Machine-learned tuning of artificial Kitaev chains from tunneling spectroscopy measurements

We demonstrate reliable machine-learned tuning of quantum-dot-based artificial Kitaev chains to Majorana sweet spots, using the covariance matrix adaptation algorithm. We show that a loss function based on local tunneling spectroscopy features of a chain with two additional sensor dots added at its ends provides a reliable metric to navigate parameter space and find points where crossed Andreev reflection and elastic cotunneling between neighboring sites balance in such a way to yield near-zero-energy modes with very high Majorana quality. We simulate tuning of two- and three-site Kitaev chains, where the loss function is found from calculating the low-energy spectrum of a model Hamiltonian that includes Coulomb interactions and finite Zeeman splitting. In both cases, the algorithm consistently converges towards high-quality sweet spots. Since tunneling spectroscopy provides one global metric for tuning all on-site potentials simultaneously, this presents a promising way towards tuning longer Kitaev chains, which are required for achieving topological protection of the Majorana modes.