An alternative AI-based inverse design paradigm for photonic devices

Abstract

This paper introduces a pioneering inverse design paradigm for the refinement and enhancement of photonic devices utilizing an AI-based algorithm known as reinforcement learning (RL). The efficacy of this novel model is examined through calibrating a selected photonic device geometry to achieve the desired device response. A comparative analysis reveals the remarkable potency of the RL framework as it outperformed conventional optimizers, such as particle swarm optimization (PSO) and a modified version of the trust region algorithm (MTR). The RL framework yielded substantial enhancement in the device performance by 41.8% for the PSO design and 71.9% for the MTR design. Additionally, the RL framework demonstrated fast convergence toward the most optimal design in only 55 iterations, which is a stark departure from the 1000 iterations required by the PSO algorithm, and the 1700 iterations required by the MTR algorithm. Furthermore, the A2C-RL model was extended to enhance the geometry of a different PCF, while simultaneously incorporating liquid crystal infiltration into selected air holes in the PCF to produce wide-band ultra-flattened dispersion compensators, showcasing the versatility and effectiveness of A2C-RL in generating unconventional design choices, while reaching the best possible design. Furthermore, the generalizability of the proposed model is rigorously illustrated by achieving a 15% enhancement in the bandwidth of a Ku-band broadband metamaterial absorber within merely seven iterations. This result underscores the efficacy and adaptability of the A2C-RL model, affirming its potential for broader application across a wide range of photonic devices.