Microwave Kinetic Inductance Detectors (MKIDs) are a cryogenically cooled, superconducting detector technology with applications in astronomical observations in the radio through gamma-ray wavelengths. In the optical and IR, MKIDs have zero read noise, single photon sensitivity, microsecond time resolution, and broadband energy resolution. These properties make MKIDs ideal for directly imaging exoplanets, as high sensitivity/low noise are required for observing faint companions, and energy and time resolution are useful for resolving atmospheric aberrations which are the dominant noise source for such observations. Two MKID cameras have been built for this purpose: the 20,000 pixel MKID Exoplanet Camera (MEC) at Subaru Observatory and 10,000 pixel DARK-speckle Near-infrared Energy-resolving Superconducting Spectrophotometer (DARKNESS) camera at Palomar Observatory.
The first chapter of this thesis will discuss the development and testing of a second generation digital readout system for large format optical/IR MKID arrays. Our system retains much of the core signal processing architecture from the first generation system, but with a significantly higher bandwidth, enabling readout of kilopixel MKID arrays. Each set of readout boards is capable of reading out 1024 MKID pixels multiplexed over 2 GHz of bandwidth; two such units can be placed in parallel to read out a full 2048 pixel microwave feedline over a 4 -- 8 GHz band. As in the first generation readout, our system is capable of identifying, analyzing, and recording photon detection events in real time with a time resolution of order a few microseconds. We describe the hardware and firmware, and present an analysis of the noise properties of the system. We also present a novel algorithm for efficiently suppressing IQ mixer sidebands to below -30 dBc.
The second chapter will focus on the development of a machine learning pipeline for automating the calibration of large MKID arrays. This process involves determining the resonant frequency and optimal drive power of every pixel (i.e. resonator) in the array, which is typically done manually. We instead use a deep-learning based object detection scheme to localize correctly-driven resonators in the 2D (power x frequency) frequency response data. Our method has performance equal to that of manual tuning, and only takes 12 minutes of computational time per 2000 pixels, as opposed to 4-6 hours for the manual method.
The final chapter will focus on the development of wavefront sensing and control algorithms to enable more sensitive exoplanet imaging. We implement an additional feedback loop between the MKID focal plane camera and adaptive optics (AO) system deformable mirror (DM) to correct residual optical aberrations in real time. Using a simple Fourier mode based technique (speckle nulling), we demonstrate in-lab convergence times $<1$ second at a control rate of 30 Hz. We also show on-sky suppression of quasistatic speckles by $25\%$ with a convergence time of 10 seconds. Preliminary work on a system identification scheme for improved calibration and control will also be discussed.
