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

TL;DR

The paper tackles the high computational cost of predicting spectra for MXene-based metasurfaces using full-wave solvers. It introduces a transfer-learning–driven MobileNetV2 architecture augmented with a Multi-Channel Spectral Refinement module and Savitzky-Golay smoothing to predict 102-point absorption spectra from 64x64 designs, achieving RMSE 0.0245, R^2 0.9578, and PSNR 32.98 dB. The approach outperforms CNN and deformable CNN baselines, with 2.93 million parameters, and includes Grad-CAM-based explainability that aligns predictions with physical resonances. These results offer a scalable, efficient alternative for rapid nanophotonic design and optimization, with potential extension to multi-layer metasurfaces and inverse-design workflows.

Abstract

The prediction of electromagnetic spectra for MXene-based solar absorbers is a computationally intensive task, traditionally addressed using full-wave solvers. This study introduces an efficient deep learning framework incorporating transfer learning, multi-channel spectral refinement (MCSR), and Savitzky-Golay smoothing to accelerate and enhance spectral prediction accuracy. The proposed architecture leverages a pretrained MobileNetV2 model, fine-tuned to predict 102-point absorption spectra from $64\times64$ metasurface designs. Additionally, the MCSR module processes the feature map through multi-channel convolutions, enhancing feature extraction, while Savitzky-Golay smoothing mitigates high-frequency noise. Experimental evaluations demonstrate that the proposed model significantly outperforms baseline Convolutional Neural Network (CNN) and deformable CNN models, achieving an average root mean squared error (RMSE) of 0.0245, coefficient of determination \( R^2 \) of 0.9578, and peak signal-to-noise ratio (PSNR) of 32.98 dB. The proposed framework presents a scalable and computationally efficient alternative to conventional solvers, positioning it as a viable candidate for rapid spectral prediction in nanophotonic design workflows.

Figures (6)

• Figure 1: (a) Three-layer MIM metasurface-based absorber comprising a top MXene layer, middle SiO$_2$ substrate, and bottom silver reflector. (b) Corresponding absorption spectra demonstrating dual-band absorption.
• Figure 2: Proposed Multi-Channel Spectral Refinement and Savitzky-Golay smoothing framework integrated with MobileNetV2.
• Figure 3: (Top): Training (left) and validation (right) loss curves for the proposed model. The (bottom) plots illustrate the confidence interval across multiple runs, indicating consistent performance.
• Figure 4: Comparison of the predicted vs. actual spectra for six representative metasurface designs. The model accurately captures both broadband and narrowband absorption features, benefiting from the MCSR and smoothing layers.
• Figure 5: Model complexity vs. accuracy: Number of parameters vs. $R^2$ score for CNN, Deformable CNN, and Proposed Model.