Speaker
Description
Hybrid pixel detectors traditionally separate sensing from advanced signal processing, constraining pixel-level functionality to analog conditioning and threshold-based interpretation [1,2]. Our earlier work [3] introduced the concept of embedding artificial neural networks (ANNs) directly within pixels to enable local inference and mitigate the impact of analog non-idealities such as noise, gain non-linearity, and baseline offsets. Building on this idea, we then demonstrated [4] an in-pixel ANN architecture designed for pulse amplitude reconstruction, improving photon energy estimation from digitized signals.
In this work, we extend this concept by demonstrating post-fabrication functional reconfiguration of the in-pixel neural network. We show that a fabricated pixelated readout ASIC, originally designed and optimized for pulse amplitude (photon energy) estimation, can be re-trained to perform time-of-arrival (ToA) estimation without any modification to the underlying hardware. Each pixel contains a compact ANN with four hidden layers and a total of over 400 trainable weights, which processes the ADC-digitized pixel signal locally. By updating only the network parameters and processor instructions, the same in-pixel network can be repurposed to perform a fundamentally different measurement task.
We present experimental results of this re-training process, demonstrating accurate ToA estimation directly at the pixel level. Modifications to the network parameters, processor instructions, and the associated test and measurement setup [5] are described. These results demonstrate the capability of pixels with embedded neural networks to be flexibly reconfigured and dedicated to various application-specific measurement tasks.