A3I: an AI-assisted optical alignment framework for spectrographs

Abstract

The optical alignment of modern spectrographs is still largely driven by iterative trial-and-error procedures, which become particularly costly given the long vacuum and cooldown cycles of cryogenic systems. We present A3I (Artificial Intelligence Assisted Alignment and Integration), a deep-learning framework that analyzes detector images to identify misalignments in key optical elements and suggest corrections to improve the overall instrument performance. The goal of A3I is to provide instrument teams with rapid, data-driven feedback during the Assembly, Integration and Testing (AIT) phase, reducing the number of cycles needed to converge towards an optimal alignment. In its first application, A3I is trained on simulated calibration frames generated with the End-to-End (E2E) spectrograph simulator for the NIR arm of the SOXS instrument. The simulator reproduces the full optical path and is used to model multiple perturbations of different optical elements. Three key components (echelle grating, collimator mirror, first camera lens) are systematically perturbed over 15 alignment-related degrees of freedom. On unseen simulated data, A3I reaches ∼79% overall accuracy and nearly 90% when focusing on the most impactful DOFs, demonstrating its potential as a portable, AI-based decision-support tool for alignment in a broad class of spectrographs. Future work will focus on validating A3I in a controlled laboratory setup using a dedicated benchmark spectrograph, where intentional misalignments of individual elements will enable a systematic assessment and refinement of its performance on real data.