Automated effective circuit Hamiltonian learning with autoencoders

High-precision superconducting quantum processors require delicate circuit QED analysis to capture interactions originating from the low frequency domain (<10 kHz). By block-diagonalizing the circuit's Hamiltonian, effective parameters are deduced for the degrees of freedom of interest. Conventional methods, such as the Schrieffer-Wolff transformation, are limited to the perturbative regime, while alternatives, though more accurate, do not scale well beyond two qubits. With the aim to enable circuit analysis on larger systems to improve the current experiments, we propose a novel numerical approach relying on machine learning. An autoencoder network is trained to determine the effective Hamiltonian on multiple qubits in order to simulate the behavior of different gates.