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

Technical Summary: Lens Descriptor-Guided Evolutionary Algorithm for Optimization of Complex Optical Systems with Glass Choice

Problem Statement
Designing high-performance optical lenses involves navigating a high-dimensional, tightly constrained, and highly multimodal search space comprising surface curvatures, glass material choices, element thicknesses, and spacings. Standard optimization workflows, typically relying on gradient-based local search or single-run evolutionary strategies, often converge to a single local optimum. This approach overlooks the existence of numerous comparably good alternative designs, which are critical for downstream engineering decisions regarding manufacturability, cost, glass availability, and tolerance characteristics. While existing multimodal optimization methods exist, they often fail to systematically optimize glass selection or require computational budgets (days to weeks) that exceed industrial practicality (often targeting hour-scale solutions).

Methodology: LDG-EA
The authors propose the Lens Descriptor-Guided Evolutionary Algorithm (LDG-EA), a two-stage framework designed to discover diverse, high-quality local minima within practical computational budgets.

1. Formalization via Behavior Descriptors:
   The core innovation is the definition of a "behavior descriptor" $D(\theta)$, which maps a full lens design $\theta$ to a qualitative representation consisting of:

   • The signs of surface curvatures.
   • The discrete indices of the selected glass materials.
   • Note: Thicknesses and distances are excluded from the descriptor to focus on structural patterns, as defined in prior optical comparison methods.
     This partitions the design space into distinct "niches" based on structural patterns rather than continuous parameter values.

2. Two-Stage Optimization Loop:

   • Stage 1: Descriptor Sampling and Subspace Search:
     The algorithm maintains a probability distribution $p(t)(x)$ over the set of behavior descriptors. At each iteration, it samples $\lambda$ descriptors. For each sampled descriptor, an internal optimization algorithm (Hill-Valley Evolutionary Algorithm, or HV-EA) is executed within the subspace constrained by that specific descriptor. The HV-EA uses Covariance Matrix Self-Adaptation Evolution Strategy (CMSA-ES) to find local minima while holding material indices fixed and varying curvatures and thicknesses within sign constraints.
   • Stage 2: Distribution Update:
     Based on the best objective values found within each sampled descriptor, the algorithm selects the top $\mu$ performing descriptors. It then updates the sampling distribution $p(t)(x)$ using a Univariate Marginal Distribution Algorithm (UMDA) rule, increasing the probability mass on high-performing structural patterns. This allows the algorithm to learn which "families" of designs are most promising.

3. Refinement:
   After the evolutionary search concludes, the algorithm optionally applies gradient-based local optimization (e.g., BFGS) to the candidate solutions to refine them further.

Key Contributions

• Formalization of Lens Descriptors: The paper formalizes the use of curvature-sign patterns and material indices as behavior descriptors to structure and navigate the multimodal optical design landscape.
• Development of LDG-EA: The authors integrate descriptor learning, the Hill-Valley Evolutionary Algorithm, and optional gradient-based refinement into a cohesive adaptive search strategy that balances exploration of new structural patterns with exploitation of high-quality regions.
• Validation on Double-Gauss Topology: The method is validated on a six-element Double-Gauss lens system (24 variables: 18 continuous, 6 integer), demonstrating its ability to generate a diverse set of solutions under an hour-scale computational budget.

Experimental Results
The algorithm was tested on a six-element Double-Gauss lens system with a Schott glass catalog of 120 materials.

• Diversity: Over five independent runs, LDG-EA generated an average of 14,741 candidate local minima spanning 636 unique behavior descriptors. This represents an order-of-magnitude increase in distinct high-quality designs compared to a baseline CMA-ES (Covariance Matrix Adaptation Evolution Strategy) with restarts, which found only ~400 solutions.
• Quality: The best LDG-EA solution, after gradient-based refinement, achieved an objective value (RMS spot size squared) of $3 \times 10^{-4}$. This is competitive with the CMA-ES baseline's best of $1 \times 10^{-3}$ and approaches the performance of a finely tuned reference design ($7 \times 10^{-5}$), though it does not surpass the reference.
• Efficiency: The LDG-EA runs were completed in approximately one hour (wall-clock time) by leveraging parallel execution across descriptors. The baseline CMA-ES, lacking this parallel structure, required over 50 hours for the same evaluation budget.
• Learning Effect: Ablation studies replacing the learned descriptor distribution with uniform random sampling showed that LDG-EA's adaptive learning mechanism rapidly improves solution quality in the first few generations by focusing search effort on high-yield descriptors.

Significance and Claims
The paper claims that LDG-EA successfully addresses the limitation of current industrial tools (like CODE V's Global Synthesis) which struggle to systematically generate a varied portfolio of high-quality designs, particularly when glass choice is involved.

• Practical Utility: The authors emphasize that while the absolute best solution found may not beat a manually tuned reference, the method's primary value lies in producing a diverse set of solutions within a practical time frame. This diversity supports downstream engineering choices that cannot be easily encoded into the optimization objective (e.g., glass availability, cost, tolerance).
• Scalability: The approach is designed for modern computing infrastructures, utilizing parallelism to achieve hour-scale runtimes for complex systems.
• Heuristic Initialization: The authors position LDG-EA as a fast, configurable method to generate high-quality initial designs, where RMS spot size serves as a heuristic starting point for subsequent, more rigorous optimization steps.

The paper concludes modestly, noting that the method does not yet overtake the best reference designs but provides a robust framework for exploring the design space. Future work is proposed to include adaptive learning rates, multi-objective optimization, and the integration of patent-based priors.