⚡ electrical engineering

Standard-Compliant Neuromorphic Integrated Sensing and Communications Aided by an Intelligent Reflecting Surface

This paper proposes a standard-compliant neuromorphic integrated sensing and communications (N-ISAC) system that utilizes a spiking neural network receiver to jointly demodulate data and detect targets from IEEE 802.15.4z UWB waveforms, while analyzing the performance-energy trade-offs introduced by a frequency-selective reconfigurable intelligent surface (RIS) in a realistic ray-traced channel.

Original authors: Jiho Park, Jiechen Chen, Joonhyuk Kang, Osvaldo Simeone

Published 2026-08-17

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiho Park, Jiechen Chen, Joonhyuk Kang, Osvaldo Simeone

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of wireless communication as a bustling, noisy city. In this city, devices like your phone or a smart tag are constantly shouting messages to each other, trying to be heard over the roar of traffic and the echo of tall buildings. For decades, engineers have built these devices to be like diligent accountants: they constantly check their math, calculate every possible path a signal could take, and use heavy, clock-driven processors to make sure the message gets through. But this approach is like trying to solve a complex math problem every time a single drop of rain hits a puddle—it uses a lot of energy and takes time.

Enter a new idea called "neuromorphic computing." Think of this not as a calculator, but as a brain. Instead of constantly crunching numbers, a brain only fires when something interesting happens—a "spike" of activity. This is incredibly efficient. Now, imagine combining this brain-like efficiency with a special kind of radio signal called Ultra-Wideband (UWB). UWB is like a super-fast, super-short flash of light; it's so precise it can tell you exactly how far away something is, almost like a radar, while also sending data. The big question scientists are asking is: Can we build a wireless device that uses this "brain-like" processing to do both talking and sensing at the same time, without burning through its battery? This is the challenge of "Integrated Sensing and Communications" (ISAC), and it's crucial because our future devices need to be smarter, faster, and much more energy-efficient.

The Brainy Radio That Listens to the City

In this paper, a team of researchers builds a prototype for a wireless system that acts like a nervous system. They call it a "Standard-Compliant Neuromorphic Integrated Sensing and Communications" (N-ISAC) system. To understand what they did, let's break down the cast of characters in their story.

First, there's the Signal. Instead of using a complex, continuous wave, they use a standard radio format called IEEE 802.15.4z. Think of this as a very specific, standardized language that real-world devices (like the ones in your phone) already speak. It uses "Impulse Radio," which means it sends data in tiny, ultra-short bursts, like a rapid-fire series of Morse code dots.

Second, there's the Environment. In the real world, signals bounce off buildings, cars, and trees. To make things even more interesting (and difficult), the researchers added a "Reconfigurable Intelligent Surface" (RIS) to their simulation. Imagine a giant, high-tech mirror on a wall that can be programmed to bend radio waves. However, this mirror isn't perfect; it's made of special materials that react differently to different frequencies, kind of like how a prism splits white light into a rainbow. This causes the radio pulses to stretch out and get a bit messy, which is a problem for both sending data and detecting objects.

Third, there's the Receiver. This is the star of the show. Instead of a traditional computer chip that calculates everything, they use a Spiking Neural Network (SNN). This is a computer program that mimics a biological brain. It only "fires" or does work when it receives a specific event (a spike). This makes it incredibly energy-efficient because it doesn't waste power doing math when there's nothing to do.

The Problem: The Messy Mirror and the Tired Brain

The researchers wanted to see if this brain-like receiver could handle the messy reality of the real world. Specifically, they were worried about two things:

The Mirror's Distortion: Because the RIS (the smart mirror) is frequency-selective, it smears the radio pulses. This makes it harder to tell exactly when a signal arrived, which is critical for both reading a message and spotting a target (like a car or a person).
The Energy Cost: Traditional receivers try to fix this mess by doing heavy calculations (channel estimation and equalization) before they can even start reading the message. This uses a lot of power.

The team asked: Can a brain-like receiver look at the messy, distorted signal, figure out what's going on, and decode the message and spot a target, all while using very little energy?

What They Did: The "Hypernetwork" Trick

To answer this, they didn't just build a simple brain; they built a brain that could learn on the fly. They created a system with two parts:

The Main Brain (SNN): This part processes the actual data and decides if a target is present.
The Coach (Hypernetwork): This is a smaller neural network that looks at the beginning of the message (the "preamble" or "SYNC" field)—a known pattern sent before the real data. Based on how that known pattern got distorted by the channel and the mirror, the Coach instantly adjusts the settings of the Main Brain.

Think of it like a musician playing a song in a room with bad acoustics. Instead of stopping to measure the room's echo and rewrite the sheet music (which takes time and energy), the musician listens to the first few notes (the preamble), instantly adjusts their playing style, and continues the song perfectly.

They also added a "sparse" feature. Since the brain only works when there's a spike, they programmed the system to ignore weak, noisy signals and only pay attention to the strong, important ones. This is like ignoring the background chatter in a crowded room and only listening to the person shouting your name.

The Findings: A Trade-Off and a Win for Efficiency

After running thousands of simulations in a virtual city with their smart mirror, the researchers found some fascinating results:

1. The Mirror is a Double-Edged Sword
They discovered that the "smart mirror" (RIS) creates a tricky trade-off.

For Communication: If the mirror is tuned to be very "sharp" (high quality factor), it distorts the signal more, making it harder to read the data. A "softer" mirror is better for sending clear messages.
For Sensing: Surprisingly, a "sharper" mirror actually helps with sensing! By filtering out some of the background noise and clutter, the target (like a car) stands out more clearly against the background.
The Conclusion: You can't have the best of both worlds with a single setting. The researchers showed that you have to balance the mirror's settings depending on whether you care more about sending a text message or spotting a moving object.

2. The Brain Wins on Energy
When they compared their brain-like receiver to traditional "accountant" receivers, the difference was huge.

Traditional receivers that try to mathematically fix the signal before reading it used about 4.2 millijoules of energy per message frame.
Their brain-like receiver, which adapts directly to the signal without heavy math, used less than 91 microjoules.
That's a massive drop in energy usage! By adding the "sparse" feature (ignoring weak signals), they could cut the energy usage even further, down to about 51 microjoules, with only a tiny drop in performance.

3. No Magic, Just Smart Design
The paper explicitly rules out the idea that you need a separate, perfect channel map to make this work. Their system doesn't need to know the exact layout of the city or the mirror beforehand. It learns from the "preamble" (the start of the message) every single time. This is a big deal because it means the system can work in a changing environment without needing to be re-calibrated constantly.

The Bottom Line

This paper suggests that we can build wireless devices that are both smart and energy-efficient by mimicking how brains work. By using a brain-like processor that adapts instantly to the environment and ignores unnecessary noise, we can send data and detect objects simultaneously without draining the battery.

However, the authors are careful to note that these results come from simulations in a controlled, virtual city. While the math looks promising and the energy savings are theoretically massive, this is a step toward a real-world technology, not the finished product itself. They showed that the "brain" approach is viable and that the "smart mirror" adds a layer of complexity that requires careful balancing. It's a proof of concept that says, "Hey, this brain-like way of doing things could be the future of wireless tech, provided we can build the hardware to match the software."

Problem Statement

The paper addresses the challenge of implementing Integrated Sensing and Communications (ISAC) in resource-constrained wireless devices using neuromorphic computing. While Impulse Radio Ultra-Wideband (IR-UWB) signaling is naturally suited for neuromorphic processing due to its sparse, event-driven nature, prior research on Neuromorphic ISAC (N-ISAC) has relied on simplified models. Specifically, previous work assumed basic Pulse Position Modulation (PPM) interfaces and simplified statistical multipath channels, failing to account for standardized physical layer protocols or realistic propagation environments.

Furthermore, the integration of Reconfigurable Intelligent Surfaces (RIS) to enhance ISAC coverage introduces a critical complication: RIS elements are typically metamaterial-based and exhibit frequency-selective responses. For wideband IR-UWB signals, this frequency selectivity causes amplitude attenuation and phase dispersion, which can distort pulses, degrade time-of-arrival resolution for sensing, and impair data demodulation. The paper seeks to evaluate N-ISAC performance under these realistic constraints: a standardized IEEE 802.15.4z High-Rate Pulse Repetition Frequency (HRP) UWB waveform, a ray-traced channel with a frequency-selective RIS, and a single Spiking Neural Network (SNN) receiver performing joint tasks.

Methodology

The authors propose a system model and receiver architecture that bridges the gap between theoretical neuromorphic concepts and commercial standards.

1. System Model:

Waveform: The system utilizes the full IEEE 802.15.4z HRP UWB physical layer. This includes a Synchronization Header (SHR) with a known ternary preamble code (SYNC and SFD fields) and a PHY payload using Burst-Position Modulation and Binary Phase-Shift Keying (BPM-BPSK).
Channel: A realistic ray-traced channel model is employed, aligned with 3GPP TR 38.901 ISAC standards. It decomposes the channel into background clutter (static and dynamic) and target-induced echoes. Crucially, the RIS is modeled with a frequency-selective Lorentzian response, which disperses wideband pulses based on the RIS quality factor ( $Q$ ).
Target: A passive radar target is modeled with direct and RIS-assisted interaction paths, where the target state (presence/absence) is a binary variable.

2. Receiver Architecture (HNet SNN):
The proposed receiver replaces traditional explicit channel estimation and equalization with a Hypernetwork-based Spiking Neural Network (HNet SNN):

Sparsification: Received chip-rate observations (both SYNC and payload) are sparsified based on magnitude to retain only the most significant chips, reducing input dimensionality.
Hypernetwork Adaptation: A hypernetwork takes the sparsified, synchronized SYNC observations as input and generates frame-specific scaling coefficients for the weights of the main SNN backbone. This allows the receiver to adapt to the current channel conditions (including RIS dispersion) without explicit channel estimation or pilot signaling beyond the standard SHR.
Joint Inference: The adapted SNN processes the sparsified payload symbols. It produces two outputs:
- Communication: Soft Log-Likelihood Ratios (LLRs) for BPM and BPSK bits, fed into a standard-aligned Forward Error Correction (FEC) decoder.
- Sensing: A symbol-wise target presence probability, which is averaged over the frame to make a final target detection decision.
Energy-Aware Training: The training objective combines cross-entropy losses for communication and sensing with a sparsity-promoting regularizer. This regularizer controls the membrane potential activity to minimize the number of spikes, thereby tuning the trade-off between ISAC performance and event-driven computation energy.

Key Contributions

The paper claims the following specific contributions:

Standard-Compliant N-ISAC: Development of an N-ISAC system based on the complete IEEE 802.15.4z HRP UWB physical layer, moving beyond the basic PPM and simplified channels of prior work.
Frequency-Selective RIS Modeling: Adoption of a ray-traced channel model incorporating a frequency-selective metamaterial RIS (via a Lorentzian response) to analyze how dispersion affects joint UWB communication and sensing.
Preamble-Driven Channel Adaptation: Proposal of a receiver where a hypernetwork generates per-frame neuron-wise scaling coefficients directly from the standardized SYNC preamble, eliminating the need for explicit channel estimation or inversion.
Energy-Aware Optimization: Formulation of a training objective that jointly optimizes ISAC performance and spike-driven synaptic activity, enabling a tunable trade-off between accuracy and energy consumption.
Comprehensive Evaluation: Extensive numerical experiments quantifying the impact of RIS frequency selectivity, standardized payload parameters, and the energy efficiency of the proposed receiver against channel-estimation-based baselines.

Experimental Results

The authors evaluated the system using site-specific urban ray-traced channel realizations (generated via Sionna RT) with a 6.4896 GHz carrier frequency.

Preamble-Driven Adaptation: The HNet SNN achieved favorable joint performance, reaching a Block Error Rate (BLER) of $\approx 10^{-2}$ at -5 dB SNR and reducing target detection error below 0.1. While channel-estimation-based baselines (LS/LMMSE) showed slightly better BLER at high SNR, they failed at target detection (error $\approx 0.5$ ), indicating that equalized payload representations are ineffective for sensing in this context.
RIS Frequency Selectivity: A trade-off was observed based on the RIS quality factor ( $Q$ $Q$ ).
- Low $Q$ : Provided a broadband response, yielding the best communication BLER but poorer target detection due to strong background clutter.
- High $Q$ : Increased off-resonant attenuation and dispersion, degrading communication but improving target detection by suppressing target-unrelated background components.
- Neural Capacity: Wider SNNs showed a systematic decrease in firing rates as $Q$ increased, suggesting a coupling between RIS selectivity and neuromorphic processing efficiency.
Standardized Payload: Higher data rates (shorter symbol durations) increased the ratio of channel delay spread to symbol duration, worsening communication BLER due to inter-symbol interference, though they offered modest sensing advantages.
Computation Energy:
- The HNet SNN reduced computation energy to < 91 $\mu$ J per frame**, significantly lower than the **> 4.2 mJ required by baselines using explicit channel estimation and equalization.
- Chip-domain sparsification (reducing input retention ratio $\rho$ ) further lowered energy (e.g., from 89 $\mu$ J to 51 $\mu$ J at $\rho=0.5$ ) with minimal performance loss.
- Sparsity regularization ( $\lambda_r$ ) reduced the empirical firing rate (e.g., from 0.21 to 0.05) and energy consumption, though with a slight degradation in communication reliability at high regularization levels.

Significance and Claims

The paper concludes that the proposed system demonstrates the feasibility of deploying neuromorphic ISAC on standard-compliant hardware. The primary significance lies in the ability to perform joint sensing and communication without explicit channel estimation or payload equalization, relying instead on preamble-driven hypernetwork adaptation.

The authors claim that this approach achieves a favorable joint ISAC operating region with substantially lower computation energy than traditional channel-estimation-based receivers. Furthermore, the study highlights that RIS frequency selectivity is not merely a propagation artifact but a critical design parameter that creates a fundamental trade-off between communication reliability and sensing accuracy, necessitating co-design of the RIS response and the neural receiver's capacity. The work establishes a framework for evaluating neuromorphic systems under realistic, wideband, and frequency-selective propagation conditions.