Physics-Based versus Data-Driven Classification of Single-Photon Quantum Emitters from Sparse Autocorrelation Data

This paper asks a practical question for quantum-emitter screening: when autocorrelation data are sparse, should one rely on a physics-based fit, a Bayesian update rule, or a data-driven neural network to decide whether a candidate source is a true single-photon emitter? The motivation is that measuring the second-order autocorrelation function, g(2)(τ), takes time, and long acquisitions limit how many candidates can be screened in practice.

The authors benchmark three classifiers on the same binary task: single emitter versus multi-emitter. The physics-based methods are Levenberg–Marquardt fitting of a reduced autocorrelation model and a new sequential Bayesian inference scheme. The data-driven method is a feedforward neural network trained on compact histogram features.

A key strength of the study is its data construction. The team calibrated synthetic training and test data using real Hanbury Brown–Twiss measurements from hexagonal boron nitride emitters, so the simulated histograms include realistic background, noise, and detection inefficiency while still having known ground-truth emitter numbers. That lets them compare methods fairly across a range of integration times, from very sparse counts to well-converged measurements.

All three methods eventually reach high, often near-perfect, accuracy when enough photons are collected. The important difference is how quickly they get there and how they behave when the data are still sparse.

The Bayesian classifier is the fastest to reach stable, near-perfect performance once enough evidence accumulates, and it keeps the physical meaning of the autocorrelation model. The neural network is the most robust at the shortest integration times, where the physics-based fits are most fragile. Levenberg–Marquardt converges the slowest, but it achieves the highest recall, meaning it is especially good at finding true single emitters.

The paper also finds that no single metric tells the full story: a method can look strong on one measure and weak on another. Combining all three predictions with a simple majority vote improves overall performance, suggesting that the methods make partly independent errors.