⚛️ quantum physics

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

This paper benchmarks sequential Bayesian inference, Levenberg-Marquardt fitting, and a feedforward neural network for classifying single-photon emitters from sparse autocorrelation data, demonstrating that while no single method dominates all metrics, physics-based and data-driven approaches are complementary tools that offer distinct advantages in convergence speed, interpretability, and robustness under varying photon statistics.

Original authors: Nhat Minh Nguyen, Md Shakhawath Hossain, Duc Anh Ngo, Chaohao Chen, Xiaoxue Xu, Toan Trong Tran, Carlo Bradac

Published 2026-08-21

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Nhat Minh Nguyen, Md Shakhawath Hossain, Duc Anh Ngo, Chaohao Chen, Xiaoxue Xu, Toan Trong Tran, Carlo Bradac

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quest to build the next generation of quantum technologies, scientists rely on tiny, specialized light sources known as single-photon emitters. These are microscopic objects, such as defects in solid materials or individual atoms, that release exactly one particle of light at a time. This precise behavior is the foundation for secure communication, powerful computing, and ultra-sensitive sensing. However, finding these rare emitters within a vast, messy landscape of potential candidates is a difficult task. Researchers must distinguish a genuine single-photon source from a cluster of emitters acting together, a process that requires measuring how the light particles arrive in time. The standard method involves counting how often two detectors click simultaneously, a technique that reveals the purity of the light. Yet, this measurement is slow and statistically fragile; if the light is dim or the measurement time is short, the data becomes sparse and noisy, making it hard to tell a true single emitter from a group of impostors.

A team of researchers at the University of Technology Sydney and Trent University set out to solve this identification problem by comparing three different ways to analyze these sparse light measurements. They wanted to know if modern machine learning could outperform traditional physics-based methods when data is scarce, or if the two approaches offered different strengths. To test this fairly, they created a massive library of simulated measurements based on real experiments with hexagonal boron nitride, a material known to host these quantum emitters. Because they generated the data themselves, they knew the exact number of emitters in every single case, allowing them to see exactly which method made the right call and which made a mistake. They then pitted a classic mathematical fitting technique against a new, step-by-step statistical method and a neural network, a type of artificial intelligence trained to recognize patterns.

The results revealed that no single method is perfect for every situation, and the best choice depends entirely on how much time the researcher has to wait for data. When the measurement time is extremely short and the data is very sparse, the neural network proved to be the most robust. It was the only method that avoided a specific trap where the other two approaches would mistakenly guess that almost everything was a single emitter simply because the data was too thin to support a different conclusion. In this early, difficult stage, the neural network made far fewer false alarms, correctly identifying that many candidates were actually groups of emitters rather than single ones. However, as the measurement time increased and more data accumulated, the playing field leveled out. All three methods eventually reached a state of near-perfect accuracy, correctly identifying the emitters with high reliability.

The differences between the methods became clear when looking at how they reached that accuracy. The neural network was a black box; it learned to recognize the shape of the light pattern without explaining why, offering speed and resilience but no insight into the physical properties of the light. In contrast, the new sequential Bayesian method offered a different kind of advantage. It treated the measurement as a story that unfolds over time, updating its confidence with every new piece of evidence as it arrived. This approach allowed it to converge on the correct answer faster than the traditional method and with a smoothness that the others lacked, all while keeping a clear, physical explanation for its decision. The classic mathematical fitting method, while the slowest to reach a reliable conclusion and most prone to errors when data was scarce, ultimately proved to be the most thorough at finding every single true emitter once enough data had been collected, even if it occasionally made more false alarms along the way.

The study suggests that relying on just one metric, such as how often a method finds a true single emitter, can be misleading. A method might find many true emitters but also falsely label many groups as singles, which is a critical failure for building quantum devices. The researchers found that by combining the predictions of all three methods—essentially taking a vote where the majority decides the outcome—they could reduce the total number of mistakes below what any single method could achieve alone. This hybrid approach leverages the unique strengths of each strategy: the neural network's ability to handle sparse data, the Bayesian method's rapid and interpretable convergence, and the classic method's high sensitivity. Ultimately, the work demonstrates that physics-based and data-driven approaches are not rivals but complementary tools, and that the most effective strategy for screening quantum emitters may well be to use them together.

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

Problem Statement
The scalable realization of quantum technologies (e.g., quantum key distribution, computing, and sensing) requires the rapid and reliable identification of single-photon emitters (SPEs) from large, inhomogeneous populations of candidate nanoscale sources. Standard identification relies on measuring the second-order autocorrelation function, $g^{(2)}(\tau)$ , using a Hanbury Brown–Twiss (HBT) interferometer. The purity of emission is quantified by the zero-delay value, $g^{(2)}(0)$ , where an ideal single emitter yields 0, and an ensemble of $N$ emitters yields $1 - 1/N$ .

However, statistical reliability is fundamentally limited by acquisition time. Because coincidence rates scale with the square of the emission rate, acquiring sufficient counting statistics for low-efficiency systems requires long integration times, creating a trade-off between measurement throughput and the number of surveyable candidates. While machine learning (ML) has been proposed to accelerate identification from sparse data, its performance relative to established physics-based inference methods has not been systematically examined under realistic noise and background conditions with known ground truth.

Methodology
The authors benchmark three distinct classification approaches for distinguishing single ( $N=1$ ) from multi-emitter ( $N \ge 2$ ) sources:

Levenberg–Marquardt (LM) Fitting: A standard physics-based approach that fits a parametric model of the autocorrelation function to the coincidence histogram via weighted nonlinear least-squares.
Sequential Bayesian Inference: A newly introduced physics-based approach that treats classification as sequential inference. It updates the posterior probability of the "single" vs. "multi" hypotheses as new coincidence data accumulates incrementally, utilizing a reduced single-parameter model for each data increment.
Feedforward Neural Network (MLP): A data-driven approach using a multilayer perceptron to classify emitters directly from the shape of the coincidence histogram without an explicit physical model.

Data Generation and Benchmarking
To enable rigorous benchmarking with an exactly known ground truth (which is unavailable in real experimental data), the authors generated 5,000 synthetic autocorrelation datasets. These were calibrated against real HBT measurements from hexagonal boron nitride (hBN) emitters. The simulation incorporated realistic background-to-signal ratios, photon-extraction efficiencies, and detection noise derived from experimental data. The dataset included 1,002 single-emitter and 3,998 multi-emitter cases, spanning integration times from sparse (iteration 10) to fully converged (iteration 300).

Key Contributions

Introduction of Sequential Bayesian Inference: The paper proposes a Bayesian framework specifically tailored to the incremental nature of HBT data acquisition, allowing for evidence chaining without refitting the entire dataset at every step.
Systematic Comparison of Paradigms: The study provides a direct comparison between physics-based inference (LM and Bayesian) and data-driven inference (MLP) across multiple performance metrics: recall, precision, and leakage (the rate of falsely certifying multi-emitters as single).
Analysis of Failure Modes: The authors identify a specific "collapse" failure mode in physics-based methods under extreme sparsity, where low counts bias the fitted $g^{(2)}(0)$ toward a deep antibunching dip, artificially inflating recall while simultaneously increasing leakage.
Ensemble Strategy: The paper demonstrates that combining the three methods via a simple majority vote leverages their partially independent failure modes to improve overall classification performance.

Results

Sparse Data Regime (Short Integration Times):
- The MLP is the most robust, exhibiting the lowest leakage (0.0623) compared to LM (0.2016) and Bayesian (0.1913). It avoids the "collapse" failure mode seen in the physics-based methods.
- However, the MLP has lower recall (0.6317) at this stage.
- The higher recall of LM and Bayesian at sparse times is identified as a misleading artifact of the collapse mechanism rather than genuine performance.
Converged Regime (Long Integration Times):
- All three methods achieve high, near-perfect accuracy.
- Bayesian Inference reaches stable, near-perfect precision (0.99) fastest (by iteration 60) and exhibits the smoothest convergence curve (only 1 reversal), owing to its sequential evidence accumulation.
- LM Fitting achieves the highest recall (0.9870) at full convergence but is the slowest to converge and suffers from the highest leakage among the three at sparse times.
- MLP achieves high precision (0.9979) and low leakage (0.0005) but requires more iterations (70) to reach 0.99 precision and shows a jagged convergence curve.
Error Analysis:
- No method achieves perfect recall even at full convergence; a small fraction of single emitters (1.3% to 3.6%) remains misclassified.
- Misclassifications are only partially correlated between methods. A majority-vote ensemble reduces the total number of missed single emitters to 22 (from 25 for Bayesian and 36 for MLP alone), demonstrating the value of combining approaches.

Significance and Conclusions
The paper concludes that no single method dominates across all performance metrics. The choice of classifier depends on the specific operational priorities of the application:

MLP is preferred for minimizing false verification at the shortest possible integration times (critical for high-throughput screening).
Sequential Bayesian Inference is preferable when rapid, smooth convergence to high precision is required while maintaining full physical interpretability.
LM Fitting remains a valuable, fully interpretable baseline that achieves the highest recall once sufficient data is accumulated.

The authors emphasize that relying on a single metric (such as recall alone) can be misleading, particularly under sparse photon statistics. Instead, physics-based and data-driven methods are presented as complementary rather than competing. The study suggests that combining predictions through simple ensemble methods can effectively mitigate individual weaknesses, providing practical guidance for scalable single-photon-source screening.

More like this