LEEX · Lattice QCD simulations at the dawn of European Exascale Computing

Large-scale simulations of lattice Quantum Chromodynamics (QCD) can provide crucial input for searches of new physics at the precision frontier, like in the calculation of the hadronic contribution to the anomalous magnetic moment of the muon. To find evidence of new physics beyond the current knowledge provided by the Standard Model will require to reach per-mille precision. Attaining such accuracy in the hadronic contribution to the anomalous magnetic moment of the muon in lattice QCD can only be achieved by controlling lattice systematics, such as logarithmic terms in the continuum extrapolation, finite volume effects and exponentially increasing noise at large time distances. Simulating gauge ensembles at lattice spacing a<0.04 fm is impossible with current algorithms due to topological freezing. Generative models based on gauge equivariant flows can unfreeze the charge but scale badly with the volume, limiting, up to now, their applicability to toy models. We propose a solution to overcome topological freezing and suppress exponential noise by developing scalable algorithms for lattice QCD simulations closer to the continuum limit and at physical quark masses based on domain decomposition. By combining machine-learned flow proposals with hierarchical accept/reject steps of the factorized fermion determinant, ensembles at very small lattice spacings can be generated using upcoming european exascale supercomputers. We will implement multi-level sampling techniques within a flexible framework to enable good performance of the here developed novel Markov Chain Monte Carlo algorithm on exascale systems. By the newly generated gauge ensembles lattice QCD will provide a leap in the precision frontier thereby critically contributing in the unraveling of new physics.