A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization

arXiv:2609.08482 · eess.SP · Submitted 2026-09-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization".

Dev: Here's the gist of the paper. Wireless cellular networks provide critical infrastructure for communication, transportation and industry, making reliable connectivity essential to modern society <ref:2609.08482#pg2>.

Rosa: Here we introduce ChaRT, a foundation model that learns transferable radio representations from measurement reports routinely generated by operational cellular networks <ref:2609.08482#pg4>.

Dev: With a single set of weights, ChaRT reconstructs radio environments in unseen cities and transfers to new-site prediction, radio map construction and network parameter tuning <ref:2609.08482#pg4>.

Rosa: The Core Problem Addressed Wireless cellular networks require accurate models of the radio environment shaped jointly by network infrastructure and their surroundings, yet city-scale radio environments remain difficult to capture <ref:2609.08482#pg2>.

Title and authors: Rosa: So we're diving into this paper called "A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization." It’s written by Qin, Pu, Dong, Peng, Xue, Shen, Yan and Chang.

Dev: Yeah. The title tells you a lot about what they're trying to achieve: building a foundation model for planning and optimizing wireless networks on a large scale. It sounds like it's aiming to solve that big problem where city-scale radio environments are just too complex for the current tools we have.

Taro: I wonder how they plan to handle the sheer variety of those environments across different cities and deployment types, because that seems like a huge generalization problem.

Rosa: Exactly. The paper says this model learns transferable radio representations from measurement reports that are routinely generated by operational cellular networks, which is a pretty big shift from what we usually train models on. It’s using real data collected from actual city operations to learn the underlying physics of how signals move around complex urban areas.

Dev: That’s the core idea—transferring knowledge across cities without retraining for every new city or deployment scenario. They claim this single set of weights can reconstruct radio environments in unseen cities and transfer to things like new-site prediction and network parameter tuning.

Taro: But how do they handle the fundamental challenge of high-dimensional, time-varying channel state information, which is what makes these wireless problems so hard to model?

Rosa: Well, they tackle that by pretraining ChaRT using context-aware masked beam modelling and self distillation. They break the training objective down into four parts: beam reconstruction, representation alignment, expert routing, and a missing-beam penalty.

Dev: The student teacher framework they use combines beam-level reconstruction with representation-level alignment across physically consistent views of the data. It seems like they are trying to build a very robust way to learn what a radio signal actually looks like in context.

Taro: If it’s learning from measurement reports, does that mean it can handle the messiness of real-world operational data, or is that just an illusion? What happens when the world misbehaves outside the lab environment?

Rosa: The paper claims that operational measurement reports actually contain transferable network-scale propagation structure. They show that a single set of pretrained weights generalizes without retraining to cities, frequency bands, or deployment scenarios they didn't see during pretraining.

Dev: That’s significant because it suggests we might not need dedicated site descriptions for every new place; the model learns the general rules of propagation from what operators are already collecting. The cross-city new-site evaluation shows an R squared of zero point six two, which is better than other models they compared, like TabPFN v2 with an R squared of zero point two four or XGBoost with negative values.

Taro: So it’s not just about getting a plausible average signal level; ChaRT is doing something more nuanced by inferring cell-specific coverage geometry, which is the basis for their zero-shot network planning tasks. That sounds like it’s moving past simple estimation towards real structural understanding.

Title and authors: Rosa: Right, and that structural understanding feeds into five downstream applications they've attached using lightweight task-specific heads to the frozen backbone. These include user localization, beam prediction, propagation scenario classification, SINR estimation, and beamspace clustering.

Dev: That’s a lot of tasks derived from one core model. The idea is that you train the backbone once on this massive dataset—over one billion samples and eighteen point two billion beam-level observations from just one city—and then you just plug in these small heads for specific jobs <ref:2609.08482#pg3,over one billion samples and 18.2 billion beam-level observations>.

Taro: If we look at the setup, they mention that cross-city transfer is evaluated by applying a City A trained head directly to the target city without any adaptation required for that new city. That’s a strong claim about true generalization across geography.

Rosa: They also use a one percent setting of labeled data as their primary few-shot adaptation setting for these downstream tasks <ref:2609.08482#pg1>. It suggests that even when you need some small fine-tuning, the foundation model has already learned so much about the physics that it adapts quickly.

Dev: The scaling results are interesting too. As the model size and pretraining data volume increase, like going from ten million to one billion measurement reports, the mean absolute error stays relatively stable at around six point three eight dB, with a total improvement of about zero point nine one dB across that grid of settings.

Taro: That saturation point suggests that just dumping more same-city data isn't the only way forward; they imply diversity across cities, frequency bands, and network architectures is what actually matters for pushing performance further.

Rosa: That’s the future direction they suggest—extending pretraining to more cities or frequency bands might expose the model to a broader range of propagation conditions than just increasing sampling density within one city.

Dev: And you also mentioned that calibrated uncertainty and explicit physical constraints will be necessary later on to figure out when those predictions can actually be trusted, which is a necessary safety step for any real-world deployment.

Taro: So, if we look at the overall picture of this paper, "A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization," it’s about moving beyond the need for massive site descriptions by learning universal propagation rules from operational data collected in one city to help plan and optimize networks everywhere.

Rosa: Exactly. It shows that we can learn a transferable substrate for the city-scale radio environment using data operators already have, which opens up practical paths toward foundation models for network planning, operation, and optimization.

Dev: It’s a lot of learning about how complex physical systems behave when you try to model them with data-driven methods rather than purely physics-based tools.

Taro: I'm curious to see if this approach holds up when we introduce more dynamic elements into the network, like human activity or rapid changes in infrastructure.

Rosa: That’s where the real test will be, pushing it beyond just static city environments to handle those evolving conditions.

The paper's summary: Rosa: So, to recap, this paper is about taking something that operators already collect—measurement reports from cellular networks—and using that to build one big model called ChaRT for planning and optimizing whole cities.

Dev: Right. It’s moving away from needing specific site descriptions for every single spot in a city, which is a huge headache because city radio environments are just too complex to map out manually or with old tools.

Rosa: Exactly. The big point they’re making is that this single model, once trained on these reports, can predict how signals will behave in totally new cities or under different network setups without needing any extra training for those new places.

Dev: That generalization capability is what caught my attention from a controls angle. We usually need a whole new set of parameters or a complete re-tuning when you move to a different deployment area, but this model seems to handle that transfer automatically through its learned representations.

Rosa: It’s not just guessing the signal level; the paper says ChaRT learns how those measurements connect to the overall network structure across different areas. They show it can reconstruct the radio environment in unseen cities and then use that knowledge for things like predicting where you need a new cell or tuning existing antenna parameters.

Dev: And they back up that claim with numbers, which is always good. They report an R squared value of point six two in a cross-city evaluation, which beats other models they tested like TabPFN v2, which got a point two four, and XGBoost got negative values. That's a pretty solid comparison for generalization.

Rosa: It really shows that the structure of propagation across a city is something you can learn from real operational data instead of just relying on physics-based tools that require tons of detailed input.

Dev: If we think about the engineering side, this means we could potentially skip those weeks or months spent gathering detailed site descriptions for every new deployment area and jump straight to optimization based on a pre-trained model.

Rosa: That’s the world it opens up for network planning—it makes using data that operators already collect a practical path toward building models that work everywhere, not just where they were originally measured.

The paper's improvements: Rosa: So, looking at what they suggest improving, the main idea is that we can do several things without needing brand new measurements for every single deployment or city we want to plan for.

Dev: Right. They focus on beam-level reconstruction first, which means you get a radio map. That lets you monitor coverage and even do counterfactual evaluations, meaning you can test what *would* happen if the network was set up differently before you even put anything in the ground.

Rosa: That’s huge because it means we don't need dedicated measurement campaigns just to see how a city looks under different conditions.

Dev: Then there's new-site deployment prediction, which they say lets operators forecast coverage for sites they haven't even built yet, using only the context of the surrounding network.

Rosa: That’s powerful for planning because you can make decisions about where to put infrastructure based on what the existing signals are already doing around it.

Dev: They also have antenna parameter tuning prediction, so operators can test different antenna settings—like mechanical azimuth or transmit power—and predict the resulting coverage using just that new setting and the surrounding network context.

Rosa: It’s like a simulation tool that runs on real-world data, which is much more useful than just a pure physics model that doesn't know about actual urban clutter.

Dev: And for those tasks where you only have a tiny bit of labeled data, they claim user localization and beam prediction accuracy can get really good with just one percent of those labels.

Rosa: That’s because the foundation model has already learned the directional structure of signals so well that it doesn't need a massive dataset to learn basic things like where a user is located.

Dev: Scenario classification also gets high macro F1 scores, say eighty-three point zero percent with just one percent labeled data, which shows it's effective even in those very data scarce situations.

Rosa: It really changes how we approach network optimization because instead of needing perfect, complete datasets for every scenario, we can use this model to get a really solid starting point quickly.

Dev: But they do have limitations they mention—it still needs calibrated uncertainty and explicit physical constraints to tell you when a prediction is actually safe to trust in the real world.

Rosa: Exactly. It’s not magic; it's a tool that gives us strong predictions, but we still need those safety checks before we deploy anything based on its output.

Conclusion: Tom: So we’re wrapping up this look at "A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization." The main thing is that this model learns to map operational data from one city onto another so you don't have to train a whole new system for every single place.

Rosa: That’s right. It shows that we can use existing measurement reports from real cities as a foundation to build tools for planning and optimizing networks everywhere.

Dev: I think the implications are that this could make network planning much faster, skipping the huge effort of gathering detailed site descriptions for every new deployment area.

Taro: It means autonomy researchers could test things in new environments more gracefully because the system already understands how signals behave in a broad context before it even sees a specific setup.

Rosa: The caveat is that they still need calibrated uncertainty and explicit physical constraints to know when their predictions are actually safe to use in the real world.

Dev: That’s true, you can’t trust an AI prediction blindly without knowing the confidence level, so those safety checks are crucial for any controls engineer.

Taro: I just wonder how well this model handles those sudden, unpredictable changes that happen in a city—like a temporary construction site or a new building going up overnight.

Rosa: That’s what we need to see next; whether it stays robust when the world misbehaves outside the perfectly controlled lab environment.

Dev: Well, for now, this paper gives us a really strong starting point for using operator data in network optimization tasks.

Taro: It definitely moves us toward building systems that can adapt and recover without needing a complete overhaul every time they encounter something unexpected.

Shenzhen Loop Area Institute · Shenzhen Research Institute of Big Data Shenzhen China The Chinese University of Hong Kong Shenzhen Sun Yat-Sen University Huawei Technologies

eess.SP

Submitted: 2026-09-08

Updated: 2026-11-04

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 100/100

The gist: Wireless cellular networks provide critical infrastructure for communication, transportation and industry, making reliable connectivity essential to modern society <ref:2609.08482#pg2>.

Key concepts

ChaRT
A foundation model trained on operational cellular measurement reports. It learns transferable radio representations from these reports to generalize knowledge across different cities, frequency bands, and deployment scenarios without needing retraining.
Transferable Radio Representations
The core learned features that ChaRT extracts from measurement data. These representations capture the underlying physical structure of the radio environment, allowing the model to apply this understanding effectively to new locations or situations where it was not explicitly trained.
Context-Aware Masked Beam Modelling
A pretraining technique used to build ChaRT. It involves reconstructing missing beam information by considering surrounding network context and using self-distillation between different views of the radio environment, ensuring the learned representations are physically consistent.

Terminology

Summary

Wireless cellular networks provide critical infrastructure for communication, transportation and industry, making reliable connectivity essential to modern society <ref:2609. p. 0

Where the paper says it

cn, wpu@sribd. cn Abstract Wireless cellular networks provide critical infrastructure for communication, transportation and industry, making reliable connectivity essential to modern society.

08482#pg2>. Here we introduce ChaRT, a foundation model that learns transferable radio representations from measurement reports routinely generated by operational cellular networks <ref:2609. p. 0
Where the paper says it

Physics-based tools require detailed site descriptions and computation, whereas task-specific models need dedicated measurements and transfer poorly across deployments. Here we introduce ChaRT, a foundation model that learns transferable radio representations from measurement reports routinely generated by operational cellular networks.

08482#pg4>. With a single set of weights, ChaRT reconstructs radio environments in unseen cities and transfers to new-site prediction, radio map construction and network parameter tuning <ref:2609. p. 0
Where the paper says it

2 billion beam-level observations from 3,503 cells in one city. With a single set of weights, ChaRT reconstructs radio environments in unseen cities and transfers to new-site prediction, radio map construction and network parameter tuning.

08482#pg4>.

The Core Problem Addressed

Wireless cellular networks require accurate models of the radio environment shaped jointly by network infrastructure and their surroundings, yet city-scale radio environments remain difficult to capture <ref:2609.08482#pg2>. Classical operator practices cannot scale to this coupled, city-wide complexity <ref:2609. p. 1

Where the paper says it

Wireless-system behaviour is jointly shaped by physical propagation, engineering configuration and human activity, and evolves as these factors change. Classical operator practices, such as drive-test campaigns, trial-and-error parameter tuning and offline ray-tracing simulation, cannot scale to this coupled, city-wide complexity.

08482#pg2>. Physics-based tools require detailed site descriptions and computation, whereas task-specific models need dedicated measurements and transfer poorly across deployments <ref:2609. p. 0
Where the paper says it

Such models underpin base-station deployment, network operation and parameter optimization, yet city-scale radio environments remain difficult to capture. Physics-based tools require detailed site descriptions and computation, whereas task-specific models need dedicated measurements and transfer poorly across deployments.

08482#pg2>. The core obstacle for existing wireless foundation models is dimensional scaling of CSI, which is high-dimensional and time-varying <ref:2609. p. 1
Where the paper says it

On the data axis, the core obstacle is dimensional scaling. CSI is a high-dimensional, time-varying quantity: a single-cell base station may carry hundreds of antenna elements, yielding a channel matrix of comparable dimension per user, and its values fluctuate rapidly with user movement and environmental change.

08482#pg4>. Furthermore, single-link CSI measurements lack information about the surrounding network, which requires joint observations across multiple cells <ref:2609. p. 1
Where the paper says it

A single-link CSI measurement is by construction an observation of one transmitter-receiver pair in isolation, encoding the link’s multipath structure produced by radio signals arriving along reflected, diffracted and scattered paths, but carrying no information about the surrounding network. Quantities that govern network performance require joint observations across multiple cells.

08482#pg4>.

How the Training Signal is Built

ChaRT is pretrained through context-aware masked beam modelling and self-distillation <ref:2609.08482#pg4>. The model learns transferable radio representations from measurement reports routinely generated by operational cellular networks <ref:2609. p. 0

Where the paper says it

Physics-based tools require detailed site descriptions and computation, whereas task-specific models need dedicated measurements and transfer poorly across deployments. Here we introduce ChaRT, a foundation model that learns transferable radio representations from measurement reports routinely generated by operational cellular networks.

08482#pg4>. The pretraining objective decomposes into four terms, including beam reconstruction, representation alignment, expert routing, and missing-beam penalty <ref:2609. p. 9
Where the paper says it

The 10th–90th percentile MAE of ChaRT is roughly half that of XGBoost across parameter types, indicating more reliable predictions under diverse propagation conditions. Moreover, the channel-modeldriven augmentation during pretraining exposes the model to power scaling and beam-pattern variations consistent with (1), enabling it to learn the structural effects of parameter changes through representation alignment rather than direct RSRP regression—bypassing the estimation error inherent in the statistical channel model.

08482#pg4>. The student–teacher framework combines beam-level reconstruction with representation-level alignment across physically consistent views <ref:2609. p. 9
Where the paper says it

The 10th–90th percentile MAE of ChaRT is roughly half that of XGBoost across parameter types, indicating more reliable predictions under diverse propagation conditions. Moreover, the channel-modeldriven augmentation during pretraining exposes the model to power scaling and beam-pattern variations consistent with (1), enabling it to learn the structural effects of parameter changes through representation alignment rather than direct RSRP regression—bypassing the estimation error inherent in the statistical channel model.

08482#pg4>.

Where the Gains Come From

ChaRT demonstrates that operational measurement reports contain transferable network-scale propagation structure <ref:2609.08482#pg4>. The results show that MR data contain transferable network-scale propagation structure, a single set of pretrained weights generalizes without retraining or architectural modification to cities, frequency bands and deployment scenarios not observed during pretraining <ref:2609. p. 2

Where the paper says it

2 billion beam-level observations collected from 5G networks in a single city (Fig. Our results show that MR data contain transferable network-scale propagation structure: a single set of pretrained weights generalizes without retraining or architectural modification to cities, frequency bands and deployment scenarios not observed during pretraining (Fig.

08482#pg4>. ChaRT achieves R2 = 0. p. 2
Where the paper says it

At an 80% masking ratio under cross-city zero-shot transfer, ChaRT reduces beam-level and cell-level RSRP reconstruction mean absolute error (MAE) by 25% and 34%, respectively, relative to XGBoost. In the cross-city new-site evaluation, ChaRT achieves R2 = 0.

62 in the cross-city new-site evaluation, compared with R2 = 0. p. 2
Where the paper says it

In the cross-city new-site evaluation, ChaRT achieves R2 = 0. 62, compared with R2 = 0.

24 for TabPFN v2 and negative values for XGBoost <ref:2609. p. 2
Where the paper says it

62, compared with R2 = 0. 24 for TabPFN v2 and negative values for XGBoost, Sionna and 3GPP TR 38.

08482#pg4>. The model's ability to infer cell-specific coverage geometry, not merely a plausible average signal level, is the basis for the zero-shot network planning and optimization tasks evaluated next <ref:2609.08482#pg4>.

Downstream Applications and Transferability

The pretrained representation supports five downstream applications through lightweight task-specific heads attached to the frozen backbone <ref:2609.08482#pg4>. These include user localization, beam prediction, propagation scenario classification, SINR estimation and beamspace clustering <ref:2609. p. 0

Where the paper says it

With only 1% of labelled data, it supports user localization, beam prediction, propagation scenario classification and estimation of the signal-to-interference-plus-noise ratio. The learned representation further enables beamspace clustering for reusable radio-grid construction.

08482#pg4>. Cross-city transfer is evaluated by applying the City A-trained head directly to the target city without target-city adaptation <ref:2609. p. 6
Where the paper says it

ChaRT infers cell-specific coverage geometry, not merely a plausible average signal level. When every beam of a target cell is hidden, mean imputation retains some information about the overall RSRP scale but cannot distinguish one unobserved cell from another: its cross-city R2 drops from 0.

08482#pg4>. The 1% setting of labelled data serves as the primary few-shot adaptation setting for these tasks <ref:2609. p. 3
Where the paper says it

Each MR sample records multi-cell, multi-beam RSRP observations that are tokenized and fed into the model. The frozen model directly serves zero-shot network planning and optimization tasks (radio map construction, new-site deployment prediction, parameter tuning), or adapts to supervised downstream tasks (localization, beam prediction, scenario classification, SINR estimation) with 1% of labelled data.

08482#pg4>.

Scaling and Future Directions

ChaRT improves consistently as model size and pretraining data volume increase, although the marginal gains diminish in both cases <ref:2609.08482#pg4>. The observed saturation with additional same-city data suggests that measurement volume cannot replace diversity across cities, frequency bands, network architectures and propagation regimes <ref:2609. p. 2

Where the paper says it

2 billion beam-level observations collected from 5G networks in a single city (Fig. Our results show that MR data contain transferable network-scale propagation structure: a single set of pretrained weights generalizes without retraining or architectural modification to cities, frequency bands and deployment scenarios not observed during pretraining (Fig.

08482#pg4>. Extending pretraining to additional cities, frequency bands and deployment configurations may expose the model to a broader range of propagation conditions than further increasing the sampling density within one city <ref:2609. p. 2
Where the paper says it

2 billion beam-level observations collected from 5G networks in a single city (Fig. Our results show that MR data contain transferable network-scale propagation structure: a single set of pretrained weights generalizes without retraining or architectural modification to cities, frequency bands and deployment scenarios not observed during pretraining (Fig.

08482#pg4>. Calibrated uncertainty, explicit physical constraints and out-of-distribution detection will also be necessary to determine when predictions can be trusted <ref:2609.08482#pg4>. The results provide evidence that the city-scale radio environment can be learned as a transferable substrate, opening a practical path towards a foundation model for network planning, operation and optimization using data that operators already collect <ref:2609. p. 2
Where the paper says it

Across these downstream applications, the pretrained backbone remains frozen. These results provide evidence that the city-scale radio environment can be learned as a transferable substrate, opening a practical path towards a foundation model for network planning, operation and optimization using data that operators already collect.

08482#pg4>.

Extended Data Table 1 | Per-city dataset statistics.

City MR samples Beam observations Cells Area (km2) Carrier (GHz) City A 1,003,362,798 18,207,030,406 3,503 16×18 2. p. 2

Where the paper says it

City A (16×18 km2 , 2. 5 GHz) is a dense urban macro deployment comprising 3,503 cells and over one billion MR samples.

1 / 3.5 City B 598,023 8,795,040 280 8×7 2. p. 2
Where the paper says it

City B (8×7 km2 , 2. 6 GHz) is a moderately dense urban deployment with 280 cells and 0.

6 City C 702,732 6,382,334 16 Area (km 2) Carrier (GHz)

Extended Data Table 2 | Parameter changes across collection rounds.

Parameter type Round 1 → Round 2 Round 1 → Round 3 Mechanical azimuth Mechanical tilt Transmit power Electrical tilt Beam codebook Note: Round 1 defines the reference antenna configuration. p. 9

Where the paper says it

We evaluate five parameter types from a real optimization campaign in City A—antenna mechanical azimuth, mechanical tilt, transmit power, antenna electrical tilt and beam codebook— with the number of adjusted cells in each collection round summarized in Extended Data Table 2 (Fig. Given only the post-tuning antenna configuration and surrounding network context, ChaRT achieves R2 of 0.

Entries report the number of cells for which each parameter changed in Round 2 or Round 3 relative to Round 1. p. 7
Where the paper says it

For parameter tuning, we isolate the ability to generalize across antenna configurations by using training and test data from City A collected across three temporal rounds, during which operators adjusted parameters of selected cells. The model is pretrained on Round 1 data only; test evaluation uses Round 2 and Round 3 data, where cells whose antenna parameters respectively changed relative to Round 1 are masked at cell level and their post-tuning coverage is predicted from surrounding context and the updated configuration vector.

For the parameter-tuning evaluation, the affected cells are masked in their entirety, and the frozen ChaRT model predicts their post-tuning multi-beam coverage from the surrounding network context and updated configuration vectors. p. 7
Where the paper says it

The model is pretrained on Round 1 data only; test evaluation uses Round 2 and Round 3 data, where cells whose antenna parameters respectively changed relative to Round 1 are masked at cell level and their post-tuning coverage is predicted from surrounding context and the updated configuration vector. We compare against TabPFN v2, XGBoost and the localized statistical channel model (LSCM)[25] , a physics-based data-driven approach that estimates the channel angular power spectrum from RSRP via sparse recovery and predicts post-tuning coverage through the modified beamforming pattern matrix.

Extended Data Table 3 | Ablation of ChaRT components.

Model configuration MAE (dB) Contextual self-distillation + Channel-model-constrained augmentation + Hierarchical masked attention + CSG-MoE (ChaRT)

Extended Data Fig. p. 9

Where the paper says it

901 produce negative R2 values. The radio maps in Extended Data Fig.

p. 9
Where the paper says it

901 produce negative R2 values. The radio maps in Extended Data Fig.

p. 9
Where the paper says it

901 produce negative R2 values. The radio maps in Extended Data Fig.

p. 9
Where the paper says it

901 produce negative R2 values. The radio maps in Extended Data Fig.

1 | Radio-map comparison for cross-city zero-shot new-site deployment in City B. Each row corresponds to one target cell located at the centre of the map. p. 6
Where the paper says it

ChaRT infers cell-specific coverage geometry, not merely a plausible average signal level. When every beam of a target cell is hidden, mean imputation retains some information about the overall RSRP scale but cannot distinguish one unobserved cell from another: its cross-city R2 drops from 0.

Building footprints are overlaid to indicate the surrounding urban environment. p. 2
Where the paper says it

These results provide evidence that the city-scale radio environment can be learned as a transferable substrate, opening a practical path towards a foundation model for network planning, operation and optimization using data that operators already collect. Results A large-scale MR dataset from operational 5G networks in China is collected to pretrain and evaluate ChaRT, spanning three major cities with distinct urban morphologies, network densities and frequency bands (Fig.

Columns show the measured RSRP and predictions from ChaRT, TabPFN v2, XGBoost, Sionna and the 3GPP TR 38. p. 8
Where the paper says it

0 R 2 Mechanical azimuth Mechanical tilt Transmit power Electrical tilt 0 5 10 15 20 25 MAE (dB) Beam codebook 0 20 40 60 80 NMSE (×10 −3 ) -0. 8 R 2 a b ChaRT TabPFN v2 XGBoost Sionna 3GPP TR 38.

901 model. RSRP is shown in dBm, and the insets magnify regions exhibiting pronounced spatial variation. p. 4
Where the paper says it

We therefore evaluate whether ChaRT can reconstruct missing RSRP values from partial multi-cell, multibeam context, which is the core foundation capability underlying the network lifecycle tasks considered below. To disentangle the axes of generalization, we evaluate under three scenarios of increasing severity: in-distribution evaluation (random split within City A), spatial extrapolation (geographically disjoint City A regions) and cross-city zero-shot transfer (City B and City C, no fine-tuning).

Extended Data Fig. 2 Radio-map comparison under antenna parameter tuning. p. 9

Where the paper says it

Moreover, the channel-modeldriven augmentation during pretraining exposes the model to power scaling and beam-pattern variations consistent with (1), enabling it to learn the structural effects of parameter changes through representation alignment rather than direct RSRP regression—bypassing the estimation error inherent in the statistical channel model. The radio-map comparison in Extended Data Fig.

Each row presents a representative cell for which one operational parameter was adjusted. p. 7
Where the paper says it

For parameter tuning, we isolate the ability to generalize across antenna configurations by using training and test data from City A collected across three temporal rounds, during which operators adjusted parameters of selected cells. The model is pretrained on Round 1 data only; test evaluation uses Round 2 and Round 3 data, where cells whose antenna parameters respectively changed relative to Round 1 are masked at cell level and their post-tuning coverage is predicted from surrounding context and the updated configuration vector.

From top to bottom, the changes concern mechanical azimuth, mechanical tilt, transmit power, electrical tilt and beam codebook. p. 8
Where the paper says it

The comparison includes ChaRT, TabPFN v2, XGBoost, Sionna and 3GPP TR 38. b, Prediction of post-tuning coverage in City A after changes to mechanical azimuth, mechanical tilt, transmit power, electrical tilt and beam codebook.

The corresponding settings change from 0◦ to 350◦, 4◦ to 7◦, −0. p. 4
Where the paper says it

Whiskers denote the 10th–90th percentiles of per-sample errors and therefore characterize reconstruction reliability across heterogeneous radio conditions. We evaluate two complementary masking settings.

2534 to 1.747 dBm per resource element, and beam-codebook index 0 to 7, respectively. p. 2
Where the paper says it

7 beams from neighboring cells per sample. The beam-level RSRP spans a wide dynamic range: beams from serving cells center around −90 dBm, while beams from neighboring cells are distributed around −98 dBm with a long tail below −130 dBm (Fig.

The target cell is located at the centre of each map, and building footprints indicate the surrounding urban environment. p. 9
Where the paper says it

The LSCM establishes a statistical mapping between RSRP and the channel angular power spectrum, but recovering a high-dimensional sparse channel structure from low-dimensional and incomplete RSRP observations of a single cell is inherently ill-posed. ChaRT circumvents this limitation by conditioning on the multi-cell network context: the RSRP patterns of surrounding cells collectively constrain the propagation environment of the target cell, a relationship that the model has learned to exploit from millions of operational MR samples.

Columns show the measured RSRP before and after tuning and the post-tuning predictions from ChaRT, TabPFN v2, XGBoost and LSCM. p. 7
Where the paper says it

We compare against TabPFN v2, XGBoost and the localized statistical channel model (LSCM)[25] , a physics-based data-driven approach that estimates the channel angular power spectrum from RSRP via sparse recovery and predicts post-tuning coverage through the modified beamforming pattern matrix. The anonymized UE coordinates used to render the maps are obtained from a location-tagged subset of MR samples collected through a consenting-user crowdsourcing programme [26] ; the corresponding measurements and predictions are spatially interpolated only for visualization.

Extended Data Fig. 3 | Latent organization of beamspace-clustering representations on the held-out Round 3 data. t-SNE projections are shown for 50 representative radio grids from each method, with colours denoting grid assignments. p. 0

Where the paper says it

With only 1% of labelled data, it supports user localization, beam prediction, propagation scenario classification and estimation of the signal-to-interference-plus-noise ratio. The learned representation further enables beamspace clustering for reusable radio-grid construction.

The projections provide a qualitative comparison of inter-grid mixing and the organization of the learned representation spaces; they are not used to quantify cluster separation or downstream performance. p. 0
Where the paper says it

With only 1% of labelled data, it supports user localization, beam prediction, propagation scenario classification and estimation of the signal-to-interference-plus-noise ratio. The learned representation further enables beamspace clustering for reusable radio-grid construction.

Extended Data Fig. 4 | Spatial organization of expert routing for a representative cell. The same MR locations are shown in all panels, and the red star marks the serving cell. p. 3

Where the paper says it

a Distribution of no. of observed cells and beams per sample Distribution of beam-wise RSRP of serving and neighboring cell Large-scale cross-city measurement report datasets No.

a, Observed RSRP distribution. p. 3
Where the paper says it

a Distribution of no. of observed cells and beams per sample Distribution of beam-wise RSRP of serving and neighboring cell Large-scale cross-city measurement report datasets No.

b, Strongest-beam assignments, revealing directionally organized propagation regions around the cell. c, Expert assignments produced by vanilla MoE without the auxiliary strongest-beam prediction objective. p. 4
Where the paper says it

Mean imputation fills a missing beam using the mean RSRP of the available beams in the same cell, or the available beams from the remaining cells when the target cell is fully missing; its performance indicates how much of the missing RSRP can be explained by the average signal level of the available context alone, without modeling beam-specific or inter-cell propagation structure. XGBoost is a strong gradient-boosted tree baseline, whereas TabPFN v2 is a transformer-based foundation model for tabular prediction.

d, Expert assignments produced by CSG-MoE. p. 1
Where the paper says it

On the model axis, the core obstacle is representational scope. A single-link CSI measurement is by construction an observation of one transmitter-receiver pair in isolation, encoding the link’s multipath structure produced by radio signals arriving along reflected, diffracted and scattered paths, but carrying no information about the surrounding network.

Extended Data Fig. 5 Scaling of ChaRT with model size and pretraining data volume. p. 9

Where the paper says it

Moreover, the channel-modeldriven augmentation during pretraining exposes the model to power scaling and beam-pattern variations consistent with (1), enabling it to learn the structural effects of parameter changes through representation alignment rather than direct RSRP regression—bypassing the estimation error inherent in the statistical channel model. The radio-map comparison in Extended Data Fig.

a, Validation MAE as a function of model size at five pretraining data scales ranging from 10 million to 1 billion MR samples. p. 2
Where the paper says it

5 GHz) is a dense urban macro deployment comprising 3,503 cells and over one billion MR samples. This single-city corpus is used for foundation model pretraining.

b, Validation MAE as a function of the number of pretraining samples for five model sizes. p. 2
Where the paper says it

5 GHz) is a dense urban macro deployment comprising 3,503 cells and over one billion MR samples. This single-city corpus is used for foundation model pretraining.

c, Joint scaling grid, with colour indicating MAE (dB). p. 9
Where the paper says it

The 10th–90th percentile MAE of ChaRT is roughly half that of XGBoost across parameter types, indicating more reliable predictions under diverse propagation conditions. Moreover, the channel-modeldriven augmentation during pretraining exposes the model to power scaling and beam-pattern variations consistent with (1), enabling it to learn the structural effects of parameter changes through representation alignment rather than direct RSRP regression—bypassing the estimation error inherent in the statistical channel model.

The MAE ranges from 6. p. 2
Where the paper says it

Our results show that MR data contain transferable network-scale propagation structure: a single set of pretrained weights generalizes without retraining or architectural modification to cities, frequency bands and deployment scenarios not observed during pretraining (Fig. At an 80% masking ratio under cross-city zero-shot transfer, ChaRT reduces beam-level and cell-level RSRP reconstruction mean absolute error (MAE) by 25% and 34%, respectively, relative to XGBoost.

38 to 5.47 dB across the grid, corresponding to a total improvement of 0. p. 4
Where the paper says it

In total, the dataset covers diverse propagation conditions, network scales, carrier frequencies and urban environments, providing a practical testbed for evaluating the generalization of network-level wireless foundation models. Built on this large-scale operational dataset, we evaluate ChaRT on the key capabilities that determine whether a foundation model can serve the network management lifecycle.

91 dB.

Extended Data Table 3 | Ablation of ChaRT components.

Model configuration MAE (dB) Contextual self-distillation + Channel-model-constrained augmentation + Hierarchical masked attention + CSG-MoE (ChaRT) Note: Components are added cumulatively to the preceding configuration. p. 7

Where the paper says it

The model is pretrained on Round 1 data only; test evaluation uses Round 2 and Round 3 data, where cells whose antenna parameters respectively changed relative to Round 1 are masked at cell level and their post-tuning coverage is predicted from surrounding context and the updated configuration vector. We compare against TabPFN v2, XGBoost and the localized statistical channel model (LSCM)[25] , a physics-based data-driven approach that estimates the channel angular power spectrum from RSRP via sparse recovery and predicts post-tuning coverage through the modified beamforming pattern matrix.

All variants are evaluated using cell-level RSRP reconstruction at a 40% masking ratio under spatial extrapolation in City A. p. 2
Where the paper says it

Our results show that MR data contain transferable network-scale propagation structure: a single set of pretrained weights generalizes without retraining or architectural modification to cities, frequency bands and deployment scenarios not observed during pretraining (Fig. At an 80% masking ratio under cross-city zero-shot transfer, ChaRT reduces beam-level and cell-level RSRP reconstruction mean absolute error (MAE) by 25% and 34%, respectively, relative to XGBoost.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Beam-level reconstruction from incomplete observations allows for radio map construction, coverage monitoring, and counterfactual network evaluation without dedicated measurement campaigns or site-specific model retraining.

  2. New-site deployment prediction enables operators to predict coverage for undeployed sites by using only the surrounding network context, which supports planning decisions before any physical change is executed.

  3. Antenna parameter tuning prediction allows operators to evaluate candidate configurations before deployment, forecasting post-adjustment coverage using only the post-tuning antenna configuration and surrounding network context.

  4. User localization can be performed with minimal label data, as ChaRT reduces mean localization error from 111.0 m to 62.9 m when only 1% of labelled data is used for supervised fine-tuning.

  5. Beam prediction accuracy improves significantly, reaching 75.4% with only 1% of labelled data, indicating that the learned representation captures beam-level directional structure that task agnostic baselines cannot recover from the same input features.

  6. Scenario classification can achieve a macro F1 of 83.0% with only 1% of labelled data, showing that ChaRT is effective in the data scarce regime where labelled samples are scarce.

  7. SINR estimation accuracy improves substantially, reaching an MAE of 2.48 dB with 1% labelled data, providing direct evidence that the pretrained representation captures network-level interference structure.

  8. Beamspace clustering enables the construction of spatially compact and radio-homogeneous grids for downstream network modelling and analysis by mapping MR samples into reusable radio grids.

Abstract

Wireless cellular networks form the connective tissue of human society, sustained by a continuous physical dialogue between engineered infrastructure and its surroundings. Radio signals emitted from base stations traverse terrain, diffract around buildings and scatter through streets before reaching billions of users. Together, these interactions produce the city-wide radio environment on which every network decision rests. Shaping this environment through deployment and optimization determines the connectivity societies rely on, yet learning it effectively at city scale and generalizing across diverse cities and deployments remain open challenges. Here we answer positively by introducing ChaRT, a foundation model that learns transferable radio representations from measurement reports generated by deployed cellular networks. These reports provide abundant multi-cell, multi-beam observations without dedicated campaigns, forming a scalable data foundation for city-scale learning. ChaRT embeds beam-level angular structure, network hierarchy and propagation-regime diversity in its architecture, and is pretrained through context-aware masked beam modelling and self-distillation with channel-model-constrained augmentation. We pretrain ChaRT on over one billion reports comprising 18.2 billion beam-level observations from 3,503 cells in one city. With a single set of weights, ChaRT reconstructs radio environments in unseen cities and transfers to radio map construction, new-site prediction and network parameter tuning. With only 1% of labelled data, it supports user localization, beam prediction, propagation scenario classification and estimation of the signal-to-interference-plus-noise ratio. The learned representation further enables beamspace clustering for reusable radio-grid construction. These results establish ChaRT as a transferable foundation for network-wide intelligence.

Sources

Related papers