Presenters: Annika Böhler AND Anka Van de Walle
Probing real-time dynamics is a key goal of quantum simulation, but experiments with cold atoms, Rydberg arrays, or trapped ions often yield limited, or hard-to-interpret data. We introduce time-dependent neural quantum states (t-NQS): a machine learning method that reconstructs the full quantum state evolution from time-evolved snapshots. At the core of t-NQS is a neural network that represents the quantum wave function as a function of time. By using Variational Monte Carlo, the network is trained globally to satisfy the time-dependent Schrödinger equation across all times at once. This continuous representation enables efficient simulation using tools like automatic differentiation, while avoiding explicit time integration. Crucially, t-NQS naturally integrates with experimental data: snapshots of the system at various times can be used to pre-train the model, significantly improving convergence and accuracy in practice. This synergy goes further. By training on various time-evolved measurements at once, t-NQS can reconstruct the underlying wave function and phase structure, enabling access to observables which are experimentally inaccessible. It even allows for reconstruction of complex initial states, including entangled and topological ones, from data collected only during the system’s evolution. We demonstrate these capabilities by simulating quantum state preparation protocols relevant to current experiments, including the evolution of a 2D Heisenberg anti-ferromagnet under a staggered field and the preparation of a quantum spin liquid in a Rydberg atom array. These results highlight t-NQS as a versatile tool for extracting physics from limited data and enhancing the interpretability of quantum simulation experiments.
AND
Simulating strongly correlated electron systems remains a major challenge in condensed matter physics. Quantum simulation platforms have become powerful tools for exploring these systems beyond the limits of classical computation. In particular, ultracold alkaline-earth atoms and molecules now enable experimental access to Fermi-Hubbard models with SU(N)-symmetric interactions. However, theoretical understanding of these models remains limited, especially at finite dopings. In this talk, I present a neural network-based variational approach to the strongly correlated limit of the Fermi-Hubbard model, using a combination of Gutzwiller projection and a Hidden Fermion Determinant State ansatz (G-HFDS). I will show that this method captures magnetic and polaronic correlations in the SU(2) t-J model across the full doping range, and extend the framework to the SU(3) symmetric case, demonstrating that it can access finite-doping regimes that are challenging for conventional approaches yet highly relevant for state of the art quantum simulators. Our results provide a first step towards studying finite doping regimes of Fermi Hubbard models with higher SU(N) symmetries.
