Operational Measurements Extend Radio Models Across Cities

City-scale cellular planning depends on estimating how signals vary across buildings, antenna configurations, frequencies, and neighboring cells. Physics-based simulators need detailed site descriptions and can be expensive to run, while task-specific learned models often require location-labelled measurements and may not transfer beyond the deployment on which they were trained. The paper introduces ChaRT, a foundation model trained on routine 5G measurement reports (MRs), arguing that these reports are a scalable source of joint observations of radio conditions.

Core Contribution

ChaRT treats a measurement report as structured evidence about a local radio environment rather than as an independent tabular row. The model is pretrained on more than one billion MR samples containing 18.2 billion beam-level observations from 3,503 cells in City A, then evaluated both within that city and under spatial and cross-city transfer. Its central claim is that a common representation can support reconstruction, planning, optimization, and supervised downstream tasks without training a separate model from scratch for each setting.

The distinction from conventional radio-map regression is the model's use of beam-level structure and network hierarchy. Each sample contains multi-cell, multi-beam RSRP observations; available beams provide context, while absent beams become prediction queries. Figure 1 characterizes City A as dense urban macro, while City B and City C provide moderately dense urban and sparse suburban settings for evaluation. The cross-city results therefore test whether representations learned from City A operational data generalize beyond a location-specific interpolation setting; they do not establish that pretraining itself covered all three propagation environments.

Technical Approach

The architecture combines four components. A measurement-report tokenizer encodes beam configuration and observed RSRP into embeddings. Hierarchical masked attention processes correlations among beams within a cell in shallow layers, then dependencies among cells in deeper layers; masked entries act as queries rather than context. A channel-structure-guided mixture of experts routes each cell token to one routed expert while retaining a shared expert for every token. An auxiliary strongest-beam prediction loss guides that routing toward dominant propagation direction.

Pretraining combines beam reconstruction with multi-granularity masking, a regularizer for missing beams, and contextual self-distillation between differently augmented views. The teacher augmentation is constrained by channel structure rather than being an arbitrary feature perturbation. This is a sensible match to MRs: the useful signal is not merely an RSRP value, but the pattern of relative strengths across beams and cells. Figure 6 lays out this division between tokenization, hierarchical attention, expert routing, and the reconstruction, regularization, and distillation objectives.

Results and Analysis

The reported experiments compare ChaRT with TabPFN v2, XGBoost, mean imputation, Sionna, the 3GPP TR 38.901 channel model, and an LSCM baseline depending on the task. Reconstruction is tested by masking individual beams or all beams belonging to selected cells, then measuring MAE, NMSE, and R² under in-distribution evaluation, spatial extrapolation, and zero-shot cross-city transfer. This is stronger evidence than a single held-out split because it tests missing observations and geographic shift separately.

For scaling, the validation MAE improves from 6.38 dB to 5.47 dB across a grid spanning 10 million to 1 billion MR samples and five model sizes, a total reduction of 0.91 dB. The figure indicates that most of that improvement occurs in the small-model, low-data regime. That tempers the scaling story: larger models and more data help, but the visible gains diminish rather than establishing an open-ended scaling law.

The planning evaluation includes co-site expansion, cross-band deployment, greenfield sites in City A, and newly deployed base stations in previously unseen City B. It also predicts post-tuning coverage after changes to mechanical azimuth, mechanical tilt, transmit power, electrical tilt, and beam codebook. Figure 3 is useful because it evaluates the operational question directly: whether the predicted radio maps retain local spatial variation after transfer or parameter changes, not simply whether average signal strength is close.

With 1% labelled data, frozen ChaRT representations are adapted using lightweight heads for localization, strongest-beam prediction, LoS/NLoS classification, and SINR estimation; the same representation is also evaluated for beamspace clustering. The paper therefore presents a broad task suite, but its strongest practical case is reconstruction and planning from already available MRs. The downstream results are promising evidence that the representation is reusable, although the paper does not show that these results replace full-field engineering measurements or detailed propagation simulation in every deployment.