Learning Design Skills as Memory Policies for Agentic Photonic Inverse Design

2026-05-28

AI Summary

This work addresses the challenges in inverse design of photonic crystal fibers, which involve multi-objective optimization and computationally expensive electromagnetic simulations, as well as the difficulty of accumulating cross-task design knowledge with existing methods. The authors propose SkillPCF, a novel framework that, for the first time, models design expertise as learnable memory strategies. SkillPCF integrates a physics-guided skill library, a reinforcement learning–driven skill selection mechanism, and a simulation-feedback–driven skill evolution mechanism into a closed-loop intelligent agent. This approach transcends the limitations of conventional single-shot prediction or surrogate modeling by enabling effective knowledge reuse across tasks. Evaluated on a real-world dataset comprising 479 expert trajectories and 553 queries, SkillPCF consistently achieves a superior trade-off between design quality and efficiency across multiple large language model backbones and classical baselines.

Abstract

Photonic crystal fiber (PCF) inverse design remains challenging because candidate geometries must satisfy coupled optical targets under expensive electromagnetic simulation. Existing pipelines improve surrogate prediction or one-shot parameter recommendation, but they do not accumulate reusable design knowledge across iterative trials. We formulate PCF inverse design as a memory-policy learning problem and propose SkillPCF, a closed-loop agent framework that combines a physics-guided memory skill bank, reinforcement-learned skill selection, and simulator-grounded skill evolution. We further construct a real-world dataset with 479 expert interaction traces (2,507 spans) and 553 memory-dependent evaluation queries covering dispersion engineering, loss optimization, and multi-objective design. Experiments across multiple LLM backbones and classical baselines show that SkillPCF achieves stronger design-quality and efficiency trade-offs under practical simulation budgets, demonstrating the effectiveness of our proposed memory-skill learning paradigm for physics-aware PCF inverse design.

Problem

• photonic inverse design
• memory policy
• design knowledge reuse
• multi-objective optimization
• electromagnetic simulation

Innovation

• memory-policy learning
• photonic inverse design
• reinforcement learning
• physics-guided memory
• skill evolution