PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Xinyu Qin, Wenqiang Pu, Hongcheng Dong, Bingsheng Peng, Ye Xue, Tsung-Hui Chang, Zhi-Quan Luo

arXiv:2609.08482v4eess.SP

TL;DR

Scaling localization and beam-management tasks remains costly because they require dedicated campaigns, site-specific retraining, or substantial labelled data. ChaRT learns a shared representation from measurement reports and physical cell configurations; its evaluation supports learning network structure across cities within one model.

  • Problem

    Scaling localization and beam-management tasks remains costly because they require dedicated campaigns, site-specific retraining, or substantial labelled data.

    Source: Results · p. 5

  • Method

    ChaRT jointly encodes measurement reports and physical cell configurations into a shared representation, pretrained with context-aware masked beam modelling.

    Source: Discussion · p. 15, Modelling · p. 19

  • Results

    Across evaluated cities, ChaRT provides evidence that cells, beams, configurations and propagation conditions can be learned within one model rather than reconstructed separately for each site and application.

    Source: Discussion · p. 15

  • Takeaways & Limitations

    The shared representation provides a common predictive basis for network decision-making.

    Source: Discussion · p. 15

    Validation covers City A and two transfer cities across sub-6-GHz deployments; broader regions and network architectures still need evaluation.

    Source: Discussion · p. 16

Abstract

from arXiv · show

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.

Results

The results assess ChaRT’s ability to reconstruct and predict radio environments across spatial, cross-city and unseen-site settings, and to transfer its pretrained representations to downstream radio tasks and reusable radio-grid construction. They also examine data-efficient prediction and scaling with model size and same-city training data.

Results

  • Pretraining uses over one billion MR samples from City A's 3,503 cells, while evaluation spans City B's 280 cells and City C's 16-cell suburban deployment.The cities differ in network density, urban morphology and carrier frequency; City A spans 2.1/3.5 GHz, City B 2.6 GHz and City C 3.5 GHz.Source: Results · p. 3

ChaRT reconstructs multi-beam RSRP from incomplete observations across cities

  • At 80% beam masking, ChaRT's MAE is 3.17 dB, versus 5.05 dB for XGBoost, 7.44 dB for TabPFN v2 and 5.95 dB for mean imputation.ChaRT achieves the lowest MAE at every masking ratio across all three evaluation scenarios.Source: ChaRT reconstructs multi-beam RSRP… · pp. 5, 7, Figure 2 · p. 6
  • At 80% beam masking, ChaRT's MAE is 3.65 dB under spatial extrapolation and 4.32 dB cross-city, where its R² is 0.75 versus 0.63 for XGBoost and 0.25 for TabPFN v2.At 20% cross-city masking, XGBoost has marginally higher R² and lower NMSE than ChaRT despite higher MAE; that advantage disappears above 20%.Source: Figure 2 · p. 6, ChaRT reconstructs multi-beam RSRP… · p. 7
  • At 80% cell masking, ChaRT's MAE is 4.35 dB in-distribution, 5.44 dB under spatial extrapolation and 5.61 dB cross-city, with 43%, 29% and 34% lower errors than XGBoost.ChaRT retains cross-city R² = 0.61, versus 0.25 for mean imputation, and infers cell-specific coverage rather than only an average signal level.Source: Figure 2 · p. 6, ChaRT reconstructs multi-beam RSRP… · p. 7

ChaRT transfers zero-shot to network planning and optimization

  • For unseen City A sites, co-site expansion yields R² = 0.72 and MAE = 4.91 dB, cross-band cells R² = 0.67 and MAE = 5.25 dB, and greenfield sites R² = 0.55 and MAE = 5.71 dB.The model predicts target-cell coverage from surrounding network context without target-site measurements; the same cell-masking mechanism supports planning and parameter tuning without labels or architectural changes.Source: Figure 3 · p. 8, ChaRT transfers zero-shot to network… · p. 9
  • ChaRT achieves greenfield coverage-prediction R2 = 0.55 (TabPFN v2: 0.28) and, in unseen City B, R2 = 0.62 and MAE = 5.28 dB, 38% below TabPFN v2.XGBoost, Sionna and 3GPP TR 38.901 have negative R2 values in both reported settings.Source: ChaRT transfers zero-shot to network… · pp. 9–10
  • Across five antenna-parameter changes, ChaRT achieves R2 = 0.60–0.80 versus TabPFN v2's 0.37–0.48; transmit-power prediction is strongest (R2 = 0.80, MAE = 4.38 dB).For mechanical azimuth and tilt, ChaRT retains R2 values of 0.60 and 0.62, while XGBoost is negative and LSCM remains below 0.28.Source: ChaRT transfers zero-shot to network… · p. 10
  • ChaRT maps reproduce the direction and magnitude of parameter-induced coverage changes and local spatial redistribution; XGBoost is noisy, while TabPFN v2 is smooth but directionally inaccurate.Azimuth and tilt change main-beam orientation, transmit power changes the coverage footprint, and beam-codebook reconfiguration modifies the directional pattern.Source: ChaRT transfers zero-shot to network… · p. 10

ChaRT empowers data-efficient adaptation to downstream tasks

  • With 1% labelled data, ChaRT's mean localization error is 62.9 m, 46% below same-fraction XGBoost (116.8 m); at 10%, it reaches 40.9 m, 40% below WKNN (68.4 m).Source: ChaRT empowers data-efficient… · p. 11
  • Zero-shot top-1 beam accuracy is 66.9% versus MLP's 44.3%; ChaRT reaches 75.4% at 1% labels and 80.8% at 10%, 18.5 points above TabPFN v2.At 1% labelled data, ChaRT exceeds all baselines trained with 10%.Source: Figure 4 · p. 12, ChaRT empowers data-efficient… · p. 13
  • At 1% labels, ChaRT reaches 83.0% macro-F1 versus MLP's 76.8% and TabPFN v2's 82.0%; at 10%, it reaches 86.7% versus XGBoost's 86.1%.Without target-city adaptation, methods have comparable macro-F1 scores of approximately 62–66%.Source: Figure 4 · p. 12, ChaRT empowers data-efficient… · p. 13
  • At 1% labels, SINR MAE is 2.48 dB, 9% below TabPFN v2's 2.71 dB; at 10%, it is 2.18 dB, 10% below XGBoost's 2.42 dB.The RSRP-based parametric model has MAE above 5 dB across data fractions, while ChaRT's result indicates its representation captures network-level interference information.Source: Figure 4 · p. 12, ChaRT empowers data-efficient… · p. 13
  • ChaRT grids have a 128.2 m mean within-grid radius, 23.8% below BeamRSRP, and the lowest CellRSRP MAE at 5.72 dB; BeamRSRP's BeamRSRP MAE is only 0.12 dB lower.ChaRT's representations form more compact, clearly separated groups and produce grids that are more compact in physical space and consistent in cell-level radio measurements.Source: Figure 4 · p. 12, ChaRT empowers data-efficient… · p. 14
  • Across four tasks, the frozen ChaRT backbone gains consistently across labelled-data fractions, with effective adaptation at 1%; beamspace clustering also yields compact, radio-homogeneous grids.The results indicate that the representation can be specialized across operational applications rather than reconstructed separately for each task.Source: ChaRT empowers data-efficient… · p. 14

ChaRT scaling approaches saturation with model size and same-city data volume

  • MAE ranges from 6.38 to 5.47 dB across model and data scales, with most improvement in the small-model, low-data regime.ChaRT improves as model size and pretraining volume increase, but marginal gains diminish.Source: Figure 5 · p. 14, ChaRT scaling approaches saturation… · p. 14
  • At 1 billion samples, increasing model size from 0.5 million to 35 million parameters reduces MAE by 0.33 dB, while increasing it to 180 million yields 0.08 dB more reduction.Across the scaling grid, the 180-million-parameter model’s average improvement over the 35-million-parameter model is only 0.07 dB.Source: ChaRT scaling approaches saturation… · p. 15
  • For the 35-million-parameter model, increasing pretraining data from 10 million to 150 million samples reduces MAE from 5.73 to 5.57 dB; it is 5.55 dB at both 500 million and 1 billion samples.The 10-million-to-150-million increase accounts for 89% of the total improvement observed from 10 million to 1 billion samples.Source: ChaRT scaling approaches saturation… · p. 15
  • Under the evaluated setting, improvement is limited beyond approximately 35 million parameters and 500 million samples, suggesting scaling should prioritize data diversity.Additional cities, frequency bands and deployment configurations may expose ChaRT to broader propagation conditions than denser sampling within one city.Source: ChaRT scaling approaches saturation… · p. 15

Discussion

ChaRT uses jointly encoded measurement reports and physical cell configurations to learn a shared representation for radio-environment reconstruction and network decision-making. The discussion frames it as a predictive foundation for planning and optimization, while identifying broader validation and prospective closed-loop evaluation as open steps.

Discussion

  • The results advance wireless machine intelligence from individual-link and task-specific modelling toward transferable learning from network-scale observations.The discussion characterizes the representation as supporting radio-environment reconstruction, network planning and data-efficient downstream inference.Source: Discussion · p. 15
  • The shared representation models coupled cells, beams, configurations and propagation conditions across evaluated cities, rather than rebuilding the model separately for each site and application.The model pairs network observations with physical configurations and supplies a common predictive basis for network decision-making.Source: Discussion · p. 15
  • Radio-map construction, new-site assessment and parameter tuning become queries to the same model, while downstream tasks use lightweight heads and limited target-city supervision.The discussion names localization, beam prediction and interference-aware estimation among these downstream tasks.Source: Discussion · pp. 15–16
  • The present study establishes the predictive component of a possible closed-loop system, but controller integration and prospective evaluation of end-to-end network objectives remain open.Prospective trials are required to establish whether decisions informed by the representation improve network-level performance under real-world changes.Source: Discussion · p. 16
  • Pretraining uses City A, with transfer evaluated in two additional cities and sub-6-GHz deployments; broader validation across regions and network architectures is still needed.Measurement reports reflect active-user locations and reporting policies, potentially under-representing low-traffic regions, weak-coverage areas and rare propagation conditions.Source: Discussion · p. 16
  • ChaRT currently produces point predictions without calibrated uncertainty, and retrospective parameter-tuning results do not establish the causal effects of prospective interventions.Prospective interventions are needed to quantify those effects.Source: Discussion · p. 16
  • The authors position calibrated uncertainty and prospective validation as next steps toward autonomous closed-loop control.Prospective trials are required to establish whether decisions based on the learned representation improve network-level performance.Source: Discussion · p. 16
  • Future work could extend pretraining across regions, frequency bands and deployment types, and add temporal, traffic and building-geometry information to model environmental change.The paper also identifies calibrated uncertainty, stronger physical constraints and out-of-distribution detection as important for determining when predictions can be trusted.Source: Discussion · p. 16

Methods

The Methods section describes ChaRT’s pretraining on routine cellular measurement reports, combining a unified masked-beam objective, propagation-aware expert routing, and contextual student–teacher learning. It also outlines transfer of the predictor or frozen representations to network-intelligence tasks.

Methods

  • ChaRT casts network-intelligence tasks, from radio-map construction to parameter tuning, as masked beam prediction and supports zero-shot operation or low-label supervised adaptation.The methodology also includes contextual self-distillation and channel-model-constrained augmentation to promote generalization across network configurations.Source: Methods · p. 16

Datasets

  • Over one billion MR samples comprising approximately 18.2 billion beam-level observations span 3,799 cells across three Chinese cities, collected automatically during routine network operation.The standardized reports include multi-beam RSRP and serving- and neighboring-cell identifiers, without dedicated measurement campaigns or specialized infrastructure.Source: Datasets · p. 17

Modelling

  • ChaRT’s context-aware masked beam modelling objective predicts hidden beam measurements from visible reports and target-beam physical configurations, supporting transferable radio representations for downstream tasks.The objective combines reconstruction with auxiliary regularization to promote representation quality.Source: Modelling · p. 19

Architecture

  • The tokenizer binds observed signal strength to each beam’s physical identity, while hierarchical masked attention models within-cell beam correlations and across-cell spatial dependencies.It encodes beam radiation patterns and cell geometry rather than fixed indices; masked entries serve as queries, not keys, to avoid information leakage.Source: Figure 6 · p. 18, Architecture · pp. 19–20
  • A channel-structure-guided mixture of experts routes each cell token to a regime-specific expert while a shared expert processes all tokens.Top-1 routing activates one routed expert, addressing heterogeneous channel structures across cells.Source: Figure 6 · p. 18, Architecture · pp. 20–21
  • Strongest-beam supervision guides CSG-MoE routing toward coherent propagation directions, yielding more contiguous, directionally organized expert regions without explicit environment labels.The strongest-beam index acts as a physical proxy for dominant propagation direction.Source: Architecture · p. 21

Pretraining ChaRT

  • ChaRT uses contextual student–teacher self-distillation: the student reconstructs masked beams and aligns its representation to an EMA teacher viewing a fully visible augmented sample.The teacher weights are updated as an exponential moving average of the student weights, with decay λ = 0.996.Source: Pretraining ChaRT · pp. 21–22
  • Training mixes beam- and cell-level masks (20%–80%) with channel-model-constrained augmentation, plus strongest-beam routing and a penalty against implausibly strong predictions for missing beams.Augmentation perturbs power and beam or antenna configurations while respecting the propagation invariant; cells with severely incomplete observations are excluded from teacher-view construction.Source: Pretraining ChaRT · pp. 22–23

Applications for network intelligence tasks

  • For radio-map construction, new-site prediction, and parameter tuning, the pretrained predictor directly estimates target beam-level RSRP without gradient updates or target-city calibration.The tasks use missing beams as radio-map targets, all beams of a planned cell as new-site targets, or a target cell under a proposed antenna configuration.Source: Applications for network intelligence… · p. 24
  • For other outputs, ChaRT freezes its pretrained backbone and adapts lightweight task-specific heads using a small labelled target-city subset, initialized from a source-city-trained head.Localization and SINR estimation use regression heads; beam prediction and propagation-scenario classification use classifiers.Source: Applications for network intelligence… · p. 24
  • For beamspace clustering, an embedding head and deep embedded clustering jointly refine cell representations and learnable grid prototypes, assigning radio data to reusable radio grids.These grids support downstream network modelling and analysis.Source: Applications for network intelligence… · p. 24

Extended Data

The Extended Data documents dataset and tuning-round context, ablation settings, and visual comparisons of radio maps, clustering, and expert routing. The routing comparison is consistent with lower reconstruction error after introducing CSG-MoE.

Extended Data Table 1

  • The per-city dataset statistics distinguish City A’s dense multi-band deployment, City B’s moderately dense single-band deployment, and City C’s sparse single-band suburban deployment.MR samples are individual measurement reports, while beam observations are beam-level RSRP records.Source: Table 1 · p. 29, Extended Data · p. 29

Extended Data Table 2

  • The parameter-change table summarizes how many cells changed each parameter in Rounds 2 and 3 relative to Round 1; affected cells are fully masked for frozen-model prediction.Predictions use surrounding network context and updated configuration vectors to estimate post-tuning multi-beam coverage.Source: Table 2 · p. 29, Extended Data · p. 29

Extended Data Table 3

  • The ablation table evaluates cumulative component additions using cell-level RSRP reconstruction at 40% masking under spatial extrapolation in City A.Lower MAE indicates better performance; bold marks the best result.Source: Table 3 · p. 29, Extended Data · p. 29

Extended Data Fig. 1

  • The cross-city zero-shot map comparison shows measured RSRP and predictions from ChaRT and four baselines for target cells in City B.Building footprints provide urban context, and spatial interpolation is used only for visualization.Source: Figure 1 · p. 30

Extended Data Fig. 2

  • The antenna-tuning maps compare measured RSRP before and after changes to five parameter types with post-tuning predictions from ChaRT and three baselines.Each row represents a cell and one adjusted parameter; spatial interpolation is only for visualization.Source: Figure 2 · p. 31

Extended Data Fig. 3

  • The held-out Round 3 t-SNE projections qualitatively compare inter-grid mixing and representation-space organization across methods, not cluster separation or downstream performance.The visualization shows 50 representative radio grids per method, with colors denoting grid assignments.Source: Figure 3 · p. 32

Extended Data Fig. 4

  • Reconstruction MAE falls from 4.98 dB to 4.72 dB after introducing CSG-MoE, whose expert assignments form more contiguous directional regions than vanilla MoE.The figure describes this qualitative organization as more closely matching strongest-beam assignments.Source: Figure 4 · p. 33
Loading 2609.08482v4…