Deep Learning-Based Coverage Map Generation for RIS-Assisted Wireless Networks

Abstract

Reconfigurable Intelligent Surfaces (RISs) offer a promising approach for improving wireless coverage in obstructed environments, but evaluating their site-specific impact requires repeated ray-tracing simulations over multiple placements and configurations. Existing learning-based radio-map methods do not explicitly model the local propagation response induced by a given RIS deployment. This thesis formulates RIS-aware local coverage prediction, where the objective is to predict an effective pathloss map over a receiver (RX)-centered region of interest from the scene layout and the transmitter (TX), RIS, and candidate RX geometries. To address this problem, this thesis proposes RISMapNet, a four-channel framework that adapts the PMNet encoder-decoder backbone with RIS- and candidate-RX-aware input conditioning, auxiliary global supervision, and global-to-local refinement. A Sionnabased dataset generation pipeline is also developed to construct effective pathloss targets through sparse virtual steering. On the University of Southern California (USC) scene, RISMapNet reduces the normalized mean absolute error from 0.06714 to 0.0406 and the physical pathloss prediction error from 18.46 dB to 10.80 dB relative to the two-channel PMNet baseline. RISMapNet also outperforms PMNet-based baselines on unseen campus scenes, demonstrating its potential as an efficient surrogate for RIS-assisted coverage prediction and candidate-placement screening.