Lens-descriptor guided evolutionary algorithm for optimization of complex optical systems with glass choice

Kirill Antonov 1 Teus Tukker 2 Tiago Botari 2 Thomas H. W. Bäck 1 Anna V. Kononova 1 Niki van Stein 1 1 Leiden University, The Netherlands 2 ASML, Veldhoven, The Netherlands · 2026

Read Lens-descriptor guided evolutionary algorithm for optimization of complex optical systems with glass choice by Kirill Antonov 1 Teus Tukker 2 Tiago Botari 2 Thomas H. W. Bäck 1 Anna V. Kononova 1 Niki van Stein 1 1 Leiden University, The Netherlands 2 ASML, Veldhoven, The Netherlands (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-22075.txt.

Introduction

Kirill Antonov and co-authors search for better camera and instrument lenses with an evolutionary method. Designing a high-performance lens means exploring a high-dimensional space: surface curvatures, glass choices, element thicknesses, and spacings, all tightly constrained. Standard optimizers tend to converge to one local optimum.

That single answer hides many comparably good alternatives a designer might prefer for cost or manufacturability. The team's 2026 preprint guides an evolutionary algorithm with lens descriptors, summaries that keep the search diverse across design families. Better automated lens design feeds cameras, microscopes, and medical optics, where each design is a compromise among sharpness, size, weight, and price.

Catalog record

CREATORKirill Antonov 1 Teus Tukker 2 Tiago Botari 2 Thomas H. W. Bäck 1 Anna V. Kononova 1 Niki van Stein 1 1 Leiden University, The Netherlands 2 ASML, Veldhoven, The Netherlands
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-22075

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Lens-descriptor guided evolutionary algorithm for optimization of complex optical systems with glass choice Kirill Antonov 1 Teus Tukker 2 Tiago Botari 2 Thomas H. W. Bäck 1 Anna V. Kononova 1 Niki van Stein 1 1 Leiden University, The Netherlands 2 ASML, Veldhoven, The Netherlands Abstract Designing high-performance optical lenses entails exploring a high-dimensional, tightly constrained space of surface curvatures, glass choices, element thicknesses, and spacings. In practice, standard optimizers (e.g., gradient-based local search and evolutionary strategies) often converge to a single local optimum, overlooking many comparably good alternatives that matter for downstream engineering decisions. We propose the Lens Descriptor-Guided Evolutionary Algorithm (LDG-EA), a two-stage framework for multimodal lens optimization. LDG-EA first partitions the design space into behavior descriptors defined by curvature-sign patterns and material indices, then learns a probabilistic model over descriptors to allocate evaluations toward promising regions. Within each descriptor, LDG-EA applies the Hill-Valley Evolutionary Algorithm with covariance-matrix self-adaptation to recover multiple distinct local minima, optionally followed by gradient-based refinement. On a 24-variable (18 continuous and 6 integer), six-element Double-Gauss topology, LDG-EA generates on average 14.5 × 10 3 14.5\times 10^{3} candidate minima spanning 636 unique descriptors, an order of magnitude more than a CMA-ES baseline, while keeping wall-clock time at one hour scale. Although the best LDG-EA design is slightly worse than a fine-tuned reference lens, it remains in the same performance range. Overall, the proposed LDG-EA produces a diverse set of solutions while maintaining competitive quality within practical computational budgets and wall-clock time. K eywords Optical Design, Evolutionary Algorithms, Multi-modal optimization, Multi-modal Search Landscapes, Quality-Diversity 1 Introduction Optical lens systems enable a broad spectrum of advanced technologies, from smartphone and machine‐vision cameras to the ultra‐precise optics used in lithography and metrology equipment [ 4 ] . Designing these and similar systems involves selecting optical surface curvatures, glass types, element thicknesses, and inter‐element spacings to meet key performance metrics, such as root mean square (RMS) spot size, focal length, and field number while respecting manufacturing constraints on element size, weight, and material availability [ 34 , 23 , 46 ] . The resulting design space is both high‐dimensional and highly multimodal [ 13 ] , with numerous local optima in the objective function [ 44 ] . Figure 1 illustrates this phenomenon on a two-dimensional slice through a real lens-design landscape: multiple separated peaks represent different high-performing families of designs rather than minor perturbations of a single solution. Figure 1: Two views of the optimization landscape for a four‐element camera lens, where variables are curvatures, distances, and materials of glasses, visualized using the method from work [ 2 ] . This method selects points between three local minima, identified through a local-search algorithm. The horizontal axes correspond to a plane in the high-dimensional design space, and the vertical axis shows the reversed merit function (the higher, the better). Multiple optima indicate distinct families of high-quality designs. In industrial practice, optical engineers typically work in a designer-in-the-loop manner: they propose a starting layout (e.g., from patents or their own expertise) and then apply local numerical optimization using commercial software (e.g., OpticStudio [ 48 ] or CODE V [ 41 ] ). This workflow is highly effective for refinement, but it is biased toward only several attraction basins and therefore explores only a tiny fraction of the design space. CODE V’s Global Synthesis (GS) module [ 9 ] is a widely used alternative intended to broaden exploration. However, GS provides limited control over the diversity of the returned solutions and offers only partial support for systematic glass optimization, making it difficult to deliberately generate a varied portfolio of high-quality designs. The ability to produce multiple, diverse high-quality designs is practically valuable. A diverse solution set supports downstream engineering choices that cannot be reliably incorporated into the optimization itself, including validation of previously unknown structural constraints, consideration of glass availability, and selection of designs with advantageous cost, manufacturability, and tolerance characteristics. Motivated by these needs, a broad range of learning-based methods [ 47 ] and metaheuristics [ 43 , 18 ] have been proposed. Neural networks have been used to extrapolate reference designs taken from the literature [ 10 , 11 , 12 ] , and surrogate models combined with Differential Evolution [ 14 ] can reduce expensive ray-trace evaluations (e.g., by around 10%) [ 17 ] . Deep models have also been explored for tolerancing and uncertainty quantification [ 25 , 38 ] and for modeling specialized aberration behavior [ 49 ] . From the evolutionary computation perspective, classical global optimization methods have been successfully applied to lens design [ 32 , 30 , 36 ] . Despite these advances, these methods are still limited in glass selection and exploration since they do not exploit the specific structure of lens-design landscapes and therefore do not aim to systematically cover diverse high-quality designs. Multimodal optimization methods offer a natural framework for addressing this problem, as they are explicitly designed to identify multiple optima within a single run [ 35 ] . Recent studies have indeed investigated multimodal optimization for optical design [ 21 , 22 , 1 , 3 ] . However, the existing applications of multimodal methods optimize only surface curvatures, which limits practical relevance because glass choice is a crucial element of real designs. Moreover, reported wall-clock times span multiple days to weeks even on moderate lens systems (three to four elements), whereas industrial tools such as GS can often deliver solutions within about an hour on such systems on the same machine [ 3 ] . In this paper, we close this gap by introducing a multimodal evolutionary algorithm that optimizes glasses, curvatures, element thicknesses, and inter-element spacings for a practical six-element Double-Gauss lens system under an hour-scale budget. We propose the Lens Descriptor-Guided Evolutionary Algorithm (LDG-EA), a two-stage framework that explicitly exploits domain structure by partitioning designs into interpretable behavior descriptors and learning an adaptive sampling distribution over them. At each iteration, LDG-EA samples promising descriptors, applies the Hill-Valley Evolutionary Algorithm [ 27 ] to locate local optima within the subspace induced by each descriptor, and updates descriptor probabilities based on the observed performance, balancing exploration of new regions with exploitation of high-quality design regions. Several leading multimodal optimizers also allocate search effort across landscape regions, for example BIPOP-CMA-ES [ 16 ] , DPI-CMA-ES [ 39 ] , DPI-MIES [ 24 ] , LADE [ 26 ] and NetCDE [ 8 ] . In contrast to these largely domain-agnostic partitions, LDG-EA defines lens-specific niches a priori through behavior descriptors that capture curvature and glass patterns. This descriptor-driven partition explicitly enforces diversity across design patterns. As a result, LDG-EA balances between solution quality and diversity to enable robust multimodal search tailored to lens design. Contributions. • We formalize lens-design behavior descriptors and use them to structure and navigate the lens-design multimodal optimization landscape. • We develop LDG‐EA, which integrates descriptor learning, Hill-Valley Evolutionary Algorithm, and optional gradient‐based refinement into an adaptive search strategy. • We validate LDG‐EA on a six‐element Double‐Gauss lens, demonstrating an order‐of‐magnitude increase in distinct high‐quality designs over a baseline algorithm and competitive RMS performance under equal computational budgets. 2 Formalization of the optimization problem Let \Theta_{0}\subseteq\mathds{R}^{n_{c}}\times\mathds{R}^{n_{d}}\times\mathds{Z}^{n_{m}} be the full design space for an

Read the complete text