[Paper Review] A novel guided deep learning algorithm to design low-cost SPP films

Yingshi Chen, Jinfeng Zhu|arXiv (Cornell University)|Dec 7, 2019

Plasmonic and Surface Plasmon Research19 references4 citations

TL;DR

This paper proposes a novel guided deep learning algorithm that designs low-cost surface plasmon polariton (SPP) films by training a deep convolutional neural network (CNN) with a low-cost sample replacement strategy, enabling the model to learn cost-effective material substitutions—such as replacing precious metals with ordinary metals—while maintaining high accuracy, achieving an average relative spectral error of less than 10%.

ABSTRACT

The design of surface plasmon polaritons (SPP) films is an ill-posed inverse problem. There are many-to-one correspondence between the structures and user needs. We present a novel guided deep learning algorithm to find optimal solutions (with both high accuracy and low cost). To achieve this goal, we use low cost sample replacement algorithm in training process. The deep CNN would gradually learn better model from samples with lower cost. We have successfully applied this algorithm to the design of low-cost SPP films. Our model learned to replace precious metals with ordinary metals to reduce cost. So the the cost of predicted structure is much lower than standard deep CNN. And the average relative error of spectrum is less than 10%. The source codes are available at https://github.com/closest-git/MetaLab.

Motivation & Objective

• To address the ill-posed inverse problem in SPP film design, where multiple structures can satisfy user requirements.
• To reduce the cost of SPP films without compromising spectral performance, particularly by minimizing reliance on expensive precious metals.
• To develop a deep learning framework that prioritizes low-cost material configurations during training.
• To improve the generalization and cost-efficiency of deep learning models in nanophotonic design through guided training with cost-aware data sampling.

Proposed method

• The algorithm employs a deep convolutional neural network (CNN) to map structural parameters to SPP spectral responses.
• A low-cost sample replacement strategy is integrated into the training process, where high-cost samples are progressively replaced with lower-cost alternatives during optimization.
• The model is trained iteratively on a dataset that prioritizes cost efficiency, allowing the network to learn cost-effective material choices such as aluminum or silver over gold or platinum.
• The training process is guided by a loss function that balances spectral accuracy and material cost, with cost-weighted sampling to emphasize low-cost solutions.
• The framework enables end-to-end learning of optimal SPP structures that are both spectrally accurate and economically viable.
• The model is validated using a benchmark dataset of SPP film configurations with known spectral responses and material costs.

Experimental results

Research questions

• RQ1 Can a deep learning model be trained to prioritize low-cost material choices in SPP film design while maintaining high spectral accuracy?
• RQ2 How effective is the low-cost sample replacement strategy in guiding the model toward economically viable solutions without degrading performance?
• RQ3 To what extent can the model replace precious metals with ordinary metals in SPP films without increasing spectral error?
• RQ4 What is the trade-off between material cost reduction and spectral accuracy in the predicted SPP structures?
• RQ5 How does the guided training approach compare to standard deep learning models in terms of cost and performance?

Key findings

• The proposed algorithm successfully replaced precious metals such as gold with ordinary metals like aluminum and silver in SPP film designs, significantly reducing material cost.
• The average relative error in the predicted SPP spectrum was less than 10%, indicating high spectral accuracy.
• The model achieved lower-cost predictions compared to standard deep CNNs, which typically do not incorporate cost-aware training.
• The low-cost sample replacement strategy effectively guided the network to learn cost-efficient material substitutions during training.
• The trained model generalized well to unseen SPP structures, maintaining both low cost and high accuracy.
• Source code for the model is publicly available, enabling reproducibility and further development in low-cost nanophotonic design.