Explainable Models for Quantum Dot Qubit Readout

Many complex, real-world phenomena resist rigorous modeling, particularly when measurements are multivariate. In tuning gate-defined quantum dots (QDs), the connection between multiple gates and the dot behavior is very challenging to model theoretically, hindering automation. We propose relaxing the requirements for full theoretical models, focusing instead on phenomenological modeling. Such models, composed of simple functions such as sigmoids and Gaussians, can approximate a wide range of systems' behaviors. We then couple these surrogates with explainable machine learning (ML) to drive closed-loop autotuning, leveraging the open nature of explainable ML models to produce actionable predictions and suggestions for end users working in the lab. Focusing on the automation of Pauli spin-blockade readout, we use GAMs (generalized additive models) and GAMCoach to produce counterfactual explanations that map desired device states to actionable gate voltage changes needed to navigate device operational regimes towards high-fidelity readout configurations. We demonstrate this approach in controlled experiments on HRL-fabricated six-QD SLEDGE devices, showing the efficacy of explainable ML models in informing live tuning decisions pursuing optimal Pauli spin blockade. Our approach overcomes tuning challenges due to insufficient theoretical basis, offering a practical solution that leads to actionable tuning decisions.