Co-design framework for the joint optimization of astronomical optical systems and deep learning algorithms

We present an optical system optimization framework that directly bridges astronomy, computational imaging, and optical engineering. Traditionally, optical systems of telescopes are optimized using classical image quality metrics (e.g., RMS spot radius), which are often decoupled from the performance of modern, deep learning-based object detection algorithms. This paper introduces an end-to-end, interdisciplinary optimization framework where the optical system is optimized directly for the detection performance of its subsequent deep learning algorithm. Our method integrates an optical system simulator-which models specific optical designs, hardware, and observation modes-with a deep learning-based detection algorithm. In this joint optimization scheme, the detection accuracy of a pre-trained, weight-fixed network on simulated imagery serves as the primary evaluation signal. By combining this signal with classical optical merit functions, we facilitate a closed-loop, iterative refinement of optical design parameters. To validate this co-design approach, we optimized both a Primary-Focus and a Ritchey-Chrétien telescope for a wide-field astronomical survey scenario. The resulting systems demonstrated gains in celestial object detection efficiency and accuracy, proving the value of our method. This work provides an innovative pathway for jointly optimizing optical systems and their bespoke algorithms for specific scientific tasks.