Overview of Quantum Dots and the need for automation

This project addresses the challenge of automating the identification of reliable Quantum Dots (QDs) for quantum research, focusing on reducing manual effort in spectrum classification. QDs are nanoscale semiconductor particles important for quantum computing applications, particularly in single-photon generation. By leveraging classical data-science methods and variational autoencoders, we aim to develop a classification pipeline with consistent performance from minimal data, improving scalability and efficiency in QD research. This framework has potential applications across fields that face challenges with small datasets.

QDs are nanoscale semiconductor particles with quantum mechanical properties like discrete energy levels. Their tunable characteristics make them useful for optoelectronics, bioimaging, and quantum computing. At Tyndall, QDs are notably important for generating single photons used in entanglement experiments, a essential part of quantum technology development. Automating their identification addresses scalability and resource challenges, enabling broader experimentation and reducing manual labor in the research.

Experimental Setup

Figure: Lab bench where the chips are mounted and cooled to <10K.

QDs are made on a semiconductor chip, cooled, and studied under a microscope. The experimental setup is outlined in the image above. A 650 nm red laser is shone on the chip which excites electrons near the QD, which emit light as the electron hole pairs captured inside the QD recombine. This light is captured captured by the spectrometer where we see sharp peaks on a wavelength vs. intensity graph corresponding to states in the quantum dot. QDs can handle multiple electron-hole pairs, resulting in several peaks (up to 20) on the spectrometer graph.

Figure: Simplified Lab Setup: the laser light excites states in the quantum dot which then themselves emit light which we collect on the spectrometer

We then vary the laser power and take a spectrum at each power to create a power dependence heatmap, this is our working piece of data for analysis.

The Need for Automation

Measuring 50 heatmaps took 2.5 days in the lab; scaling to 500 would take nearly a month, which highlights a need for automation to improve scalability, free up resources, and enable more extensive experimentation. Note that the overall research program at Tyndall involves measurements of thousands of QDs across hundreds of chips. This manual process is slow and labor-intensive, automation data collection and analysis would free up lab time and allow for a greater range of experiments.

The automation involves two phases:

Automatic Spectrum Classification: Using 50 heatmaps, we plan to generate synthetic data and train a classifier using variational autoencoders and classical methods.

Microscope and Data Collection Automation: This will address hardware and data acquisition in a future project.

We propose using classical data-science methods and variational autoencoders to classify spectra due to their ability to handle small datasets effectively and capture complex patterns in the data. Compared to traditional rule-based or purely statistical approaches, these methods offer flexibility and improved accuracy by leveraging both structured feature engineering and representation learning. Our goal is to both automate the lab experiments and publish a scientific paper on the work. The quality of the publication and how ambitious we can be with journal submission depends on the quality of the results and interest to the community.

If we manage to create an automation pipeline with just 50 initial samples we believe it would be of create interest to the wider community because we would show that even with small data automation can be achieved with modern machine learning techniques.