CityMPC: Environment-Aware Multi-Path Channel Dataset

CityMPC is a large-scale synthetic dataset of multi-path wireless channels for 5 US cities (Austin TX, New York NY, Dallas TX, Fort Worth TX, Denver CO), generated via Sionna RT ray tracing on DeepMIMO v4 building geometries at 3.5 GHz.

Each sample pairs:

• A full multi-path channel description — up to 25 paths per link with complex gains, delays, and angles of departure/arrival.
• An environment-aware POV rendering of the propagation scene — RGB image, depth map, surface normals, RF material maps (per-channel permittivity, conductivity, roughness, scattering), and global building height map.

The dataset is intended for training and evaluating environment-aware channel-prediction models that replace ray tracing at inference time. It accompanies a NeurIPS 2026 Evaluations & Datasets Track submission. The full code release is at the citympc repository (link withheld for double-blind).

---

What's in this release (~700 GB)

Tier	Cities	Per-city contents	Total
Sionna scenes (all 20 DeepMIMO US cities)	20 cities	<city>.tar.gz (scene.xml + meshes/*.ply)	244 KB
Mini bundle (all 5)	Austin, NY, Dallas, Fort Worth, Denver	train_2000.{npz,pt}, val_500.{npz,pt}, test_500.{npz,pt}, norm_stats.json	~23 GB
Full bundle (all 5)	Austin, NY, Dallas, Fort Worth, Denver	train.h5, val.h5, test.h5, train.npz, val.npz, test.npz	~680 GB

Per-city full sizes: Austin ~127 GB, NY ~101 GB, Dallas ~156 GB, Fort Worth ~134 GB, Denver ~162 GB.

---

Sample for reviewers (small subset, ~4.6 GB)

Per NeurIPS Datasets-track guidance, large datasets should provide a smaller sample. The per-city mini bundle is exactly that: 2000 training links + 500 validation + 500 test, baked as PyTorch tensors with normalisation statistics — same schema as the full data, ~3,000× smaller per city.

# Smallest single download — Austin mini bundle, ~4.6 GB:
hf download neurips2026citympc/CityMPC \
  --repo-type dataset \
  --include "manifests/city_10_austin_3p5_s/{train_2000,val_500,test_500}.* manifests/city_10_austin_3p5_s/norm_stats.json" \
  --local-dir .

How the sample was created: deterministic seed-42 random subset of links from the filtered training/validation/test splits (size N=2000/500/500), then the same bake_dataset.py pipeline used for the full splits stacks normalised tensors into .pt files with the identical schema as train.h5 / val.h5 / test.h5. The same procedure is available for all 5 cities.

Quick start (full dataset)

pip install huggingface_hub
hf download neurips2026citympc/CityMPC --repo-type dataset --include "manifests/city_10_austin_3p5_s/*" --local-dir .

For a one-line download + verify experience, use the citympc scripts/download_hf.py helper (see the citympc repo).

Croissant metadata

A valid Croissant 1.1 metadata file is at the dataset repo root: croissant.json. It contains all FileObject SHA-256 hashes, the RecordSet schema (channel sample fields + norm-stats fields), Responsible AI properties, and provenance. Validated against mlcroissant (0 errors, 0 warnings).

---

Data schema

Each HDF5 file (train.h5, val.h5, test.h5) and each baked tensor file (*_2000.pt, *_500.pt) shares the same record schema. One record = one TX-RX link. Field shapes use N for the number of links in a split and 25 for the maximum number of multi-path components.

Field	Shape	Description
excess_delay	(N, 25)	Excess delay (seconds) of each path relative to first arrival τ₀. Zero-padded for inactive paths.
path_presence	(N, 25)	Binary indicator of active paths (0.0 or 1.0).
path_coeff	(N, 25, 2)	Complex path coefficient (real, imag) per path.
path_dirs	(N, 25, 6)	Departure + arrival 3D unit direction vectors [AoD_x, AoD_y, AoD_z, AoA_x, AoA_y, AoA_z].
path_loss_db	(N,)	Total received path loss (dB) aggregated over all active paths.
first_arrival	(N,)	Absolute propagation delay (seconds) of the first-arriving path.
rx_pov	(N, 12, 128, 128)	Receiver POV image stack: RGB (3) + depth (1) + normals XYZ (3) + RF material props (5). Float32.
tx_pov	(N, 12, 128, 128)	Transmitter POV image stack — same 12-channel layout as rx_pov.
global_map	(N, 1, 128, 128)	Global building height map centred on the TX-RX midpoint.
scalars	(N, 6)	TX + RX 3D world coordinates [tx_x, tx_y, tx_z, rx_x, rx_y, rx_z] in metres.

norm_stats.json (per city) contains:

Key	Description
first_arrival.mean_log_ns	Mean of log(first_arrival × 1e9) across training-split links.
first_arrival.std_log_ns	Std of log(first_arrival × 1e9) across training-split links.
path_loss_db.mean	Mean of path_loss_db across training-split links.
path_loss_db.std	Std of path_loss_db across training-split links.
n_links_valid, n_links_total	Link counts.

---

Provenance

• Geometry source: DeepMIMO v4 — building footprints (OpenStreetMap-derived) and TX/RX user grids for 20 US city scenarios at 3.5 GHz.
• Ray tracing: Sionna RT — full TX-RX path enumeration with up to 25 paths per link, complex coefficients, delays, AoD/AoA, computed for all link pairs.
• POV rendering: Mitsuba 3 — 12-channel image stacks rendered from both TX and RX viewpoints, plus a global height map per link.
• Filtering & splits: link-level power filter + path-level prune + L_max truncation; 80/10/10 train/val/test split via deterministic seed; norm_stats.json recomputed on the filtered training set.
• Full pipeline reproducibility: see PIPELINE.md in the citympc code release.

---

Responsible AI

Synthetic data

This dataset is fully synthetic. It contains only ray-traced wireless channel parameters and 3D environment renderings derived from publicly available building geometry models. No personal or sensitive information is present.

Limitations

• Frequency: 3.5 GHz only. Results may not generalise to mmWave (28+ GHz), sub-GHz, or other bands without retraining.
• Geography: 5 US cities. Other cities, non-US urban layouts, and rural/suburban environments are out of distribution.
• Static environments: No dynamic objects (vehicles, pedestrians, foliage), no weather, no atmospheric effects.
• Geometry fidelity: DeepMIMO v4 building footprints are extruded uniformly from OpenStreetMap data and may not match real building heights or shapes.
• Source code: The generation pipeline is open-source; see the citympc repository for full reproducibility.

Biases

• All cities are mid-to-large US urban centres with dense building layouts; receiver placements follow DeepMIMO v4 user grids and may oversample regular street patterns.
• Material RF properties are assigned from a fixed per-material lookup; no spatial variation within a material class.

Intended uses

• Training and benchmarking ML models for environment-aware multi-path channel prediction.
• Channel charting, site-specific radio resource management, digital-twin radio simulation.
• Evaluation of generative models that consume image/geometric scene representations.

Out-of-scope uses

• Operational deployment of trained models without site-specific validation.
• Generalisation claims beyond the cities, frequency, and conditions covered.

Social impact

Open release of the dataset and the generation toolchain reduces reliance on proprietary or computationally expensive ray tracers in 5G/6G research, and supports reproducibility in environment-aware wireless communications.

---

File integrity (SHA-256)

File	SHA-256
manifests/city_10_austin_3p5_s/train.h5	89830838f697674d993390862e8ae99adf1ebc16de0a6db9ded898ae9b32245c
manifests/city_10_austin_3p5_s/val.h5	2a89a446355cebb5a347891a5a69b2cd8a67455e09c0efa53020638d9ef7ae60
manifests/city_10_austin_3p5_s/test.h5	551cffd696a7a259ac1f9b318479880f886b08722c8fcca535e1ec39d050dda6
manifests/city_10_austin_3p5_s/norm_stats.json	5a9f73b8692464b75f6fb54a228fb7fb51c58e837be0a7c4e01608c03ed8b8e1
manifests/city_10_austin_3p5_s/train_2000.pt	3c996049dc3ca61c90c20f72366f18ad1fbceca1b5717b754558f0af9af6dd45
manifests/city_0_newyork_3p5_s/train.h5	89d34d6d70b16eb226e5b0e7a0bc459115d03eb45112e246e742a957664f7035
manifests/city_0_newyork_3p5_s/val.h5	a75efb4d9e9c2a162750c6b35f6848639c8b4e1a33ff9bff3683d4ff5646d567
manifests/city_0_newyork_3p5_s/test.h5	00872e53d95895dba15dacfb6a675046db3b84166ee3a0c516eb668ac881fafe
manifests/city_0_newyork_3p5_s/norm_stats.json	f74a82ca84aa3cd5577bcb0f29f37c386268c787d7c3201e4e09b8154ea4a1b6
manifests/city_0_newyork_3p5_s/train_2000.pt	028e16a072ba6fe9589ce43ae8b02a4d084e3e125ab9371c6e9ef22dc2b8c7a6
manifests/city_8_dallas_3p5_s/train.h5	ec9ff0c5df4feb3549bff3ef6341a1949363c40ce624976d9b272db73db71ef5
manifests/city_8_dallas_3p5_s/val.h5	6981d2b15bb71a2ee8f74f11210f8d891efed1b177c2922c54d267fd9a0d4bad
manifests/city_8_dallas_3p5_s/test.h5	9876d5b8a65ee42f99b084a44f22f4e3be5be5aee6a4742c9783b144437daa33
manifests/city_8_dallas_3p5_s/norm_stats.json	f4bf3a71288e74d364f463ee85ef0377bd43b8faccfcf5b7fbda5afcc5fc7294
manifests/city_8_dallas_3p5_s/train_2000.pt	670b0266756b308e56991bb54f9581adbef7fec51525b74217fd0d4b4c7638ca
manifests/city_12_fortworth_3p5_s/train.h5	d7b0a591f60f10b0000a45c8d18320b78d0d8264cdb9d03401bfcab1288aaae4
manifests/city_12_fortworth_3p5_s/val.h5	d13e41ae0fe852e58a01f692c55b7c6a89e8143d82430a84c89b83657c9d4623
manifests/city_12_fortworth_3p5_s/test.h5	cde6c439f7ec918d8fa77ce524e8d096737b5b5824bcceb851919e1522ac5681
manifests/city_12_fortworth_3p5_s/norm_stats.json	d400be7f4fcc59cda63b0eafe88bd3330a8a3866c19a9ac6ae59c86c97fb3ef5
manifests/city_12_fortworth_3p5_s/train_2000.pt	ce7a01b737021521f215162d904cae70c6079b3951bbbf3bbffcdbb52e4fba7d
manifests/city_18_denver_3p5_s/train.h5	0de729b9165b97f67835c04745b5e604d495fcd3c1c3fa27e8203460905e58cd
manifests/city_18_denver_3p5_s/val.h5	d3eb526e35fe4c57eace7d4f8628c496f87ce7f8ae17f83d625458fb80daa8ca
manifests/city_18_denver_3p5_s/test.h5	e4cdd747079b0027bb44533395881fa3f931fa14182f0fe12e4567f95d925c2f
manifests/city_18_denver_3p5_s/norm_stats.json	6d9fd65de6bccb346b29204726392fd710276a6a2e763245c4c80794eb96cd7d
manifests/city_18_denver_3p5_s/train_2000.pt	a08371d2ccc1d8cc5a5494880bcb388db155ff7c5bf589fd163e74e94363325c

---

Citation

@inproceedings{citympc2026,
  title     = {CityMPC: Environment-Aware Multi-Path Channel Prediction via Conditional Generation},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026},
  note      = {Evaluations \& Datasets Track}
}

License

Apache 2.0.