Experience · Wireless Centre of Excellence
BT Group
Wireless Digital Twin & 5G Channel Modelling
Wireless Centre of Excellence · Digital Twins · RF Propagation · 5G
Developed a Python-based wireless digital-twin workflow using NVIDIA Sionna RT, OpenAirInterface and OAIBOX hardware to model radio propagation, generate time-varying channels and reproduce them through a hardware channel emulator for repeatable 5G testing.
Field trials are hard to reproduce
Real wireless field trials are expensive, time-consuming, hard to reproduce and sensitive to changing conditions. The aim: a digital-twin workflow where realistic propagation environments can be simulated, reproduced and measured repeatedly.
Simulation wired to real hardware
The core of the work connects a physics-based propagation model to actual 5G hardware emulation — it does not stop at plots from a propagation model.
Ray-traced radio environments in Sionna RT
I built the simulation workflow in Python with NVIDIA Sionna RT — one transmitter and multiple indoor / outdoor receiver positions in a reproducible scenario.
- 3D scene configuration
- Transmitter / receiver placement
- Carrier frequency
- Bandwidth
- Antenna configuration
- Ray tracing
- Radio maps
- 3.5 GHz
- 7.05 GHz
- 100 MHz bandwidth
The framework supported the 3.5 / 7.05 GHz comparison, but broader 7.05 GHz hardware sweeps were limited by time and equipment access — not all were completed.
Uniform planar arrays, aperture-preserving scaling
Transmitters were modelled as Uniform Planar Arrays rather than point sources. Moving from 3.5 GHz to 7.05 GHz, element count could increase while keeping a comparable physical aperture — so the comparison of lower-frequency coverage vs higher-frequency spatial / array gain stays meaningful.
Conventional · Zero-Forcing · MMSE
Engineering metrics, not just rays
Throughput estimates came from SINR with a capped spectral-efficiency model; hardware runs provided actual 5G link KPIs.
Stochastic launching + deterministic paths
Monte Carlo ray launching
Rays sampled and launched from the transmitter, scored on whether they reached a capture region around a receiver — efficient orientation-dependent statistics.
Deterministic path tracing
Sionna RT's path solver for physically valid paths, up to five interactions.
From geometry to channel representations
- Path gains + propagation delays
- Channel behaviour across frequency
- Discrete channel for baseband sim / emulation
Time-varying channels
Kinematic recomputation
Re-run ray tracing as positions change.
Doppler-only evolution
Reuse a ray solution and evolve its complex path coefficients using Doppler shifts.
The time-dependent channel H(f, t) was evaluated across an OFDM grid (1024 subcarriers, 30 kHz spacing): delay–Doppler spectra, subcarrier and CFR evolution, average gain vs time and frequency.
Doppler-only evolution closely tracked full kinematic recomputation for small per-step displacements. When geometry changes significantly (new occlusions or reflection points), full ray re-tracing is still required.
Reproducing channels on real RF
Simulated channel behaviour was taken beyond software and reproduced on hardware, so controlled channels could be measured without relying on uncontrolled over-the-air propagation.
A reproducible golden baseline
A UE moves around a transmitter on a circular trajectory at 3.5 GHz, establishing a stable baseline before more complex interference and mobility scenarios.
Measured hardware baseline
Stable SINR, RSRP, throughput and BLER showed the simulation-to-hardware chain produced a controlled, repeatable baseline suitable for further experimentation — not full validation of every 5G scenario.
Across the whole stack
Simulation
Signal Processing
Propagation
Hardware
What the internship built up
Wireless Modelling
Antennas & Beamforming
Numerical / Software
Digital Twins
5G / RF Systems
Hardware Integration
Potential extensions, not completed results
The completed hardware results establish a controlled single-cell, single-UE, 3.5 GHz baseline. Due to internship time and intermittent equipment access, these broader experiments were not fully executed: