Adversarial Hamiltonian learning of quantum dots in a minimal Kitaev chain

Determining Hamiltonian parameters from noisy experimental measurements is a key task for the control of experimental quantum systems. An experimental platform that recently emerged, and where knowledge of Hamiltonian parameters is crucial to fine-tune the system, is that of quantum dot-based Kitaev chains. In this work, we demonstrate an adversarial machine learning algorithm to determine the parameters of a quantum dot-based Kitaev chain. We train a convolutional conditional generative adversarial neural network (Conv-cGAN) with simulated differential conductance data and use the model to predict the parameters at which Majorana bound states are predicted to appear. In particular, the Conv-cGAN model facilitates a rapid, numerically efficient exploration of the phase diagram describing the transition between elastic co-tunneling and crossed Andreev reflection regimes. We verify the theoretical predictions of the model by applying it to experimentally measured conductance obtained from a minimal Kitaev chain consisting of two spin-polarized quantum dots coupled by a superconductor-semiconductor hybrid. Our model accurately predicts, with an average success probability of 97%, whether the measurement was taken in the elastic co-tunneling or crossed Andreev reflection-dominated regime. Our work constitutes a stepping stone towards fast, reliable parameter prediction for tuning quantum-dot systems into distinct Hamiltonian regimes. Ultimately, our results yield a strategy to support Kitaev chain tuning that is scalable to longer chains.