This research addresses optical system optimization challenges where local optimization algorithms with random starting points frequently converge to suboptimal local solutions, while heuristic global optimization algorithms demonstrate low efficiency in solution space exploration. To effectively utilize parallel computing capabilities and minimize computational resources spent on invalid solutions, this study implements a monitoring system that analyzes configuration differences between systems based on selective guidance of global optimal information. When global information guides systems from poor local minima toward improved regions, system aggregation inevitably occurs. This aggregation results in uneven distribution within the solution space, potentially overlooking valid configurations that might yield superior results.

A diversity-enhanced selective guided gradient descent (DESG-GD) algorithm is proposed. Building upon the global information selective guided (GISG) framework, DESG incorporates a solution space density assessment indicator. This indicator quantifies distribution density through calculating the minimum Euclidean distance between system parameters and implements a dual-guidance point strategy: directing low-performing systems toward the loss-optimal point, while guiding systems with acceptable performance but high distribution density toward the diversity-optimal point. By separating systems with similar configurations, the algorithm explores more configurations within the solution space, enhancing system performance and generating effective, diverse results.

Validation through 4-lens and double Gaussian lenses demonstrates that initial structures optimized by DESG-GD achieve or exceed patent-level performance. In complex double Gaussian systems, the proposed method demonstrates more than 10% improvement in spot diagram size, efficiency, and diversity compared to existing advanced methods. DESG-GD substantially enhances the exploration of diverse optical system results while maintaining solution quality, offering novel approaches for reference-free optical design.

This study demonstrates how DESG-GD selectively guides high-performing but densely distributed solutions toward sparser regions of the solution space, facilitating exploration of previously unexplored areas. This approach enhances both diversity and performance in ab initio lens design. The root mean square (RMS) spot sizes for both the 4-lens system and the double Gaussian match or exceed patent performance, demonstrating its effectiveness in obtaining high-quality starting points for optical system optimization without references. The proposed DESG-GD, which enhances local optimization through gradient descent with system interaction strategies, presents significant opportunities for further development and research.