Tolerance-Aware Deep Optics

TL;DR

The paper tackles the design-to-manufacturing gap in deep optics caused by unavoidable tolerances in fabrication and assembly. It introduces an end-to-end tolerance-aware optimization framework that embeds tolerance modeling into differentiable ray tracing and jointly optimizes optical elements with the downstream decoder, using a two-stage training regime. Two novel losses, Spot Loss and PSF Similarity Loss, stabilize training under random tolerance patterns, enabling robust performance under manufacturing variability and yielding substantial improvements in simulated deblurring (over $2\text{dB}$) and real-world reconstructions. This approach advances the practicality of deep optical systems for mass production by reducing sensitivity to tolerances and improving downstream imaging robustness.

Abstract

Deep optics has emerged as a promising approach by co-designing optical elements with deep learning algorithms. However, current research typically overlooks the analysis and optimization of manufacturing and assembly tolerances. This oversight creates a significant performance gap between designed and fabricated optical systems. To address this challenge, we present the first end-to-end tolerance-aware optimization framework that incorporates multiple tolerance types into the deep optics design pipeline. Our method combines physics-informed modelling with data-driven training to enhance optical design by accounting for and compensating for structural deviations in manufacturing and assembly. We validate our approach through computational imaging applications, demonstrating results in both simulations and real-world experiments. We further examine how our proposed solution improves the robustness of optical systems and vision algorithms against tolerances through qualitative and quantitative analyses. Code and additional visual results are available at openimaginglab.github.io/LensTolerance.

Figures (14)

• Figure 1: When subjected to random tolerances perturbations, the performance of conventional deep optics degraded severely, whereas deep optics with our tolerance-aware optimization maintained excellent computational imaging performance. Camera Captured and Reconstructed, representing the simulated camera imaging result and the reconstructed result after decoder, respectively.
• Figure 2: The tilt, decentration and central thickness tolerances can be modeled as Rotation and Translation of Lens. The spatial transformations of Lens can be equivalent to the opposite spatial transformations of ray's coordinate. The PSFs of the lens system change drastically due to tolerances. Scale bar: 15$\mu m$.
• Figure 3: Tolerance-aware optimization for deep optics. We integrate tolerances into differentiable ray tracing. Every kind of tolerances are randomly sampled from its distribution $\mathcal{N}($0$, \tilde{\sigma_{i}^2})$, use ray tracing with tolerances to render perturbed spatially-variant PSF maps and simulated the imaging results by spatially-variant convolution, then noise is added to simulate sensor-captured images. These images are then passed to a computational decoder for reconstruction. During forward simulation, we track gradients of optical parameters, We can subsequently back-propagate the errors from either the reconstruction images quality and tolerance loss terms. The framework jointly optimizes the optics and the computational decoder in a tolerance-aware manner.
• Figure 4: Spot diagrams and RMS spot sizes produced by our framework with and without perturbations are highly resemble those by Zemax, the tested lens is Cooke Triplet and for visualization, the scale bars are different.
• Figure 5: Deblurring results comparison of deep optics with and without tolerance-aware optimization under tolerances perturbations. The deep optics with tolerances optimization maintain better deblurring performances than its counterpart.