5G NR Neural Receiver Dashboard

DeepRx-style receiver vs LS+LMMSE baseline on a Sionna TDL-C link. The dashboard shows where neural wins, where classical holds, and what is still outside the evidence boundary.

Landing page README Business case Technical brief

Technical Decision Summary

What does the AI-PHY experiment prove?

BER @ 5 dB

0.062

classical 0.122

BER @ 10 dB

0.014

classical 0.026

BLER @ 12.5 dB

0.585

classical 0.938

ONNX parity

PASS

max_diff=1.38e-05

Decision field	Evidence
Experiment	DeepRx-style neural receiver vs classical LS+LMMSE baseline
Link model	Sionna 5G NR, TDL-C, CP-OFDM, QPSK, SISO
Neural receiver size	373,154 trainable parameters
Main finding	Neural receiver improves BER around 5-10 dB
Moderate-SNR gain	about 2-3 dB effective BER advantage
BLER finding	12.5 dB neural BLER 0.585 vs classical 0.938
High-SNR boundary	Classical baseline remains competitive above 15 dB
Export evidence	ONNX parity PASS, max_diff=1.38e-05
Deployment boundary	Simulated link evidence only, not live RAN or SDR deployment

Problem -> What I Built -> What I Found -> What I Would Validate Next

The AI-PHY story in one pass

Problem

Classical 5G receivers separate channel estimation, equalization, and demapping. Neural receivers test whether those functions can be learned jointly under realistic channel models.

What I Built

I built a DeepRx-style residual CNN receiver on a Sionna-modeled 5G NR TDL-C link and compared it against an LS+LMMSE+soft-demap baseline on the same channel conditions.

What I Found

The neural receiver improves BER by about 2x around 5-10 dB and shows a 12.5 dB BLER improvement from 0.938 to 0.585. At high SNR, the classical baseline remains competitive.

What I Would Validate Next

I would test higher-order modulation, MIMO, LDPC-coded BLER, mobility/Doppler sweeps, channel mismatch, and SDR/hardware-loop integration before making deployment claims.

Visual Evidence

Curves and distributions from committed artifacts

BER/BLER comparison

Neural receiver gains are clearest in the moderate-SNR region; high-SNR limits remain visible.

Open source artifact

LLR distribution comparison

The LLR distribution shows receiver confidence shape at a measured operating point.

Open source artifact

Classical baseline

The classical LS+LMMSE path is a strong reference receiver, not a straw man.

Open source artifact

ONNX parity

ONNX parity passes against ONNXRuntime. This proves export correctness, not production deployment readiness.

Open parity JSON

Where Neural Wins / Where Classical Holds

The measured result has boundaries

SNR region	Neural receiver behavior	Classical baseline behavior	Engineering interpretation	Boundary
Low SNR, -5 to 0 dB	Improves BER, but frames still fail heavily.	Also frame-error limited in the uncoded QPSK setup.	Useful signal, not an operating-point victory.	No coded BLER or scheduler claim.
Moderate SNR, 5 to 12.5 dB	Shows the strongest practical advantage and the clearest BLER improvement.	Falls behind on BER/BLER under the same TDL-C conditions.	This is the strongest measured neural-receiver region.	Measured only for TDL-C/QPSK/SISO.
High SNR, 15 to 20 dB	No longer dominates; BER can trail the classical chain.	Becomes competitive as channel estimation improves.	The result is not 'neural always wins.'	High-SNR limitation remains visible.

Receiver Architecture Comparison

Classical blocks versus learned joint receiver

Stage	Classical path	Neural path	Why it matters
Channel estimation	LS pilot channel estimation	Learned jointly from received grid + pilot mask	Neural path can exploit local structure beyond explicit LS estimates.
Equalization	LMMSE equalization	Implicitly learned through CNN residual blocks	Moves receiver design from separated blocks to learned joint inference.
Demapping	Max-log soft demapping	CNN outputs LLRs for data bits	Keeps the output compatible with a downstream soft-decoder path.
Training/tuning	No training; model assumptions are explicit	End-to-end BCE through Sionna channel, random SNR per minibatch	Learned receiver needs data/model discipline and validation across mismatch.
Export/deployment artifact	Algorithmic Python baseline	ONNX opset export with parity check	ONNX proves export correctness, not production readiness.
Failure/validation risk	Can be strong at high SNR	Can overfit channel/modulation assumptions	Both paths need explicit operating boundaries.

Engineering Practices That Matter

The practices that make the comparison useful

Signal	Evidence
deterministic pilot RNG bug	Found and fixed by controlling `sionna.phy.config.seed` before ResourceGrid creation.
Fair comparison	Same TDL-C/QPSK/SISO link conditions are used for the classical and neural receiver paths.
ONNX parity	PASS with max_diff=1.38e-05.
Test suite	31/31 tests passing in the recorded evidence.
Training cost	10.4 min / 100,000 steps on RTX 5090, batch=64.
Committed artifacts	CSV, SVG, JSON, dashboard, and reproducibility workflow are committed.

What This Is

The concrete system layers in this repo

Layer	What it does
AI-PHY receiver experiment	Tests learned receiver behavior against a classical baseline.
Sionna-modeled 5G NR link	Uses TDL-C/QPSK/CP-OFDM/SISO link conditions.
Classical receiver baseline	LS pilot estimation, LMMSE equalization, max-log soft demapping.
DeepRx-style neural receiver	Residual CNN that learns joint receiver behavior.
BER/BLER evidence pack	Measured curves and tables across SNR.
ONNX parity export check	Verifies export correctness against ONNXRuntime.
Static dashboard	Packages measured evidence, boundaries, and artifacts for review.

What This Is Not

Boundaries that keep the evidence honest

Claim	What is true instead	Limit it protects
Live 5G deployment	No, this is simulated link evidence.	Prevents production overclaiming.
SDR validated receiver	No, no hardware-loop validation yet.	Keeps hardware claims honest.
O-RAN/gNB integration	No, no RAN integration is claimed.	Separates link evidence from RAN integration.
NVIDIA Aerial integration	No, this uses Sionna, not Aerial.	Avoids vendor integration theater.
MIMO receiver	No, this is SISO unless implemented otherwise.	Keeps antenna scope clear.
Higher-order QAM / LDPC-coded system	No, QPSK and current uncoded scope only as implemented.	Avoids unmeasured PHY claims.
Production-ready AI-RAN component	No, this is a research-grade evidence pack.	Preserves deployment boundary.

Evidence vs Boundary

What is demonstrated, and what is deliberately not claimed

Evidence demonstrated

Sionna-modeled 5G NR TDL-C link
Classical LS+LMMSE baseline
DeepRx-style neural receiver
BER/BLER comparison
ONNX export parity
Deterministic pilot fix
Reproducible artifacts and tests

Boundary preserved

No live 5G network
No SDR/hardware-loop validation
No O-RAN/gNB integration
No NVIDIA Aerial integration
No MIMO/higher-order QAM claim
No LDPC-coded system claim
No production AI-RAN deployment claim

Technical Workflow

How the evidence is produced

Stage	Output
Sionna link config	Shared TDL-C/QPSK/SISO link setup.
Classical baseline sweep	`reports/bler_classical.csv` and plot.
Neural receiver training	`reports/training_log.json`.
Head-to-head BER/BLER comparison	`reports/bler_comparison.csv` and plot.
ONNX export	Model export when checkpoint artifacts are available.
Parity check	`reports/onnx_parity_test.json`.
Dashboard generation	`reports/index.html` and `reports/dashboard.html`.
Tests / verify	pytest, ruff, and `make verify`.

Measured BER / BLER

Same channel, same SNR points, two receiver paths

SNR	BER classical	BER neural	BLER classical	BLER neural
-5.0 dB	0.40599	0.36102	1.000	1.000
-2.5 dB	0.35186	0.27311	1.000	1.000
+0.0 dB	0.28032	0.18828	1.000	1.000
+2.5 dB	0.19960	0.11652	1.000	1.000
+5.0 dB	0.12213	0.06173	1.000	1.000
+7.5 dB	0.06147	0.02799	1.000	1.000
+10.0 dB	0.02638	0.01418	1.000	0.983
+12.5 dB	0.01038	0.00905	0.938	0.585
+15.0 dB	0.00471	0.00805	0.460	0.314
+17.5 dB	0.00234	0.00737	0.224	0.225
+20.0 dB	0.00123	0.00725	0.127	0.197

Evidence Artifacts

Files behind the dashboard

Artifact	Link	What it proves
BER/BLER CSV	reports/bler_comparison.csv	Raw measured comparison table.
BER/BLER plot	reports/bler_comparison.svg	Visual error-rate comparison.
LLR plot	reports/llr_distribution_comparison.svg	Receiver confidence distribution.
ONNX parity	reports/onnx_parity_test.json	Export correctness check.
Training log	reports/training_log.json	Training config and convergence trace.