Differentiable Imaging: Progress, Challenges, and Outlook

Differentiable imaging has emerged as a transformative paradigm in computational imaging by enabling end-to-end optimization of integrated optical-computational systems. This review examines how differentiable frameworks bridge the persistent gap between physical reality and computational models, effectively addressing fundamental challenges in uncertainty quantification, system design, and computational complexity. We analyze recent advances demonstrating how this approach simultaneously optimizes physical and computational elements while enhancing reconstruction accuracy beyond traditional limits. The framework’s capacity to systematically address both deterministic uncertainties, such as manufacturing tolerances and alignment errors, and stochastic uncertainties, including sensor noise and environmental fluctuations, has enabled unprecedented robustness in real-world imaging applications. We identify emerging challenges in numerical modeling, computational efficiency, and system integration, with particular focus on how digital twin architectures may provide solutions for next-generation adaptive imaging systems. Looking forward, we discuss critical opportunities in theoretical foundations, system innovation, and scientific applications, highlighting how differentiable imaging is poised to not only revolutionize imaging technology itself but also accelerate scientific discovery through artificial intelligence-enabled, physics-consistent computational methods.