ReQuestNet: A Foundational Learning model for Channel Estimation

Kumar Pratik 1 , Pouriya Sadeghi 2 , Gabriele Cesa 1 , Sanaz Barghi 2 , Joseph B. Soriaga 2 , Yuanning Yu 2 , Supratik Bhattacharjee 2 3 , Arash Behboodi 1 · 2025

Read ReQuestNet: A Foundational Learning model for Channel Estimation by Kumar Pratik 1 , Pouriya Sadeghi 2 , Gabriele Cesa 1 , Sanaz Barghi 2 , Joseph B. Soriaga 2 , Yuanning Yu 2 , Supratik Bhattacharjee 2 3 , Arash Behboodi 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-08790.txt.

Introduction

Kumar Pratik and co-authors presented ReQuestNet in 2025, a learning model that estimates wireless channels in 5G and beyond. Channel estimation tells a receiver how the radio path distorts a signal. Classic methods lean on extra reference signals and stall when many antennas and unknown precoding enter the picture.

One unified model handles varying block counts, layer counts, bundling sizes, and pilot patterns. A coarse stage first estimates each stream block by block. A refinement stage then uses links across blocks and across antenna paths. In tests, the model beat even an idealized classic estimator across many channel conditions, gaining up to ten decibels at high signal strength. It also held up on channel types it never saw in training.

Catalog record

CREATORKumar Pratik 1 , Pouriya Sadeghi 2 , Gabriele Cesa 1 , Sanaz Barghi 2 , Joseph B. Soriaga 2 , Yuanning Yu 2 , Supratik Bhattacharjee 2 3 , Arash Behboodi 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-08790

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ReQuestNet: A Foundational Learning model for Channel Estimation Thanks: Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. Corresponding author: Kumar Pratik ( [email protected] ) ©2025 Qualcomm Technologies, Inc. and/or its affiliated companies. All Rights Reserved. Thanks: 3 Work completed while affiliated with Qualcomm Technologies Inc., USA. Kumar Pratik 1 , Pouriya Sadeghi 2 , Gabriele Cesa 1 , Sanaz Barghi 2 , Joseph B. Soriaga 2 , Yuanning Yu 2 , Supratik Bhattacharjee 2 3 , Arash Behboodi 1 Abstract In this paper, we present a novel neural architecture for channel estimation (CE) in 5G and beyond, the Recurrent Equivariant UERS Estimation Network (ReQuestNet). It incorporates several practical considerations in wireless communication systems, such as ability to handle variable number of RB ( RB ), dynamic number of transmit layers, physical resource block groups (PRGs) bundling size (BS), demodulation reference signal (DMRS) patterns with a single unified model, thereby, drastically simplifying the CE pipeline. Besides it addresses several limitations of the legacy linear MMSE solutions, for example, by being independent of other reference signals and particularly by jointly processing MIMO layers and differently precoded channels with unknown precoding at the receiver. ReQuestNet comprises of two sub-units, CoarseNet followed by RefinementNet. CoarseNet performs per PRG, per transmit-receive (Tx-Rx) stream channel estimation, while RefinementNet refines the CoarseNet channel estimate by incorporating correlations across differently precoded PRGs, and correlation across multiple input multiple output (MIMO) channel spatial dimensions (cross-MIMO). Simulation results demonstrate that ReQuestNet significantly outperforms genie MMSE ( MMSE ) CE across a wide range of channel conditions, delay-Doppler profiles, achieving up to 10dB gain at high SNRs. Notably, ReQuestNet generalizes effectively to unseen channel profiles, efficiently exploiting inter-PRG and cross-MIMO correlations under dynamic PRG BS and varying transmit layer allocations. 3GPP 3rd Generation Partnership Project 5G fifth-generation ADC analog to digital converter AFE analog front end AGC automatic gain control AGV automated guided vehicle AMP approximate message passing API Application Programming Interface AWGN additive white Gaussian noise AE auto-encoder BCE binary cross-entropy BER bit error rate BB baseband bpcu bits per channel use BP belief propagation BPSK binary phase shift keying BS base station CB codebook CE channel estimation CL contrastive-loss CFAR Constant-False-Alarm-Rate CDF cumulative distribution function CDM code division multiplexing CFO carrier frequency offset CoSaMP compressive sampling matching pursuit CP cyclic prefix CS compressive sensing CSI channel state information CNN convolutional neural network DBSCAN Density-Based-Spatial-Clustering-of-Applications DA domain adaptation DBA Distance-based Accuracy DAC digital-analog-converter DC direct current DMRS demodulation reference signal DE distance error DeepL deep-learning DoF degree-of-freedom DFT discrete Fourier transformation DL deep learning DS delay spread DGPS Differential Global Positioning Systems DSP digital signal processing EPS Constant-False-Alarm-Rate ECC error-correcting code ENoB effective number of bits ERP effective radiated power EVM error vector magnitude EVD eigenvector decomposition FB feedback FP false positive FN false negative FC fully connected FDD frequency division duplexing FDM frequency division multiplexing FIR finite impulse response FFT fast fourier transform FT fine tuning FPGA field programmable gate array GAN Generative adversarial network gNB next generation node B GPIO general-purpose input/output GPS global positioning system GPSDO GPS disciplined oscillator GPU graphical processing unit HDF Hierarchical Data Format HDD hard decision decoding IC integrated circuit ICI inter-carrier-interference ISAC Integrated Sensing And Communication I2C Inter-Integrated Circuit ICSP in-circuit serial programming IF intermediate frequency i.i.d. independent and identically distributed IIR infinite impulse response IMU inertial measurement unit IoT Internet of Things IPS indoor positioning system IR infrared JSDM Joint Spatial Division and Multiplexing LiDAR Light Detection And Ranging LLR log-likelihood ratio LP leakage precoder LMMSE Linear Minimum Mean Square Error LO local oscillator LoS line of sight LiDaR Light Detection and Ranging LM likelihood module LS least squares LSTM long-term short-term memory LTE Long Term Evolution LTI linear time invariant LTV linear time variant MAP maximum a posteriori ML maximum likelihood MMSE minimum mean squared error mmWave millimetre Wave MUSIC Multiple Signal Classification NN Neural Network MLP multilayer perceptron NNI Neural Network Intelligence NLoS non-line of sight KNN k-nearest neighbors OCC orthogonal cover codes OOD out of distribution OFDM orthogonal frequency division multiplex PRB physical resource block PRG physical resource block group RADAR Radio Detection And Ranging RGB Red-Green-Blue ReLU rectified linear unit RB resource block RE resource element RIM Recurrent inference machine RF radio frequency RMS-DS Root Mean Square - Delay Spread RNN recurrent neuronal network RSSI received signal strength indicator R-ZF regularized zero-forcing SDD soft decision decoding SDR software defined radio SE spectral efficiency SFO sampling frequency offset STO sampling time offset SLAM Simultaneous Localization and Mapping SGD stochastic gradient descent SISO single input single output SINR signal-to-interference-and-noise-ratio SIR signal-to-interference-ratio SLNR signal-to-leakage-and-noise ratio SNR signal-to-noise-ratio SP subspace SQR signal-to-quantization-noise-ratio SQNR signal-to-quantization-noise-ratio SVD singular value decomposition SU single-user TDD time division duplexing TRIPS time-reversal IPS TRP transmission and reception point TP true positive TN true negativ UE user equipment UL uplink ULA uniform line array URLLC ultra-reliable low-latency communication US uncorrelated scattering USRP universal software radio peripheral UWB ultra-wideband WiFi Wireless Fidelity WSS wide sense stationary WSSUS wide sense stationary uncorrelated scattering YOLO You-Only-Look-Once ZF zero forcing MSE Mean Squared Error NMSE Normalized MSE TDL tapped delay line CDL clustered delay line LOS Line of Sight NLOS Non-LOS SCS Subcarrier spacing I Introduction The advent of 5G NR and the anticipated evolution toward sixth-generation (6G) networks have ushered in an era of unprecedented connectivity, data throughput, and system complexity. These developments necessitate advanced techniques for low-power, compute-efficient, and reliable wireless communication. CE ( CE ) is a fundamental building block of modern wireless communication systems, and high quality CE is essential for achieving the target throughput required by many applications. Orthogonal Frequency Division Multiplexing (OFDM), a foundational modulation scheme in 5G NR, creates parallel communication channels across a large time-frequency grid. To acquire CSI ( CSI ), the pilot signals known as DMRS ( DMRS ) is used, whose time-frequency positions and values are known a priori to both transmitter and receiver. The CE task involves estimating CSI at non-DMRS resource elements (REs) from noisy DMRS observations. The standard allows for a variety of DMRS patterns [ 1 ] , tailored to different use cases and deployment scenarios. The CE needs to work across various system configurations and channel conditions, and to run efficiently for all hundreds of channels within the time-frequency resource grid. Techniques such as variable configuration for DMRS and leveraging other reference signals are additional overheads for the 5G stack, and will get more cumbersome as the channel dimensionality increases. A better channel estimation algorithm can help to trade off some of the existing overheads, for example by decreasing DMRS density, while keeping the CE quality at the desired level. Acquiring the channel statistics, adapting the underlying CE algorithm accordingly, and maximize the processing gain by joint estimation across the whole resource grid are among the challenges of traditional methods. Linear MMSE based solutions are among the best methods for the channel estimation, given their optimality for jointly Gaussian estimation problems. However, MMSE methods rely on precise knowledge of channel and noise statistics—information that needs to be acquired beforehand. It is important to note that when MMSE methods are done jointly over a grid of thousands of resouruces, they incur high computational cost due

Read the complete text