End-to-end Wireless Path Deployment with Intelligent Surfaces Using Interpretable Neural Networks

TL;DR

This paper introduces an interpretable neural network-based algorithm for orchestrating intelligent surface tiles to optimize wireless communication, power transfer, and security by controlling electromagnetic wave reflection.

Contribution

It presents a novel neural network model mapping surface tiles to network nodes, enabling optimized wave manipulation through interpretable weights for end-to-end wireless path deployment.

Findings

Neural network mapping of tiles enables effective wave control.

Optimization aligns surface functionalities with communication goals.

Interpretability allows understanding of wave manipulation strategies.

Abstract

Intelligent surfaces exert deterministic control over the wireless propagation phenomenon, enabling novel capabilities in performance, security and wireless power transfer. Such surfaces come in the form of rectangular tiles that cascade to cover large surfaces such as walls, ceilings or building facades. Each tile is addressable and can receive software commands from a controller, manipulating an impinging electromagnetic wave upon it by customizing its reflection direction, focus, polarization and phase. A new problem arises concerning the orchestration of a set of tiles towards serving end-to-end communication objectives. Towards that end, we propose a novel intelligent surface networking algorithm based on interpretable neural networks. Tiles are mapped to neural network nodes and any tile line-of-sight connectivity is expressed as a neural network link. Tile wave manipulation functionalities are captured via geometric reflection with virtually rotatable tile surface norm, thus being able to tunable distribute power impinging upon a tile over the corresponding neural network links, with the corresponding power parts acting as the link weights. A feedforward/backpropagate process optimizes these weights to match ideal propagation outcomes (normalized network power outputs) to wireless user emissions (normalized network power inputs). An interpretation process translates these weights to the corresponding tile wave manipulation functionalities.