A model combining neural networks and transfer functions (neuro-TF) is developed as an efficient way to parametric modeling electromagnetic(EM) responses. For coding metasurface, a new pole-residue tracking technique is improved to make the orders of pole-residue-based transfer functions consistent. However, the technique causes non-unique problems between poles/residues and corresponding EM responses. To address this issue, multiple cost functions are designed and a multitask neural network framework is constructed. During the training, two tasks are trained alternately in the same networks to speed up convergence by sharing the same representations such as weights and bias. Ultimately, poles/residues and corresponding EM responses are obtained concurrently. Compared with other existing neuro-TF network modeling methods, the proposed model can obtain accurate poles/residues as well as predict EM responses generated by transfer functions simultaneously in challenging applications of coding metasurface.
Traditional forward prediction of coding metasurface is highly time-consuming due to repetitive numerical calculations. In this paper, an advanced pole-residue-based neuro-transfer function (neuro-TF) technique is proposed for parametric modeling and predicting electromagnetic (EM) response of coding metasurface with respect to the changes in coding values representing the geometry of metasurface. In the proposed model, neural networks are trained to learn the mapping between poles and residues of the pole-residue transfer functions and coding values of metasurfaces, and an objective function based on joint learning is designed for the model optimization to increase the accuracy and efficiency of the model. A soft-sharing model called customized gate control (CGC) is brought in to jointly predict the poles and residues. In order to further improve the model performance and accelerate the learning process, we propose an objective function, in which the pole prediction is introduced as auxiliary task to serve for the main task of response predicting. The weights of losses of tasks are respectively determined by homoscedastic uncertainty reflecting the training difficulty of each task from the perspective of output observation noise. The proposed method allows the existence of real pole-residue pairs as well as pairs in complex format, which makes its application less limited compared to existing researches. Experiment result shows that the proposed model achieves great generalization and improves the accuracy of EM response prediction.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.