Neuro-inspired automated lens design

TL;DR

Lens design is hindered by a highly non-convex optimization landscape. OptiNeuro addresses this by a neuro-inspired workflow that starts with a diverse pool of candidate lenses and progressively prunes low-performers while refining remaining designs with gradient-based optimization, formalized as $MF = MF_{img} + MF_{\phi}$. The framework leverages physics-constrained initialization, incremental resource reallocation, and an improved Adam optimizer to navigate complex search spaces, while employing GPU-accelerated, multi-lens evaluation. Demonstrations include automated design of a six-element aspheric lens with substantial RMS-spot gains and distortion control, four nine-element aspherics with results near manual designs, and a glass-plastic hybrid fisheye under unprecedented specs, underscoring its ability to explore novel architectures at scale. While limitations remain (focus on spherical/aspheric lenses and need for downstream analyses), OptiNeuro offers a path toward large public lens databases and AI-assisted, high-throughput lens design workflows.

Abstract

The highly non-convex optimization landscape of modern lens design necessitates extensive human expertise, resulting in inefficiency and constrained design diversity. While automated methods are desirable, existing approaches remain limited to simple tasks or produce complex lenses with suboptimal image quality. Drawing inspiration from the synaptic pruning mechanism in mammalian neural development, this study proposes OptiNeuro--a novel automated lens design framework that first generates diverse initial structures and then progressively eliminates low-performance lenses while refining remaining candidates through gradient-based optimization. By fully automating the design of complex aspheric imaging lenses, OptiNeuro demonstrates quasi-human-level performance, identifying multiple viable candidates with minimal human intervention. This advancement not only enhances the automation level and efficiency of lens design but also facilitates the exploration of previously uncharted lens architectures.

Figures (12)

• Figure 1: Automated lens design inspired by the synaptic pruning mechanism in mammalian neural development.a Developing mammalian brains establish redundant neural connections through synaptic overgrowth during early development. These connections are initially imprecise and become more precise through synaptic pruning, which involves weakening and eliminating some synaptic connections and strengthening others. Ultimately, high-efficiency and low-redundancy neural connections are formed. b Drawing from this biological principle, OptiNeuro generates a group of redundant initial structures given a set of design specifications. During the iterative optimization, low-quality lenses are progressively eliminated. Ultimately, a set of high-quality lens designs is obtained. Here, we employ spot diagrams as a critical metric for evaluating lens quality. A smaller RMS (Root Mean Square) spot radius typically indicates superior imaging performance.
• Figure 2: Automated design of a six-element aspheric lens.a The baseline design and top-$5$ lenses designed by OptiNeuro. For each involved lens, its layout, spot diagrams across five sampled fields (with corresponding RMS spot radii labeled below and zoomed patches of simulated images illustrated on the right), and grid distortion diagram (annotated with maximum distortion value below) are illustrated from top to bottom. b ISO 12233 resolution chart and five zoomed patches corresponding to the five sampled field positions. c The ranking progression of five selected lenses (labeled as FR1, FR3, FR5, IR1 and IR3) throughout entire $7$ steps of lens elimination and refinement. Lenses with a final ranking of $j$ are labeled as FR$j$, and those with an initial ranking of $j$ are labeled as IR$j$. $N^{(i)}_L$ represents the number of remaining lenses in the $i_{th}$ step.
• Figure 3: Design results of nine-element aspheric lenses. Four sets of design specifications are labeled respectively as A1, A2, A3, and A4, and here we list the five primary design specifications: FOV, F-number, EFL, TTL, and distortion. The design results including the top-$4$ final lenses designed by OptiNeuro and reference manual design results, and below each lens is labeled with the corresponding Avg RMS spot radius (unit is ${\mu}m$).
• Figure 4: Design results of the glass-plastic hybrid fisheye lens. We configure six potential Design Forms (DFs): GGPPSPPPP, GGGPSPPPP, GGPGSPPPP, GGPPSGPPP, GG(GG)SPPPP, and GGPPS(GG)PP. Each DF is named after its sequence of Glass elements, Plastic elements, and aperture Stop. (GG) denotes two glass elements cemented together. We illustrate the optimal solution along with the corresponding spot diagrams and MTF curves for each DF.
• Figure 5: Efficient and accurate Merit Function evaluation.a The general process of ray tracing. This process begins by initializing a ray (blue arrow) at the object surface with position $\mathbf{O}_0$ and direction vector $\mathbf{D}_0$. Given $N_{S}$ lens surfaces (excluding the object and image surfaces), the ray propagates through the system, and the process terminates once the intersection $\mathbf{O}_{N_{S}+1}$ on the image surface is determined. b Three-step ray aiming strategy. Firstly, rays are sampled at the center of the aperture stop and then traced backward to the first lens surface. The approximate entrance pupil center for a specific sampled field angle can be determined. Subsequently, rays are uniformly sampled around the entrance pupil centers along meridional and sagittal directions, and two iterative ray tracings are conducted to determine the entrance pupil edge. c Rays are sampled within the entrance pupils based on the Gaussian Quadrature (GQ) algorithm and then traced forward to the image surface to evaluate the image quality function ($MF_{img}$) and the physical constraint function ($MF_{phi}$), and then further evaluate the Merit Function ($MF$). d GPU-accelerated evaluation of Merit Function based on parallel ray tracing. Assuming that there are $N_L$ lenses, where each lens consists of $M$-dimensional lens parameters, and $x^{(j)}_{i}$ represents the $j_{th}$ lens parameter in the $i_{th}$ lens. All lenses form a matrix, which is converted into a ray matrix. Assuming that evaluating the Merit Function of a lens requires tracing $N_R$ rays, a total of $N_L \times N_R$ rays are traced in parallel to obtain the tracing results, and these results are used to compute the MFV (Merit Function Value) vector, in which $y_i$ represents the MFV of the $i_{th}$ lens.