Problem

The design of intelligent metasurfaces faces significant challenges in achieving rapid, automated adaptation to user-specific functionalities and dynamic environmental conditions. Traditional methods for designing freeform metasurfaces demand extensive computational resources due to the high dimensionality of the solution space and the intricate characterization of meta-atoms. Integrating deep learning models, while promising, is hindered by the need for large labeled datasets, frequent model fine-tuning, and inefficiencies in managing designs with a high degree of freedom. Furthermore, automating the deployment of these intelligent metasurfaces in complex and unpredictable environments remains a substantial hurdle.

Method

The paper proposes an MLOps-assisted inverse design framework for metasurfaces, implemented using Red Hat OpenShift AI (RHOAI). The core of the approach involves designing lossless impenetrable metasurfaces for optimal anomalous reflection by optimizing the number of Floquet modes and minimizing the net power density normal to the surface, aiming for a purely reactive surface impedance. A conditional Generative Adversarial Network (cGAN) is utilized for the metasurface design process. An extended cGAN, incorporating a surrogate model (Forward ResNet-50), is employed to predict spectral responses from metasurface patterns, acting as an efficient simulation tool. The cGAN implicitly learns the underlying physical relationships governing electromagnetic responses. The RHOAI platform provides an MLOps framework that automates the entire machine learning model lifecycle, including data gathering, cleansing, training, evaluation, and deployment, ensuring scalability and reproducibility. This system is integrated with Software-Defined Metasurfaces (SDMs) within Programmable Wireless Environments (PWEs) for control and interaction.

Results

The developed MLOps platform on Red Hat OpenShift AI (RHOAI) successfully streamlines the entire lifecycle of the cGAN model used for metasurface design. Performance evaluations of ResNet-50, a component within the generator's surrogate model, demonstrated near-native performance when trained on ImageNet within OpenShift 4.13+, achieving results within 0.4% of bare-metal deployments. Single-node training on RHOAI showed competitive Top-1 accuracy values ranging from 75.08% to 76.15%. The framework also provides real-time monitoring capabilities for training and testing errors through integrated RHOAI components like Model Monitoring dashboards and Pipelines. These results validate the feasibility and advantages of deploying such complex deep learning models in a containerized environment like Red Hat OpenShift for intelligent metasurface design.

Implications

The proposed multi-layered architecture for intelligent metasurfaces, supported by the RHOAI platform, significantly automates and simplifies the integration of machine learning models into the metasurface design workflow, from development to continuous monitoring and redeployment. This approach leverages cloud-native edge computing for creating efficient, scalable, and secure distributed applications, fostering the advancement of an Internet-of-Metasurfaces (IoM) with automated control over electromagnetic behavior. The convergence of microservices and metasurfaces, as explored in this work, represents a crucial research direction for future technologies like 6G and Open RAN (O-RAN). While the implementation simplifies MLOps for reconfigurable metasurfaces, further consideration for realistic application scenarios is highlighted as future work.