Scholarlink

<p>This paper presents a reinforcement learning (RL)-based inverse design framework for photonic device optimization, where a Deep Q-Network (DQN-RL) agent is directly coupled with three-dimensional finite-difference time-domain (3D-FDTD) simulation environment. The proposed approach eliminates the need for pre-collected training data by enabling autonomous, target-driven exploration of a discrete design space. The framework is applied to optimize a silicon-based D-shaped photonic crystal fiber optical modulator (PCF-OM) incorporating a VO₂ phase-change layer, with the objective of minimizing insertion loss (IL). The proposed DQN-RL model demonstrates fast, stable, and consistent convergence behavior. Across multiple independent training runs with different random seeds, all runs converge to the same optimal solution with negligible variation, confirming the robustness of the learned policy. A comprehensive comparative study against particle swarm optimization (PSO), random search, grid search, and Bayesian optimization (BO), performed under identical design space, computational budget, and hardware conditions, demonstrates that the proposed DQN-RL framework achieves the fastest convergence, reaching the optimal solution within only 12 iterations (~ 3&#xa0;min), compared to 42, 104, 21, and 103 iterations for PSO, random search, grid search, and BO, respectively. reaching an ultra-low IL of 0.935 dB/mm, significantly outperforming the target value of 2 dB/mm. Using the optimized geometry obtained from DQN-RL framework, the PCF-OM exhibits excellent modulation performance, with an extinction ratio exceeding 280 dB/mm, a maximum modulation depth of 99.9%, and broadband operation. Furthermore, fabrication tolerance analysis confirms robust performance, with IL remaining below 1 dB/mm under ± 5% parameter variations. These results demonstrate that DQN-RL provides a fast, robust, and highly effective inverse-design strategy for complex photonic devices.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI-enhanced inverse design of photonic crystal fiber optical modulator using deep reinforcement learning technique

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Nada Yazeed M. Dawood,
  • B. M. Younis,
  • Mohamed Farhat O. Hameed,
  • Hamdi A. El-Mikati,
  • S. S. A. Obayya,
  • Nihal F. F. Areed

摘要

This paper presents a reinforcement learning (RL)-based inverse design framework for photonic device optimization, where a Deep Q-Network (DQN-RL) agent is directly coupled with three-dimensional finite-difference time-domain (3D-FDTD) simulation environment. The proposed approach eliminates the need for pre-collected training data by enabling autonomous, target-driven exploration of a discrete design space. The framework is applied to optimize a silicon-based D-shaped photonic crystal fiber optical modulator (PCF-OM) incorporating a VO₂ phase-change layer, with the objective of minimizing insertion loss (IL). The proposed DQN-RL model demonstrates fast, stable, and consistent convergence behavior. Across multiple independent training runs with different random seeds, all runs converge to the same optimal solution with negligible variation, confirming the robustness of the learned policy. A comprehensive comparative study against particle swarm optimization (PSO), random search, grid search, and Bayesian optimization (BO), performed under identical design space, computational budget, and hardware conditions, demonstrates that the proposed DQN-RL framework achieves the fastest convergence, reaching the optimal solution within only 12 iterations (~ 3 min), compared to 42, 104, 21, and 103 iterations for PSO, random search, grid search, and BO, respectively. reaching an ultra-low IL of 0.935 dB/mm, significantly outperforming the target value of 2 dB/mm. Using the optimized geometry obtained from DQN-RL framework, the PCF-OM exhibits excellent modulation performance, with an extinction ratio exceeding 280 dB/mm, a maximum modulation depth of 99.9%, and broadband operation. Furthermore, fabrication tolerance analysis confirms robust performance, with IL remaining below 1 dB/mm under ± 5% parameter variations. These results demonstrate that DQN-RL provides a fast, robust, and highly effective inverse-design strategy for complex photonic devices.