Automating Sensor Characterization with Bayesian Optimization

TL;DR

We address the bottleneck of detector characterization in sensor development by introducing a closed-loop Bayesian optimization workflow to automate parameter-tuning during testing. The method integrates a cryogenic test stand, a carefully defined objective function combining SNR and penalties, and a Gaussian-process-based BO loop to select operating states in real time. The key contributions include the detailed objective formulation, demonstration on the SiSeRO CCD showing orders-of-magnitude reduction in tuning time (to a couple of days), and validation on other CCD architectures with plans to generalize to sensor arrays. The practical impact is a scalable framework that accelerates detector characterization and reduces reliance on expert-guided manual scans, enabling faster deployment of next-generation sensors.

Abstract

The development of novel instrumentation requires an iterative cycle with three stages: design, prototyping, and testing. Recent advancements in simulation and nanofabrication techniques have significantly accelerated the design and prototyping phases. Nonetheless, detector characterization continues to be a major bottleneck in device development. During the testing phase, a significant time investment is required to characterize the device in different operating conditions and find optimal operating parameters. The total effort spent on characterization and parameter optimization can occupy a year or more of an expert's time. In this work, we present a novel technique for automated sensor characterization that aims to accelerate the testing stage of the development cycle. This technique leverages closed-loop Bayesian optimization (BO), using real-time measurements to guide parameter selection and identify optimal operating states. We demonstrate the method with a novel low-noise CCD, showing that the machine learning-driven tool can efficiently characterize and optimize operation of the sensor in a couple of days without supervision of a device expert.

Figures (6)

• Figure 1: Schematic top and cross-sectional view of the SiSeRO-CCD output amplifier, consisting of an n-channel MOSFET (green shaded region) integrated into the CCD's p-channel (red shaded region). The p-channel acts as an internal gate (IG), coupling the charge (red dots) to the MOSFET. During readout, charge is shifted from the CCD into the MOSFET internal gate by manipulating the voltages and timing of the clock phases of the Summing Gate (SG), Output Gate (OG) and Amplifier Gate (AG). The junction coupling between the charge packet and the transistor modulates the MOSFET's drain current, providing a high-sensitivity measurement of the pixel charge. Non-destructive readout (NDR) is performed by shifting the charge between the SG and the SiSeRO (AG), further reducing the noise on the charge measurement. The isolation guard (blue shaded region) isolates the n-type MOSFET from the n-type CCD substrate, preventing parasitic flow of electrons into the MOSFET channel.
• Figure 2: The experimental test bench setup, including the major components and the light path. Light is emitted from the source (yellow) and travels first to the shutter (pink), which controls the timing of illumination. Light then travels to the integrating sphere which provides uniform exposure of the sensor, which resides in the vacuum chamber (cyan). The readout electronics (red) supply the bias voltages and clock timings to the sensor through a DB-50 port. The shutter (pink) controls illumination from the light source (yellow). The data acquisition (DAQ) module, the black cover that is used to minimize light leaks into the system, and the cryogenics are not shown in the figure.
• Figure 3: Left: An example image taken with a SiSeRO-CCD. The shading distinguishes regions of the image: prescan, active, overscan, and reverse overscan regions. Each region is used to diagnose an aspect of the detector's performance. Right: Histograms of the pixel-charge distribution corresponding to the active (black) and reverse overscan (blue) regions highlighted by open rectangles in the left panel. The active pixel levels are distinct from those in the reverse overscan region, and this separation defines the signal $S$. The width of the reverse overscan $\sigma_{rev}$ represents the noise in the readout system.
• Figure 4: Illustration of BO showing the true objective function (dashed line) along with the evolution of the Gaussian process surrogate model's posterior mean ($\mu$; black line) and uncertainty ($\sigma$; shaded region). The surrogate model is updated over a series of $n$ observations (filled circles) where the red circle represents the $n$th observation. The acquisition function (green) scores each candidate location in the parameter space, and the point at which it peaks (red triangle) is selected for the next observation. This illustration is inspired by Fig. 1 of Shahriari_2016.
• Figure 5: Corner plot of the objective function across the parameters used during SiSeRO-CCD optimization. Diagonal panels: one-dimensional partial dependence of the objective with respect to each parameter (others marginalized). Off-diagonal panels: pairwise partial dependence of the objective with respect to each parameter pair. The black points mark the configurations explored during burn-in and BO; dark purple indicates higher objective values, and bright yellow indicates lower values (lower objective values are better under the defined metric). The red vertical markers on the diagonal panels denote the coordinate-wise minimizers (argmin) of the partial dependence curves $PD_j(\theta_j)$ over the explored range. To make some weaker trends visible, the vertical scales of the CDS and $V_{I}$ marginals were expanded.