What question did this study set out to answer?

The aim is to develop an efficient resource allocation framework for RIS-UAV-assisted THz NOMA networks.

February 26, 2026Journal of King Saud University - Computer and Information Sciences0 citationsOpen Access

Graph neural network-based end-to-end resource allocation for energy-efficient RIS-UAV-NOMA-THz networks

AKAdil Khan BHBabar Hayat SAShabeer Ahmad

Key Points

The aim is to develop an efficient resource allocation framework for RIS-UAV-assisted THz NOMA networks.
Formulated a single optimization problem for power allocation, SIC order, and other parameters.
Utilized a graph neural network to model the system and learn optimal policies.
Applied differentiable feasibility projection to ensure hard constraints were met.
Simulated performance against DRL/MARL and greedy approaches under various conditions.
Reduced latency and lower energy per bit compared to baseline approaches.
Improved spectral efficiency and reduced outage performance.
Quantitative gains increased with larger RIS apertures and finer phase resolution.

Abstract

Terahertz (THz) access provides multi-GHz bandwidth for sixth-generation (6G) services, but its short-range propagation, molecular absorption, blockage, and narrow-beam misalignment make end-to-end (E2E) delay-energy guarantees difficult when computation tasks are offloaded to an edge computing node (ECN). When THz access is combined with non-orthogonal multiple access (NOMA), feasible successive interference cancellation (SIC) ordering must also be maintained, further coupling radio decisions with queueing and computation constraints. A reconfigurable intelligent surface (RIS) mounted on an unmanned aerial vehicle (UAV) create a controllable reflected path to mitigate blockage, yet practical deployment introduces b-bit phase quantization, UAV kinematics, and battery limitations that invalidate many idealized control assumptions. This paper proposes a constraint-aware E2E resource allocation framework for RIS-UAV-assisted THz NOMA-MEC networks. We formulate a single optimization problem that decides users’ transmit powers, SIC order, b-bit RIS phase shifts, UAV 3D waypoints, task offloading ratios, and per-user edge CPU allocations, subject to spectral masks, power limits, queue stability, CPU budgets, UAV kinematics/battery evolution, and RIS quantization constraints. To handle the high-dimensional coupled state, we represent the system as a heterogeneous graph (users, RIS-UAV, and ECN) and learn a graph neural network (GNN) policy. A differentiable feasibility projection enforces hard constraints, while a weighted time-energy objective (uplink transmission, edge execution, downlink return, and UAV propulsion) is augmented with a CVaR term to control tail latency. Simulations show reduced latency, lower energy per bit and energy-delay product, and improved spectral efficiency and outage performance compared with DRL/MARL and greedy baselines, with gains increasing for larger RIS apertures and finer phase resolution.

Perguntar à IA

Bookmark

View Full Paper