⚡ electrical engineering

DRIFT: Joint Channel Estimation and Prediction Towards Pilotless 6G Non-Terrestrial Networks

This paper proposes DRIFT, a lightweight, iterative joint channel estimation and prediction framework for 6G non-terrestrial networks that significantly reduces pilot overhead and computational complexity to enable efficient onboard implementation while achieving up to 12% spectral efficiency gains.

Original authors: Bruno De Filippo, Carla Amatetti, Alessandro Vanelli-Coralli

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bruno De Filippo, Carla Amatetti, Alessandro Vanelli-Coralli

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 you are trying to talk to a friend who is zooming past you on a high-speed train (a Low Earth Orbit satellite). Because they are moving so fast, your voice gets distorted by the Doppler effect, and the signal gets messy. To understand each other, you usually have to shout "Hello, Hello, Hello" repeatedly in between your real conversation. These "Hellos" are called pilots in engineering. They help the receiver figure out how to clean up the signal, but they take up valuable time and space that could be used for actual data.

The paper "DRIFT" proposes a clever way to stop shouting "Hello" so often, allowing for a faster, more efficient conversation.

Here is the breakdown of their idea using simple analogies:

1. The Problem: The "Hello" Tax

In current satellite systems, every time you send a chunk of data, you have to send a few "pilot" signals first. It's like sending a letter where 20% of the page is just your address and a "To Whom It May Concern" header, leaving less room for the actual story. This wastes energy and bandwidth.

2. The Solution: The "One-and-Done" Strategy

The authors suggest a new rule: Send the "Hello" only once at the very beginning.

Slot 1: You send the pilots (the "Hello") and the first batch of data.
Slots 2, 3, 4, etc.: You send only data. No more "Hellos."

The receiver then has to guess what the signal looks like for the rest of the conversation based on that first guess. The danger here is that if the receiver makes a small mistake in the beginning, that mistake gets bigger and bigger as time goes on (like a game of "Telephone" where the message gets garbled).

3. The DRIFT Engine: A Self-Correcting GPS

To stop the mistakes from getting worse, they built a smart system called DRIFT (Data-driven Refinement and Iterative Forecast). Think of DRIFT as a highly skilled navigator with two special tools:

Tool A: The "Refiner" (Cleaning the Lens)
When the receiver gets a new chunk of data, it makes a rough guess about what the message said. DRIFT looks at that guess and says, "Wait, that doesn't look quite right based on the pattern." It uses a lightweight AI to clean up the guess, making the channel estimate sharper. It's like using a lens cleaner on a camera; even if the photo is slightly blurry, the cleaner makes it crisp enough to read the text.
Tool B: The "Forecaster" (Predicting the Future)
Once the lens is clean, DRIFT predicts what the signal will look like in the next moment. Because the satellite moves in a predictable path, the signal changes in a pattern. DRIFT learns this pattern and says, "Based on where the satellite is now, the signal will look exactly like this in the next second."

4. The "Lightweight" Secret

Most AI models used for this are like supercomputers: heavy, power-hungry, and too big to fit on a satellite. Satellites have strict power limits (they run on solar panels and batteries).

The authors designed DRIFT to be tiny and efficient.

They used simple building blocks (Convolutional Neural Networks and TCNs) that are like a Swiss Army knife: small, simple, but effective.
The paper claims the whole system is so light that it requires fewer than 200,000 mathematical operations (MACs) to run. To put that in perspective, it's like a tiny, efficient robot compared to a massive, energy-guzzling factory. This makes it possible to run directly on the satellite without draining its battery.

5. The Results: Faster and Smarter

The team tested this in a computer simulation of a satellite link.

The Win: By stopping the "Hellos" after the first slot, they achieved a 12% increase in efficiency. This means they can send 12% more actual data in the same amount of time compared to traditional systems.
Robustness: Even when the training data (what the AI learned on) didn't perfectly match the test data (the real scenario), the system still worked well.
Stability: The system managed to keep errors low for up to 4 or 5 time slots before the "Telephone game" effect started to get too strong.

Summary

The paper presents a new way to talk to satellites. Instead of constantly shouting "Hello" to stay connected, the system says "Hello" once, then uses a tiny, smart AI assistant (DRIFT) to clean up its own guesses and predict the future. This saves energy, reduces waste, and lets satellites send more data to the world, all while running on a computer small enough to fit on a satellite's power budget.

Technical Summary: DRIFT – Joint Channel Estimation and Prediction Towards Pilotless 6G Non-Terrestrial Networks

Problem Statement
Non-terrestrial networks (NTNs), particularly those utilizing Low Earth Orbit (LEO) satellites, are critical for achieving ubiquitous connectivity in sixth-generation (6G) systems. However, LEO dynamics introduce strong, time-varying Doppler shifts and delays. To counteract "channel aging," channel prediction is essential for improving spectrum utilization by reducing pilot overhead. While artificial intelligence (AI) has been proposed for this task, existing solutions often rely on complex architectures (e.g., Transformers with >100M FLOPs) that are unsuitable for satellite onboard implementation due to strict power and computational constraints. Furthermore, many studies prioritize prediction accuracy (measured by NMSE) over end-to-end metrics like effective spectral efficiency (SE) and rarely evaluate performance under training-test mismatches.

Methodology
The paper proposes an iterative joint channel estimation and prediction framework designed for 6G uplink NTNs. The core innovation is the DRIFT (Data-driven Refinement and Iterative Forecast for wireless channel Tracking) architecture.

System Framework: The system operates over a sequence of $N_{slot}$ time slots. Pilots are transmitted only in the first slot ( $t=1$ ) to initialize the process. For subsequent slots ( $t \geq 2$ ), the system operates in a pilotless mode, relying entirely on data-driven processing.
Iterative Loop:
1. Initialization: In slot 1, a standard Least Squares (LS) pilot-aided (PA) estimate is computed and used for equalization.
2. Data-Aided (DA) Estimation: For subsequent slots, received symbols are equalized using the predicted channel from the previous iteration. Hard decisions are made on the equalized symbols to reconstruct the transmitted data. These reconstructed symbols serve as "ground truth" for a new DA channel estimation step.
3. DRIFT Processing: The DA estimate is fed into the DRIFT model, which performs two functions:
  - Channel Refinement: A lightweight block corrects errors introduced by inaccurate demapping in the DA estimation, producing a refined estimate ( $\hat{H}^{(E')}_t$ ).
  - Channel Prediction: The refined estimate is used to predict the Channel Frequency Response (CFR) for the next slot ( $\hat{H}^{(P)}_{t+1}$ ).
DRIFT Architecture: The model is designed for low complexity and operates on 3D tensors (real/imaginary parts of CFR). It consists of:
- Refinement Block: Uses a temporal causal convolution (CConv1D), activation, batch normalization, and a spectral convolution (Conv1D). It employs a dynamic weighting scheme that emphasizes updates on later, less reliable time slots.
- Prediction Block: Two variants are investigated:
  1. Hybrid CNN-LSTM: Uses Conv2D/TConv2D for compression/reconstruction and an LSTM layer for temporal prediction.
  2. Temporal Convolutional Network (TCN): Uses feed-forward dilated Conv1D layers with residual connections, avoiding recurrent structures.
Training Strategy: The models are trained using a curriculum learning approach where the loss function initially prioritizes channel refinement and gradually shifts focus toward prediction. Training utilizes synthetic data based on NTN-TDL fading models.

Key Contributions

Iterative Framework: Introduction of a pilotless framework where pilot-aided estimation is replaced by a cycle of data-aided estimation, refinement, and prediction.
DRIFT Model: A lightweight architecture that jointly refines data-aided channel estimates and predicts future CFRs with minimal computational cost.
End-to-End Evaluation: Assessment of the system based on effective Spectral Efficiency (SE) rather than just NMSE, explicitly accounting for error propagation and training-test mismatches (e.g., training on 16-QAM, testing on QPSK).
Complexity Analysis: Demonstration that the proposed models require fewer than 200k multiply-accumulate (MAC) operations, making them feasible for satellite payloads.

Results
Simulations were conducted in an end-to-end MATLAB environment for an uplink LEO scenario (600 km altitude, 2 GHz carrier).

Spectral Efficiency: The TCN-based DRIFT model achieved a ~12% SE gain compared to conventional pilot-based systems when using $N_{slot}=4$ (one pilot slot, three pilotless slots). The Hybrid CNN-LSTM variant showed gains but was limited by error propagation after $N_{slot}=3$ .
Robustness: The system maintained consistent performance across different channel models (NTN-TDL-C and NTN-TDL-D) without retraining. It also demonstrated robustness to mismatches between training (16-QAM) and testing (QPSK) data.
Accuracy: The channel refinement stage successfully reduced estimation NMSE to below -19 dB for the first three slots. Error propagation became significant only after the fourth slot.
Computational Complexity: The total inference complexity was calculated at 154k MACs (Hybrid CNN-LSTM) and 171k MACs (TCN). The refinement stage accounts for only ~10% of the total complexity. The authors estimate inference latencies of approximately 1.0–1.1 ms per slot using low-power hardware accelerators.

Significance and Claims
The paper claims that DRIFT addresses the critical trade-off between prediction accuracy and computational feasibility for 6G NTNs. By significantly reducing pilot overhead while maintaining robust performance under strict power constraints, the approach enables substantial spectral efficiency gains. The authors emphasize that the low MAC count (under 200k) makes the solution suitable for real-time, onboard satellite implementation, a capability often lacking in high-complexity AI predictors designed for terrestrial networks. The work concludes that while error propagation remains a challenge for longer pilotless sequences, the proposed framework offers a viable path toward pilotless communication in resource-constrained space environments. Future work is identified as evaluating real-world hardware latency and further reducing error propagation.