Phys. Rev. Lett. (in the press); preprint at https://arxiv.org/abs/1904.08441 (2019)

The dream of future quantum machines that can calculate or simulate what has so far been beyond our reach runs into several rather practical obstacles. The experimental challenges in controlling large quantum states are well known, but the equally relevant problem of state reconstruction — finding a way to efficiently determine the state of a device — often goes unmentioned.

In an experiment with a nine-atom programmable Rydberg quantum simulator, Giacomo Torlai, Brian Timar and co-authors have now shown that neural networks can be a great asset for state reconstruction. The team parametrized the almost-pure state of a one-dimensional array of strongly interacting neutral atoms with a restricted Boltzmann machine — a two-layer stochastic neural network. After adding a third layer to account for noise, they were able to reconstruct the state of the system by training the network with real experimental data. The trained network was then able to reach beyond experimental limitations and predict complex observables whose measurement requires specialized equipment.