Direction finding by learning-assisted programmable metasurface

This study explores the assistance of supervised machine learning to a programmable metasurface for precise direction finding. The programmable metasurface plays the role of an electronic-scanning reflector to scan the azimuth plane and redirect incident waves toward a single radio frequency receiver, which records the power level of the incoming signals. The collected data are then fed into a pre-trained model. Different machine learning algorithms are considered and their suitability for direction finding is assessed through a comparative analysis under realistic conditions. The study includes classical approaches as well as deep learning methods. The algorithms are evaluated based on different metrics, such as root mean square error, mean absolute angular error, determination coefficient, precision for specific error thresholds, and prediction speed, under a wide range of signal-to-noise ratio conditions. While deep learning algorithms require large training data sets to fully leverage their capabilities, classical tree-based algorithms exhibit notable stability but do not show high accuracy and robustness to noise. Classical algorithms such as K-Nearest Neighbors and Multi-layer Perceptron can maintain high accuracy even in very noisy environments, confirming their suitability for applications in realistic scenarios. The analysis also highlights that a trade-off between accuracy and speed can be required when selecting an algorithm.