Machine Learning-Based Denoising for Single-Photon Detection With MKIDs
Viktor Makhrinov, Michael Arena, Aled Cuda, Mehdi Bagheri, Akhan Almagambetov, Aziret-Sultan Kiyazov, R. D. Martin, B. Grossan, George F. Smoot, Mehdi Shafiee
Microwave Kinetic Inductance Detectors (MKIDs) are superconducting sensors primarily used in astrophysics and quantum measurements for their energy- and time-resolution of single-photon detection. However, their sensitivity and precision are limited by various noise sources. Conventional denoising approaches enhance the signal-to-noise ratio (SNR) but often blur signal features and are ineffective with non-stationary noise. This work explores the use of machine learning (ML) denoising techniques for MKID signals in steady state resonators and in photon-induced transients. Different denoising techniques are evaluated on a dataset of frequency-domain S21 measurements and time-domain phase pulses with illumination of 660 nm, 780 nm, and 1310 nm wavelengths. Traditional filtering approaches (moving average, median, Wiener, Savitzky-Golay, wavelet, Kalman, total variation denoising) were evaluated in comparison with ML models including convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and gated recurrent units (GRUs). Due to absence of clean reference signals in experimental settings, a Noise2Noise training approach was used for self-supervised learning from noisy data. The models were assessed using mean squared error (MSE), signal-to-noise ratio (SNR), variation in peak depth and optical resolution metrics. Findings of this thesis show that machine learning models achieve substantially better denoising performance compared to conventional methods across all metrics. GRU models outperformed other models when denoising steady-state resonator signals, as they captured local features of the spectral response, and they also were the most effective at denoising transient photon signals due to their capacity to capture temporal dynamics. Conventional filters, while being computationally efficient, were found to distort the signal and were not as effective in retaining low-amplitude photon-induced features. In summary, this study demonstrates that machine learning denoising is a robust and versatile method for improving MKID signals without prior knowledge of noise properties. The results also demonstrate the opportunity to enhance superconducting detector readouts by combining data-driven approaches, leading to improved sensitivity and resolution for applications such as astrophysical imaging and quantum sensing.
Microwave Kinetic Inductance Detectors (MKIDs) are superconducting sensors primarily used in astrophysics and quantum measurements for their energy- and time-resolution of single-photon detection. However, their sensitivity and precision are limited by various noise sources. Conventional denoising approaches enhance the signal-to-noise ratio (SNR) but often blur signal features and are ineffective with non-stationary noise. This work explores the use of machine learning (ML) denoising techniques for MKID signals in steady state resonators and in photon-induced transients. Different denoising techniques are evaluated on a dataset of frequency-domain S21 measurements and time-domain phase pulses with illumination of 660 nm, 780 nm, and 1310 nm wavelengths. Traditional filtering approaches (moving average, median, Wiener, Savitzky-Golay, wavelet, Kalman, total variation denoising) were evaluated in comparison with ML models including convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and gated recurrent units (GRUs). Due to absence of clean reference signals in experimental settings, a Noise2Noise training approach was used for self-supervised learning from noisy data. The models were assessed using mean squared error (MSE), signal-to-noise ratio (SNR), variation in peak depth and optical resolution metrics. Findings of this thesis show that machine learning models achieve substantially better denoising performance compared to conventional methods across all metrics. GRU models outperformed other models when denoising steady-state resonator signals, as they captured local features of the spectral response, and they also were the most effective at denoising transient photon signals due to their capacity to capture temporal dynamics. Conventional filters, while being computationally efficient, were found to distort the signal and were not as effective in retaining low-amplitude photon-induced features. In summary, this study demonstrates that machine learning denoising is a robust and versatile method for improving MKID signals without prior knowledge of noise properties. The results also demonstrate the opportunity to enhance superconducting detector readouts by combining data-driven approaches, leading to improved sensitivity and resolution for applications such as astrophysical imaging and quantum sensing.
Superconducting micro-resonators are crucial in creating superposition states in quan tum computers and are gaining prominence as quantum technology evolves. Beyond their quantum applications, these resonators are instrumental as ultra-sensitive de tectors in astronomy, particularly Microwave Kinetic Inductance Detectors (MKIDs), which are key for both scientific exploration and industrial use. However, the is sue of electronic noise in these detectors remains a significant hurdle. To address this, our approach integrates machine learning strategies alongside conventional noise reduction methods to improve MKID signal fidelity. Our process involves gathering data and applying denoising models, with a strong emphasis on machine learning techniques. Our findings highlight the strengths of dif ferent denoising techniques, particularly deep learning architectures like Long Short Term Memory (LSTM) networks and Autoencoders, which demonstrate promising results in denoising MKID signals. These models exhibit adaptability to diverse noise sources, proficiency in identifying complex noise patterns, and continuous improve ment with additional data. This research represents a significant advancement in the field and provides es sential insights into improving MKID detectors, which in turn could lead to ground breaking developments in various scientific areas.
Superconducting microresonators are pivotal in the realm of quantum computing, essential for establishing superpositions. As quantum technology evolves, these components become even more integral, not just in quantum applications but also as precision instruments in astronomy—specifically as highly sensitive cameras. Microwave kinetic inductance detectors are notable in this field, recognized for their heightened sensitivity crucial for both scientific exploration and industrial use. Yet, electronic noise remains a formidable obstacle. Addressing this, our approach combines machine learning strategies with traditional noise reduction techniques to enhance the signal clarity of superconducting microresonators. Our process involves gathering data and applying various denoising models, placing a significant emphasis on machine learning methods. We have discovered that certain denoising approaches, especially deep learning frameworks such as long short-term memory networks and autoencoders, are highly effective in clarifying the signals from superconducting microresonators. These models demonstrate versatility in handling different noise types, skill in deciphering intricate noise structures, and they consistently improve their performance as more data is provided.
Noise is one of the main factors affecting the performance index of weak signal detection devices, and the optimal filtering algorithm is an effective method to adaptively extract various useful weak signals from the white noise background. In order to improve the performance of single photon detector (especially the photon number resolution ability), one mainly focuses on the optimization of detector hardware such as the optimization of photosensitive materials and the technology of device fabrication. However, in this paper the performance of microwave kinetic Inductance detector (MKID) in the way of data processing is improved. Considering the fact that the template of light pulse signal in the optimal filtering algorithm is obtained by taking the average, we replace the noise model in the original optimal filtering algorithm with the white noise model and the whitening noise model. Then we process the photon response data that are detected by the MKID in an extremely low temperature environment. The results show that the energy resolution (one of the main performance indexes of single photon detector) of MKID is improved by about 15%, and we achieve an infrared single photon energy resolution of 0.26 eV. In this paper, the application and development trends of superconducting single photon detector are briefed. Then, how the MKID responds to weak coherent optical signal in low temperature environment, and the process of signal conversion, acquisition and output are explained in detail. According to the optimal filtering algorithm, we use different noise models to analyze the results of the signals detected by MKID. After that, we count the optimal amplitude multiple, perform the Gaussian fitting analysis on the statistical graph, and compare the energy resolution with the photon number resolution of the optimal filtering algorithm under different noise models. As a result, we find that under the white noise model, the optimal filtering algorithm is used to obtain the best result for MKID processing, and high energy resolution can be achieved.
Microwave Kinetic Inductance Detectors (MKIDs) are remarkable photon detectors, that have single photon detection and energy resolving capabilities in the near-infra red and higher frequency range. At lower frequencies, MKIDs are excellent radiation detectors as well, because of their high sensitivity and natural multiplexing capabilities, which enable large scale detector arrays.In the (near) optical regime, MKIDs can use their single photon energy resolving capabilities for direct exoplanet detection, in missions like Hab-Ex or LUVOIR. This would enable atmospheric characterization, potentially finding habitual exoplanets. However, two aspects of the MKID still need improvement for this application: the photon absorption efficiency must go from 30% to 50% and the resolving power must go from 8 to 100, both for wavelengths of 1 µm. In this thesis, we study the single photon response and generation-recombination (GR) noise in optical NbTiN-Al hybrid MKIDs, to find knowledge gaps and opportunities to improve detector performance. For the single photon response, we set up a model starting from the Rothwarf-Taylor and Mattis-Bardeen equations, including the pair-breaking efficiency as only fit parameter. Both at high (220 mK and 250 mK) and low (120 mK) temperatures, the model predicted the single photon response of 4 different wavelengths correctly, with a fitted pair-breaking efficiency close to expected values (30%-70%), but somewhat high, which might be due to a non-thermal distribution of quasiparticles caused by read power. In particular, the effect of phonon trapping was captured by the model, which was verified by considering MKIDs on substrate and membrane.At lower temperatures (120 mK), a second exponential decay in both the amplitude and phase single photon pulse tail was observed, which is not explained by the model. This second decay was faster for higher read powers. For the GR noise, also an unexpected feature was observed: the GR noise level dropped exponentially, when lowering temperature (<250 mK), for the amplitude, phase and cross power spectral densities (PSDs). High read powers mask this effect, due to the creation of excess quasiparticles. The noise drop was present in all analysed MKIDs, except for an MKID with an 150 nm (instead of 50 nm) thick Al film, which might be due to read power induced excess quasiparticles. When assuming the GR noise to be Poissonian, we show that the behaviour could be explained by a process which limits the quasiparticle lifetime, while keeping the quasiparticle density thermal.We hypothesize the cause of both of these unpredicted measurements to be quasiparticle trapping, which is the localisation of quasiparticles. This process is known to degrade superconducting tunnel junction detectors and limit quasiparticle lifetimes in MKIDs. A secondary ion-mass spectroscopy (SIMS) analysis showed Fe contamination on the substrate-Al film interface, which is thought to be the quasiparticle trapping cause in our systems, consistent with the observation that increasing the film thickness diminishes the GR noise drop. To our knowledge, quasiparticle trapping has not be studied in steady state experiments, such as GR noise measurements.Different models including quasiparticle trapping have been set up and can predict the amplitude PSDs, when assuming that trapped quasiparticles cannot dissipate microwave power. However, comparison of the fitted model parameters with the trapping and detrapping rates calculated by Kozorezov et al., showed a orders of magnitude difference. Combining this with the fact that amplitude, phase and cross PSDs show similar behaviour, implying that they relate to Cooper-pair fluctuations, led us to conclude that the trapping process cannot be the cause of the observed behaviour. On-trap recombination most likely plays a role in the GR noise drop, but models including this process involve cyclic transitions and non-equilibrium steady states, greatly complicating the calculations. The second exponential decay could not be described by the trapping models, but also here on-trap recombination and Cooper-pair fluctuations must be considered.A qualitative analysis on the MKID detector performance (both single photon and radiation power integration) showed that quasiparticle trapping effects can increase the energy resolving power and noise equivalent power at low temperatures, when this effect replaces the usual lifetime saturation due to excess quasiparticles.
We report a self-supervised deep learning framework for Rydberg sensors that enables single-shot noise suppression matching the accuracy of multi-measurement averaging. The framework eliminates the need for clean reference signals (hardly required in quantum sensing) by training on two sets of noisy signals with identical statistical distributions. When evaluated on Rydberg sensing datasets, the framework outperforms wavelet transform and Kalman filtering, achieving a denoising effect equivalent to 10,000-set averaging while reducing computation time by three orders of magnitude. We further validate performance across diverse noise profiles and quantify the complexity-performance trade-off of U-Net and Transformer architectures, providing actionable guidance for optimizing deep learning-based denoising in Rydberg sensor systems.
Microwave Kinetic Inductance Detectors (MKIDs) are superconducting photon counting detectors that simultaneously measure the energy and arrival time of individual photons and are suitable for large-format arrays. The energy resolving power R = E/ΔE is the primary figure of merit for spectrophotometric applications. It is limited in practice by the assumptions underlying the Wiener optimal filter used to estimate photon energy: stationary noise, linear response, and energy independent pulse shapes, all of which MKIDs structurally violate. We present Venom (Very Efficient Neural Optimal-filter for MKIDs), a selective state space model based on the Mamba architecture that replaces the coordinate transform and optimal filter, operating directly on raw in-phase/quadrature timestream data, carrying 3,314 trainable parameters, and targeting deployment on the MKIDGen3 and future system-on-chip readout platforms. To deal with the small calibration datasets typical of MKID experiments, we train on synthetic pulses drawn from a streaming principal component analysis that samples photon energy continuously between calibration wavelengths. All numbers are reported on a stratified 15% held-out validation set, with R measured from a kernel density estimate. On an InHf bilayer resonator (ten wavelengths, 254-1310nm), Venom reaches a mean R of 26.4 versus 24.2 for the published per-wavelength optimal filter baseline, a 9% improvement from one model in place of ten per-laser filter templates. The R of 38.6 at 254 nm is the highest ever recorded for an ultraviolet-to-near-infrared MKID suitable for use in a dense array. On a PtSi resonator (five wavelengths, 808-1310nm), where the published analysis used one shared 920nm template, Venom matches the optimal filter, with a mean R of 9.1 versus 8.8 at the three interior wavelengths.