Robots in a condensed matter lab

My optimism about applications of machine learning to condensed matter physics is very cautious. In an earlier commentary [1] I argued that condensed matter physics is unlike identifying cat photos, and therefore we might not benefit from going further than a robust linear regression. I believe the main reason for my argument still holds: because we deal with controlled problems and know a large part of the answer in advance, our data is often too simple to warrant machine learning. When writing this, I was however mostly considering applications of machine learning to theory. Experiments are a different story, and they require a lot of work invisible to a theorist’s eye. The largest hurdle in advanced experiments is not the observation of the relevant phenomenon, but bringing the device to a correct regime and calibrating it. I have once visited an optics lab, where the Ph. D. students would spend every single morning measuring the parameter drift that occurred overnight and compensating for it. Then after lunch, an afternoon of measurements and a night of parameter drift, everything would repeat again. Likewise tuning a system of multiple quantum dots may take months of painstaking work regardless of whether the goal is to observe Kondo physics, make a spin qubit, or simulate a Hubbard model. Both works that I discuss tackle this problem using an automated approach. In the work of van Diepen et al. the authors use something I would call a physicist’s approach: their tuning protocol relies on a physics-based model of their device. The algorithm starts by identifying areas in the space of gate voltages that matches to different charge states. Then it fits all the observed anticrossings between energy levels to an effective model in order to determine couplings of different gates to individual levels and to define the so-called “virtual gates” that only control a single dot. At the last step, iterative tuning