Adversarial machine learning for modeling the distribution of large-scale ultracold atom experiments

Directly generating microstates with desired properties from the configuration space of many-body systems is infeasible due to its high-dimensional nature. Instead, traditional generation methods rely on computationally costly algorithms or carefully controlled experimental setups, which limits the number of particles that can be investigated.

We present how artificial neural networks allow for the direct and targeted generation of large-scale microstates, while restricting the time-consuming simulations or measurements to a small number of particles. Their potential is illustrated on a data set of experimental snapshots of a doped Fermi-Hubbard model realized by ultracold atoms trapped in an optical lattice. The adversarial machine learning method we develop here is broadly applicable and can also be used for speeding up computer simulations of both equilibrium and nonequilibrium physical systems.