Berlin 2024 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

TT: Fachverband Tiefe Temperaturen

TT 2: Focus Session: Artificial Intelligence in Condensed Matter Physics I (joint session TT/DY)

TT 2.5: Hauptvortrag

Montag, 18. März 2024, 11:45–12:15, H 0104

Neural quantum states for strongly correlated systems: learning from data and Hamiltonians — •Annabelle Bohrdt¹, Hannah Lange², Schuyler Moss³, Fabian Döschl², Felix Palm², Giulia Semeghini⁴, Mikhail Lukin⁴, Sepehr Ebadi⁴, Tout Wang⁴, Fabian Grusdt², Juan Carrasquilla⁵, and Roger Melko³ — ¹Universität Regensburg — ²LMU München — ³UWaterloo — ⁴Harvard University — ⁵Vector Institute

Neural quantum states have emerged as a new tool to efficiently represent quantum many-body states with two main use cases: 1.) efficiently reconstruct a quantum state by training on measured data. For states with a non-trivial sign structure, measurements in many different basis configurations are necessary. I will present an active learning scheme which adaptively chooses the next measurement basis in order to maximize the information gain. 2.) The second main application of neural quantum states is to apply variational Monte Carlo to find e.g. the ground state of a system. I will present some of our recent results on ground states of strongly correlated systems, such as t-J type systems and fractional quantum Hall states. Finally, we combine both approaches: by first training on experimental data from a Rydberg atom tweezer array, we initialize the neural quantum state closer to the ground state. By then switching to variational Monte Carlo to minimize the energy in the second stage of training, we find a speedup in convergence. This showcases how limited datasets from experiments can be combined with numerical methods in a hybrid approach to yield more accurate results than either could provide on their own.

Keywords: Neural quantum states; interacting quantum many-body systems; hybrid learning; active learning