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

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 around 14500 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.

💡 Research Summary

The paper introduces the Lens‑Descriptor‑Guided Evolutionary Algorithm (LDG‑EA), a two‑stage multimodal optimization framework specifically designed for complex optical lens design. Traditional approaches—gradient‑based local refinement or standard evolutionary strategies such as CMA‑ES—tend to converge to a single basin of attraction, thereby missing the multitude of high‑performing designs that differ in curvature sign patterns and glass material choices. LDG‑EA addresses this limitation by first mapping every candidate design θ ∈ Θ to a discrete behavior descriptor x = D(θ). The descriptor consists of the sign of each surface curvature and the integer index of the glass material for each element; distances and thicknesses are deliberately omitted because they do not define the structural family of a lens. This mapping partitions the enormous continuous‑integer design space into a finite set X of “niches” that correspond to qualitatively different optical families.

At each iteration t, LDG‑EA samples λ descriptors from a probability mass function p⁽ᵗ⁾(x). For each sampled descriptor, a sub‑search is launched using the Hill‑Valley Evolutionary Algorithm (HV‑EA). HV‑EA first performs hill‑valley clustering inside the subspace D⁻¹(x) to identify distinct attraction basins (valleys) that share the same descriptor. Within each valley, a Covariance‑Matrix Self‑Adaptation Evolution Strategy (CMSA‑ES) is run as the single‑mode optimizer. CMSA‑ES adapts its mutation covariance matrix on‑the‑fly, which is crucial for navigating the highly rugged, high‑dimensional landscape of optical design. The result of each HV‑EA run is an archive A(t,i) containing all candidate solutions found under the current descriptor, filtered by a quality window w: only solutions whose merit function value lies within w of the best found in that archive are retained. This window keeps the archive compact while preserving diversity of near‑optimal designs.

After all λ descriptors have been evaluated, the best objective value per descriptor f(t,i) = min_{θ∈A(t,i)} F(θ) is recorded. The top µ descriptors (lowest f) are selected, and the descriptor distribution is updated using a Uniform Marginal Distribution Algorithm (UMDA) rule: p⁽ᵗ⁺¹⁾(x) = (1‑α) p⁽ᵗ⁾(x) + α · (empirical frequency of selected descriptors). This adaptive sampling concentrates future evaluations on promising structural patterns while still allowing exploration of less‑tried regions. Additionally, the algorithm learns separate Bernoulli models for curvature signs and categorical models for material indices from the selected descriptors, further biasing the sampling toward high‑performing families.

The method was evaluated on a realistic six‑element Double‑Gauss lens, comprising 24 variables (18 continuous thickness/spacing parameters and 6 integer glass choices). Under a practical computational budget of roughly one hour on a 64‑core workstation, LDG‑EA generated on average 14,500 candidate local minima spanning 636 unique descriptors. This represents an order‑of‑magnitude increase in both the number of distinct high‑quality designs and the diversity of descriptor families compared with a baseline CMA‑ES run under the same budget. The best LDG‑EA design was slightly worse in RMS spot size than a finely tuned reference lens, but it remained within the same performance envelope, demonstrating that the algorithm does not sacrifice quality for diversity.

Key contributions of the work are: (1) formalization of lens‑specific behavior descriptors that capture the essential discrete structure of optical designs; (2) integration of descriptor‑guided probabilistic sampling with a multimodal search (HV‑EA) and a self‑adapting ES; (3) a quality‑windowed archive mechanism that maintains a bounded spread of merit values per niche; and (4) empirical evidence that the approach can deliver a rich portfolio of viable lens designs within industrially relevant time frames. By explicitly exploiting domain knowledge, LDG‑EA bridges the gap between exhaustive multimodal exploration and the practical constraints of commercial optical design workflows, offering engineers a powerful tool to explore trade‑offs in glass availability, manufacturability, cost, and tolerance sensitivity while retaining competitive optical performance.