跳转至

文章背景与核心概要

本文探讨了基于MXene材料的太阳能吸收器在电磁光谱预测中所面临的计算强度挑战。为了替代传统耗时且计算量巨大的全波求解器,作者提出了一种高效的深度学习新框架,能够显著加速纳米光子学的正向设计流程。

该研究的核心创新在于将预训练的 MobileNet v2 迁移学习模型与多通道光谱精细化(multi-channel spectral refinement)模块相结合,并引入了 Savitzky-Golay 平滑算法来有效降低高频预测噪声。基于 \(64\times64\) 像素的超表面设计输入,该模型能够高精度地预测 102 点吸收光谱。实验结果表明,该框架在各项关键性能指标上均显著优越于标准的 CNN 架构,为快速实现纳米光子学器件设计提供了一种极具扩展性的技术方案。


Optimizing Spectral Prediction in MXene-Based Metasurfaces Through Multi-Channel Spectral Refinement and Savitzky-Golay Smoothing

arXiv: 2602.08406
Subjects: Optics (physics.optics); Artificial Intelligence (cs.AI); Signal Processing (eess.SP)
Authors: Shujaat Khan, Waleed Iqbal Waseer, Muhammad Shahid Jabbar

arXiv: 2602.08406
Subjects: Optics (physics.optics); Artificial Intelligence (cs.AI); Signal Processing (eess.SP)
Authors: Shujaat Khan, Waleed Iqbal Waseer, Muhammad Shahid Jabbar


Summary

本研究旨在解决基于 MXene 的太阳能吸收器电磁光谱预测中的计算密集性问题。作者提出了一种新颖的深度学习框架,用更高效的预测模型取代了传统且耗时的全波求解器。通过整合迁移学习(使用预训练的 MobileNet v2)、多通道光谱精细化以及 Savitzky-Golay 平滑,该模型基于 \(64\times64\) 的超表面设计,实现了对 102 点吸收光谱的高精度预测。与标准的 CNN 架构相比,该框架表现出更优越的性能,为快速纳米光子学设计提供了一种可扩展的解决方案。

Summary

This research addresses the computational intensity of predicting electromagnetic spectra for MXene-based solar absorbers. The authors propose a novel deep learning framework that replaces traditional, time-consuming full-wave solvers with a more efficient predictive model. By integrating transfer learning (using a pretrained MobileNet v2), multi-channel spectral refinement, and Savitzky-Golay smoothing, the model achieves high-accuracy predictions for 102-point absorption spectra based on \(64\times64\) metasurface designs. The framework demonstrates superior performance compared to standard CNN architectures, offering a scalable solution for rapid nanophotonic design.


Key Technical Contributions

  • 网络架构: 利用微调后的 MobileNet v2 来处理超表面几何结构并预测光谱输出。
  • 多通道精细化: 实现了通过多个卷积通道处理特征图的模块,以改善特征提取效果。
  • 噪声抑制: 采用 Savitzky-Golay 平滑技术,有效降低预测光谱中的高频噪声。
  • 性能指标:
    • 均方根误差 (RMSE): 0.0227
    • 决定系数 (\(R^2\)): 0.9563
    • 峰值信噪比 (PSNR): 33.10 dB

Key Technical Contributions

  • Architecture: Utilizes a fine-tuned MobileNet v2 to process metasurface geometry and predict spectral output.
  • Multi-Channel Refinement: Implements a module that processes feature maps through multiple convolutional channels to improve feature extraction.
  • Noise Mitigation: Employs Savitzky-Golay smoothing to effectively reduce high-frequency noise in the predicted spectra.
  • Performance Metrics:
    • RMSE: 0.0227
    • \(R^2\) (Coefficient of Determination): 0.9563
    • PSNR (Peak Signal-to-Noise Ratio): 33.10 dB

Publication Details

Publication Details


Access & Resources

Access & Resources

license icon