Solid-state quantum simulators skip hand fitting: an autoencoder reads effective Hamiltonians from transport data

Characterizing solid-state quantum simulators shifts from hand fitting to automatic inference of effective Hamiltonians, turning measured transport curves on programmable devices into device parameters. Earlier practice relied on known or labeled Hamiltonian parameters to tie data to model, and this route needs no such labels. In numerical experiments on a triple quantum dot chain, the model accurately infers Hamiltonian parameters and generalizes beyond the training domain.

The physics constraint sits inside the decoder rather than serving as a post-hoc check, so the parameters the network emits carry physical meaning by construction. An unsupervised autoencoder is used, with the S-matrix formalism embedded directly in the decoder; the S-matrix describes scattering and is the physical language of transport measurement. Training and evaluation both run on numerical transport data from the triple quantum dot chain, including noisy measurement conditions.

The model also handles data generated outside the Hamiltonian family its physics decoder assumes, still returning meaningful effective parameters. This goes beyond both the training distribution and the assumed model family, so inference is not confined to the Hamiltonian class that generated the data. Trained on transport data generated outside that Hamiltonian family, the model still identifies meaningful effective parameters.

The demonstrated scope is numerical transport experiments on a triple quantum dot chain, covering parameter inference, generalization beyond the training domain, robustness to noisy measurements, and training on data generated outside the assumed Hamiltonian family.

Model capacity is finite as the parameter space grows broader or more diverse; training across Hamiltonian families yields meaningful effective parameters rather than parameters of the original family.