Deep learning across optical system workflow: a comprehensive review from design to assembly

Abstract Deep learning is rapidly transforming optical system engineering by introducing data-driven and differentiable modeling techniques into traditionally expert-driven design workflows. This review provides a comprehensive synthesis of recent advances in integrating deep learning across the optical system lifecycle, from initial structure generation to physical assembly. We categorize current methods into three major categories: starting point generation (SPG), End-to-End (E2E) co-optimization, and optical active alignment (OAA). SPG networks aim to explore the design space efficiently by learning mappings from system specifications to viable optical layouts, serving as high-quality initial configurations. E2E methods incorporate the optical imaging process and downstream tasks into differentiable learning pipelines, enabling joint optimization of optical and network parameters. OAA networks bridge simulation and hardware realization by using optical measurements to predict and correct fabrication and alignment deviations during assembly. For each category, we discuss the motivations, modeling strategies, and system-level insights, complemented by representative design examples that illustrate practical performance, as well as the associated challenges and emerging trends. We further analyze cross-domain synergies, highlighting reusable modules. Despite their promise, current approaches face challenges related to generalization, simulation fidelity, and interpretability. Addressing these gaps will require standardized datasets and benchmarks, physics-informed and tolerance-aware training strategies, and learning frameworks that generalize across optical architectures. This review aims to provide both a structured overview and a roadmap for researchers and practitioners working at the intersection of photonics and machine learning.