⚡ electrical engineering

CITYMPC: A Large-Scale Physics-Informed Benchmark and Tool for Generative Complete Multipath Wireless Channel Modeling

The paper introduces CITYMPC, a conditional variational autoencoder framework that generates complete multipath wireless channel parameters from 2D imagery and terrain maps without requiring 3D geometry, achieving ray-tracing-level accuracy across a large-scale, multi-city benchmark dataset.

Original authors: Ashwin Natraj Arun, David R. Nickel, Yaguang Zhang, Yunchou Xing, Jie Chen, Amitava Ghosh, Christopher Brinton, David J. Love, James V. Krogmeier

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ashwin Natraj Arun, David R. Nickel, Yaguang Zhang, Yunchou Xing, Jie Chen, Amitava Ghosh, Christopher Brinton, David J. Love, James V. Krogmeier

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

The Big Picture: Predicting the "Radio Weather" Without a Map

Imagine you are trying to predict how a shout echoes through a city. In the old days, to know exactly how that sound bounces off buildings, hits the ground, and reaches your ear, you needed a massive, detailed 3D blueprint of the entire city. You'd have to simulate every single bounce of the sound wave. This is like Ray Tracing (RT) in wireless technology. It's incredibly accurate, but it's also slow, expensive, and requires you to have the "blueprint" (the 3D map) right in front of you every time you want to make a prediction.

CITYMPC is a new tool that acts like a super-smart weather forecaster for radio waves. Instead of needing the full 3D blueprint, it just looks at a photo taken from the transmitter's perspective and a photo taken from the receiver's perspective, plus a simple topographic map (a flat map showing building heights).

Based only on these pictures and the height map, CITYMPC can guess exactly how the radio signal will travel, bounce, and arrive. It doesn't need the 3D blueprint anymore. It's like looking at a photo of a room and instantly knowing exactly how a ball would bounce off the walls, without needing to measure the room's dimensions first.

The Problem: The "Blueprint" Bottleneck

Current methods for modeling wireless signals (which are crucial for the upcoming 6G internet) fall into two camps:

The Statisticians: They guess based on average rules (e.g., "signals usually bounce this way"). They are fast but often get the specific details of a unique city street wrong.
The Architects (Ray Tracing): They build a perfect 3D simulation. They are accurate but slow and require the 3D blueprint to be available every single time.

The paper argues that for 6G to work with AI, we need a model that is as accurate as the Architects but as fast and flexible as the Statisticians, without needing the 3D blueprint at the moment of prediction.

The Solution: CITYMPC (The "Photo-Reader")

The authors built a machine learning model called CITYMPC. Here is how it works, using an analogy:

The Inputs (The Clues): Imagine you are trying to guess the path of a ball in a complex obstacle course.
- The View: You give the model a "Point-of-View" (PoV) photo from the start (Transmitter) and one from the finish (Receiver). These photos aren't just pretty pictures; they are "super-vision" photos that show not just colors, but also how hard the walls are, how bouncy they are, and how thick they are.
- The Map: You also give it a flat map showing where the tall buildings are.
The Brain (The cVAE): The model is a type of AI called a Conditional Variational Autoencoder. Think of this as a creative artist who has studied millions of radio wave paths. When shown the photos, it doesn't just draw one path; it understands the rules of how waves behave in that specific city.
The Output (The Prediction): It generates a complete list of "Multipath Components" (MPCs). In plain English, it lists every single "echo" the signal will take:
- How long did it take to arrive? (Delay)
- Which way did it come from? (Angle of Arrival)
- Which way did it leave? (Angle of Departure)
- How strong was the signal? (Gain)
- Did it bounce off a wall or go straight through? (Path existence)

The Results: How Good Is It?

The researchers tested this on five major US cities (Austin, Dallas, Denver, Fort Worth, and New York), covering over 427,000 different connection points.

Accuracy: When they compared CITYMPC's guesses to the "Gold Standard" (the slow, detailed 3D Ray Tracing), the guesses were incredibly close.
- Power: The predicted signal strength was off by only about 1.29 decibels on average (a very small error).
- Time: The predicted travel time was off by only 7.25 nanoseconds (a billionth of a second).
Speed: This is the big win. Once the model is trained, it can generate these predictions 1.9 times faster than running the slow 3D simulation, and the model itself runs nearly 10 times faster than the simulation engine.
No 3D Map Needed: Crucially, at the moment of prediction, the model does not need the 3D city map. It only needs the photos and the height map.

The "City Personality" Limitation

The paper also found something interesting about "City Personalities."

If you train the model on New York (tall, dense skyscrapers) and then ask it to predict for Dallas (different building styles), it gets a bit confused. The error goes up.
This means the model learns the specific "personality" of each city's layout. To get the best results, you need to train a specific version of the model for each city, or find a way to teach it how to adapt between cities.

The Gift to the Community

The authors didn't just build the tool; they also built the training data.

They created a massive dataset of 427,397 radio links across five cities, generated using the slow 3D simulation.
They released this dataset and the code for CITYMPC to the public.
Why? So other researchers can use this "benchmark" to build even better AI models for wireless communication without having to spend months creating their own 3D simulations.

Summary

CITYMPC is a new AI tool that looks at photos of a city and a height map to instantly predict exactly how radio signals will bounce around, without needing a complex 3D map. It is fast, accurate, and provides a new standard for researchers to build the wireless networks of the future (6G). It proves that you can learn the physics of radio waves just by looking at the "scenery" of the city.

Technical Summary: CITYMPC

Problem Statement

The fidelity of wireless channel models is a critical bottleneck for the development of sixth-generation (6G) systems, which require rich, site-specific knowledge of propagation environments to support use cases like integrated sensing and communication (ISAC) and AI-native radio access networks. Current channel modeling paradigms face a fundamental trade-off:

Geometry-Based Stochastic Models (GBSM): While computationally efficient, models like 3GPP TR 38.901 are environment-agnostic and fail to capture the distinct characteristics of specific line-of-sight (LoS) and non-line-of-sight (nLoS) links.
Deterministic Ray Tracing (RT): Tools like Sionna RT produce physically accurate multipath components (MPCs) but are computationally expensive, require explicit 3D scene geometry at inference, and generate only a single deterministic realization per link configuration.

Existing learning-based approaches attempt to bridge this gap but suffer from significant limitations: generative models often lack site-specific conditioning; neural ray tracing surrogates still require 3D scene meshes at inference; and radio map methods typically predict aggregate signal quantities rather than complete per-path MPC parameters. Consequently, there is no existing method that can generate a complete, site-specific set of MPC parameters (including Azimuth/Elevation AoD/AoA, delays, and gains) from scene imagery alone without access to 3D geometry during inference.

Methodology

The authors propose CITYMPC, a conditional variational autoencoder (cVAE) framework designed to predict the complete per-path MPC parameter set from point-of-view (PoV) imagery and terrain height maps, eliminating the need for 3D scene meshes at inference.

Data Construction and Conditioning

The system is trained on a large-scale dataset spanning five US urban environments (Austin, Dallas, Fort Worth, Denver, New York), comprising 427,397 Tx-Rx links generated via Sionna RT at 3.5 GHz.

Inputs (Conditioning $c$ ): The model takes as input:
- PoV Image Stacks: 12-channel images from both the Transmitter (Tx) and Receiver (Rx) perspectives. These channels encode RGB, depth, surface normals, and five radio material properties (relative permittivity $\epsilon_r$ , conductivity $\sigma$ , scattering coefficient, cross-polarization coefficient, and thickness) derived from the underlying 3D mesh.
- Global Height Map: A 128 $\times$ 128 macro-scale building layout.
- Coordinates: Raw Cartesian coordinates of Tx and Rx.
Outputs (Target $x$ ): The model predicts a structured set of $L=25$ $L = 25$ path slots, including:
- Path presence mask ( $m_\ell$ ).
- Normalized complex baseband gain ( $\tilde{\alpha}_\ell$ ).
- Normalized excess delay ( $\tilde{\tau}_\ell$ ).
- Unit-vector encodings of AoD and AoA in azimuth and elevation.
- Link-level Time-of-Flight (ToF) time ( $\tau_0$ ) and aggregate received power ( $P_{rx}$ ).

Model Architecture

CITYMPC utilizes a transformer-based cVAE architecture comprising four modules:

Conditioning Encoder (CE): Uses three independent ChannelViT towers to process the global height map, Tx PoV, and Rx PoV stacks, fusing them with coordinate embeddings into a compact scene embedding.
Prior Network (PN): Learns a scene-dependent prior $p_\psi(z|c)$ to capture structured variability across environments, replacing the standard fixed Gaussian prior.
Posterior Encoder (PE): Operates only during training to approximate the posterior $q_\phi(z|x,c)$ by projecting path parameters into tokens that cross-attend to scene tokens.
Decoder (DEC): Generates the output using $L$ learned "slot queries" initialized by the latent variable $z$ . These slots internalize the ordinal structure of MPCs (sorted by power). The decoder employs physics-informed routing, where separate cross-attention layers attend exclusively to Tx or Rx scene tokens to predict AoD and AoA, respectively.

Training Strategy

The model is trained by maximizing the Evidence Lower Bound (ELBO). A key innovation is the use of Kendall uncertainty weighting for multi-task learning. The total loss is a weighted sum of seven task-specific losses (path presence, ToF, received power, AoD/AoA directions, excess delay, and complex gain), where the weights are learned via per-task noise parameters ( $\sigma_k$ ) to automatically balance tasks of varying difficulty.

Key Contributions

Complete MPC Characterization without 3D Geometry: CITYMPC is the first work to predict the full per-path parameter set (including elevation angles for both AoD and AoA) conditioned solely on scene imagery and height maps, removing the inference-time dependency on 3D meshes.
Generative Matching of Ray Tracing: The framework generates physically valid channel realizations that match ray-traced ground truth across all predicted parameters, achieving a received power Mean Absolute Error (MAE) of 1.29 dB and a ToF MAE of 7.25 ns across the evaluated cities.
Large-Scale Benchmark and Dataset: The authors release a reproducible benchmark and a large-scale multi-city ray-traced dataset (427,397 links) to accelerate future research in scene-conditioned channel modeling.
Efficient Inference: Once trained, CITYMPC generates channel realizations at inference without ray tracing, offering a 1.90 $\times$ speedup in end-to-end latency and a 9.88 $\times$ speedup in model forward pass time compared to Sionna RT.

Experimental Results

Evaluated across five diverse US cities, CITYMPC outperforms a strong MLP baseline (which uses ResNet-18 and flat MLPs) on all metrics except path presence F1, where performance is comparable.

Accuracy: CITYMPC achieves lower MAE than the baseline in received power, ToF, average delay, and all four angular dimensions (AoD/AoA azimuth and elevation). The ToF error reduction is particularly significant (1.5 $\times$ to 3.2 $\times$ lower than the baseline).
Distribution Fidelity: Empirical Cumulative Distribution Functions (CDFs) of generated parameters closely overlap with ray-traced ground truth, confirming the model captures the full marginal channel statistics.
Cross-City Generalization: The study reveals significant distribution shifts between cities. Models trained on one city degrade when applied to others (e.g., received power MAE increases from <2.1 dB to >7 dB), indicating that per-city training is currently necessary for high fidelity.
Qualitative Performance: Generated channel realizations exhibit stochastic diversity (different path configurations for the same link) while maintaining aggregate statistics consistent with ground truth, effectively reproducing spatial power maps including street canyon effects and building shadows.

Significance and Limitations

The paper positions CITYMPC as a practical tool for researchers requiring large-scale, site-specific channel data without the computational burden of repeated ray tracing. It establishes a strong in-distribution baseline for future multi-city generalization research.

The authors explicitly acknowledge several limitations:

Sim-to-Real Gap: The model has not been validated against real-world channel measurements; the gap between simulation and reality remains an open challenge.
Cross-City Transfer: Performance degrades significantly across cities due to distribution shifts, necessitating per-city training.
Scope: The current model targets outdoor urban macro-cell environments at 3.5 GHz with omnidirectional antennas; extension to indoor scenarios, mmWave frequencies, and directional antennas is not yet addressed.
Training Stability: The use of Kendall uncertainty weighting can lead to training divergence (task abandonment), which the authors plan to address in future work.