Data-Driven Pretraining of Time-Dependent Neural Quantum States

Neural Quantum States (NQS) provide a powerful framework for representing many-body quantum systems, yet learning their real-time evolution remains in its early stages. A recent approach, the time-dependent Neural Quantum State (tNQS) [1], captures dynamics by optimizing time-independent parameters to satisfy the time-dependent Schrödinger equation. However, training NQS is often hindered by complex optimization landscapes and sensitivity to initialization.

Here, we introduce a data-driven pretraining strategy for tNQS, where time-resolved data from simulated experimental measurements provides a physically informed initialization for subsequent variational optimization. By incorporating observables—such as projective measurements and expectation values accessible in cold-atom platforms—we benchmark the speedup in tNQS convergence using the transverse-field Ising model in one and two dimensions.

Our results show that pretraining on minimal, physically motivated datasets yields robust initializations that accelerate and stabilize the learning of quantum dynamics, providing a scalable route for integrating experimental data with neural-network-based quantum simulations.

[1] https://doi.org/10.48550/arXiv.2412.11830