Source-linked AI summary
Adaptive RIS-aided Communications through ML-based Generation of Phase Masks
Corwin Carpenter, Thomas Daltzis, George C. Trichopoulos, Jacek Kibilda, Joao F. Santos
TL;DR
RIS phase-mask codebooks are limited by the storage and coverage demands of changing channel conditions and deployments. The demo embeds an ML model on the RIS microcontroller to generate masks at runtime, closely matching ground-truth steering with average RSS only 0.26 dB lower.
Problem
Computationally generated phase masks stored in fixed codebooks cannot feasibly cover all channel conditions and deployment scenarios within embedded microcontroller storage.
Method
A lightweight neural network embedded on the RIS microcontroller approximates an analytical phase-mask generation algorithm from TX and RX coordinates.
Results
0.26 dB lower average RSS was obtained with ML-generated masks than with ground-truth masks at target steering locations.
Takeaways & Limitations
Runtime mask generation supports adaptive RIS beam management across new operating conditions without taking the RIS offline or updating its firmware.
Abstract
from arXiv · showhide
Reconfigurable Intelligent Surfaces (RISs) are an attractive technology for Millimeter Wave (mmWave) communications due to their ability to passively reflect incident signals. However, current implementations of RIS rely on performing computationally-intensive algorithms offline to generate phase masks, which are stored as a codebook on the embedded microcontroller on the RIS. The codebook size is restricted by the embedded microcontroller's storage capacity, which limits the ability of the RIS to adapt to evolving channel conditions and deployment scenarios. In this demo, we showcase an Machine Learning (ML)-based solution for dynamically generating new phase masks during runtime. Our approach leverages a ML model deployed on the microcontroller for approximating the output of a phase mask generation algorithm, responding to new inputs while remaining smaller than a codebook.
I. INTRODUCTION
mmWave links can lose strength when obstructed, motivating RISs that redirect signals. This demo replaces storage-limited fixed codebooks with an embedded ML model that generates phase masks during runtime for adaptive steering.
- Obstructions and path loss can reduce mmWave RSS and degrade or interrupt communication links.
- RIS phase masks configure element phase shifts, but optimal masks require computationally intensive algorithms and extensive channel geometry data.
- Codebooks cannot feasibly cover all channel conditions, deployment scenarios, and use cases within embedded microcontroller storage.
- An Arduino Mega could ideally store nearly 2000 masks, yet this yields only up to 13 uniformly distributed positions per dimension across three spatial dimensions.
- The demo embeds a lightweight ML model trained on pregenerated near-field masks to generate new masks dynamically for adaptive beam management.
II. DEMO DESIGN AND FUNCTIONALITY
The demo uses a non-line-of-sight mmWave link assisted by a 1024-element RIS. A Host PC controls adaptive beam management while an Arduino microcontroller loads generated masks into the RIS.
- The setup includes two mmWave radios communicating over a non-line-of-sight link aided by an RIS, with control provided by a Host PC and testbed equipment.
- The RIS has 1024 1-bit elements in a 32×32 grid, each applying either 0° or 180° phase shift.
- An Arduino GIGA loads each phase mask, converts it to a bitstream, and shifts the bits into RIS shift registers simultaneously.
- The Arduino runs custom firmware and an ML model converted from PyTorch to TensorFlow Lite for phase-mask generation.
III. ML-BASED GENERATION OF PHASE MASKS
The approach embeds a neural network on the RIS microcontroller to replace externally maintained codebooks with runtime mask generation. The model maps radio locations to approximate binary phase masks for steering reflected signals.
- Fixed codebooks require external mask generation, firmware updates, and taking the RIS offline when new operating conditions arise.
- Runtime ML generation broadens operating-condition adaptation without taking the RIS offline or updating its firmware.
- The model is a four-layer fully connected MLP with 128 binary input features, two 256-neuron ReLU layers, and one output per RIS element.
- An analytical algorithm generates labeled 32×32 ground-truth masks from TX and RX positions, with phase shifts quantized to 0° and 180°.
IV. NUMERICAL VALIDATION
The model was validated on near-field non-line-of-sight phase masks and evaluated across a coarse RX-position grid. Its generated masks closely matched ground-truth steering behavior, with a small average RSS difference.
- Training used 1681 near-field masks over a 41×41 RX grid, while evaluation used a 3×3 grid with 50 cm spacing at fixed z = 2 m.
- The current architecture supports movement along x and y at fixed z, with variable-distance generation planned for future work.
- 0.26 dB lower average RSS was measured for ML-generated masks than for ground-truth masks at target steering locations.
- During the demonstration, users can move the RX and observe RSS, generated masks, and array factors as the system updates steering through a dashboard.