Can neural networks speed up quantum simulations on a lattice?

Researchers built a flow-based proposal mechanism—essentially a learnable transformation of quantum fields—that feeds into Monte Carlo sampling of SU(2) gauge theories. The method stays mathematically rigorous by preserving the underlying symmetry (Haar measure) and works with frozen-link backgrounds. In 2D tests, it matches conventional algorithms and shows small improvements in some configurations, establishing a foundation for scaling to larger systems and realistic QCD simulations.