Closed-Loop, Event-Driven Machine Learning for Autonomous Triage of Single-Photon Emitters

Abstract Identifying which candidate sites on a solid-state sample are single-photon emitters requires a Hanbury Brown–Twiss (HBT) measurement of the second-order correlation function g(2)(τ), which is slow because photon coincidences accumulate at a rate proportional to the square of the detected count rate. We examine how much of this time can be saved by deciding while the coincidences are still accumulating and by adapting the exposure of each site. A spiking neural network reads the coincidence record in short time slices and updates, after every slice, its estimate of the probability that the site hosts a single emitter; a reinforcement-learning controller decides, site by site, whether to expose longer, certify, or reject; and a graph that encodes the electronic level structure of each emitter class conditions one estimator on the class being measured. All components are trained on simulated data from a standard three-level photon-counting model, which we validate against the exact master-equation solution before use. In simulation, the event-driven estimator reaches 90% balanced accuracy with 6 times less acquisition time than a multistart least-squares fit, typically committing to a decision 312 ms into a 1 s exposure, and adaptive exposure screens simulated 48-site fields 1.7 times faster than fixed-dwell scanning at matched precision and recall. Applied without retraining to published experimental HBT data from an InGaAs/GaAs quantum dot, a network trained only on simulated pulsed-excitation histograms estimates g(2)(0) from 30 s records with half the error of peak-area analysis of the same records. The closed-loop results are obtained in simulation only. The code, data, and trained models are openly available.