, and prioritize configurations that minimize energy. Essentially, “the laws of physics give you a score,” Spriggs elaborates. “The network suggests what each configuration should be, computes the energy it implies, and trains the network until it reliably chooses the lower energy state.” Importantly, the researchers maintain a continuous field description, avoiding both discretization and the “sign problem” that is a problem often encountered in these simulations.

Their recent work focuses on lattice gauge theory, a particle physics model in which space is represented as a grid, and the inherent freedom to relabel grid points poses significant computational difficulties. The technique also avoids the notorious “sign problem” often encountered in quantum simulations, demonstrates success in both two-dimensional and three-dimensional models, achieves lower energies than traditional symmetry-based methods, and is consistent with established theoretical predictions.