Scaling ptVMC with autoregressive neural-network quantum states beyond L=5

Demonstrate convergence and stability of projected time-dependent variational Monte Carlo (ptVMC) for real-time dynamics using autoregressive neural-network quantum states (such as recurrent neural networks or transformers) on two-dimensional disordered transverse-field Ising models for system sizes larger than L=5, overcoming the observed optimization-convergence failures in infidelity minimization.

Background

The authors investigated ptVMC, which learns the action of the short-time propagator by minimizing an infidelity objective between successive time steps. While computationally appealing, this method requires a reliable optimization that may fail to converge.

For autoregressive architectures (RNNs, transformers), the authors report convergence difficulties that prevented simulations beyond L=5. Establishing a stable, scalable ptVMC procedure for these models would enable larger, more expressive NQS dynamics studies.