ECE PhD Thesis Defense: Shashwath Shankar Bharadwaj

ECE PhD Thesis Defense: Shashwath Shankar Bharadwaj

Title: Signal Processing for Readout-Multiplexed Single-Photon Detector Arrays

Presenter: Shashwath Shankar Bharadwaj

Advisor: Professor Vivek Goyal

Chair: Professor Alexander Sergienko

Committee: Professor Vivek Goyal, Professor Tianyu Wang, Professor David Castañón, Dr. Adam McCaughan (NIST)

Google Scholar Link: https://scholar.google.com/citations?user=bVktye0AAAAJ&hl=en

Abstract: Kilopixel and megapixel scale cameras based on superconducting nanowire single-photon detectors (SNSPDs) are desirable due to properties such as near-unity quantum efficiency, low jitter, and high timing resolution. Readout multiplexing has been critical to the first demonstrations of large-scale SNSPD arrays by mitigating the high heat loads associated with fully-addressed arrays. However, ambiguities in the spatial locations of multiple coincident photons have restricted their use to low incidence flux regimes. In this thesis, we develop novel signal processing techniques to resolve multiphoton coincidences and enable high-resolution, high-flux, and low-latency imaging with SNSPD arrays.

We first study row–column multiplexing, where each row and each column of an array are read out instead of every pixel individually. For an array of size n x n, this scheme reduces the required number of readouts to 2n. We introduce a multiphoton estimator (ME) that probabilistically resolves up to 4-photon coincidences by maximizing an approximate likelihood of observations. With this estimator, we achieve 1) 3 to 4 dB increase in reconstruction PSNR at optimal attenuation compared to conventional baselines, 2) Four-fold decrease in integration time, and 3) a close match to the Cramer-Rao bound under a wide range of incident fluxes.

While the ME provides a physically principled estimation method, it is limited in application due to complex combinatorial computations and dependence on noisy unambiguous readouts. This motivates the study of a combinatorics-free method that can work with limited readouts and at higher incidence fluxes. We introduce a novel input representation method called the context tensor that makes multiplexed readouts suitable for neural network processing. We demonstrate that a convolutional neural network is able to process the context tensor to achieve a 4 to 8 dB improvement over the ME in cases where the ME is constrained by the number of readout frames and the number of incident photons modeled.

Finally, we study time-of-flight multiplexing where an n x n array is fully addressed using just 2 readout buses. Photon detections on the row layer are thermally coupled to the column layer, leading to a stream of measured time tags at both ends of the readout buses. The time tags encode both the pixel position of the incident photon and its arrival time at the array. We then develop a novel MMSE-style soft-decoding method that exploits the array geometry to disambiguate between coincident photons. We leverage the distribution of thermal coupling times and show that the soft decoder improves over the hard baselines by 3 to 8 dB across a range of incident fluxes. We also demonstrate that pixel quantization can lead to further improvements of the soft decoder.

Together, these contributions constitute the first computational methods for multiplexed readout processing with single-photon detector arrays. Our methods can enable innovations in areas like hardware-software co-design of sensor arrays, quantum state estimation, and ghost imaging.