Map-Based Channel Model (MBCM)

• MBCM is a site-specific channel modeling paradigm that maps physical environments to wireless channel characteristics such as CSI, path loss, and multipath parameters.
• It combines deterministic geometry from CAD/GIS, statistical methods, and learned models to accurately compute location-dependent channel states.
• Applications in mmWave, UAV communications, and beamforming leverage MBCM for precise, environment-aware radio maps and adaptive channel predictions.

Searching arXiv for recent and foundational papers on map-based channel models to ground the article. Searching arXiv for "map-based channel model" and closely related "channel knowledge map" work. arXiv search query: "map-based channel model OR channel knowledge map wireless" Map-Based Channel Model (MBCM) denotes a site-specific channel-modeling paradigm in which wireless channel characteristics are represented as spatial maps of environment-dependent quantities indexed by location, and in advanced formulations by time, frequency, beam, or array dimension. In this formulation, a channel map can serve as a world model of the radio environment, associating locations with channel state information (CSI), path loss, shadowing, angles, delay spread, Doppler, or full MIMO channel matrices. MBCM contrasts with purely stochastic scenario-generic models because it ties channel generation to actual geometry, materials, and object configurations; at the same time, later work shows that statistical CKM/CGM constructions and learned location-to-channel mappings are also valid MBCM realizations when they produce location-specific channel knowledge from maps, measurements, or both (Lim et al., 2017, Li et al., 2021, Qi et al., 1 Apr 2026).

1. Conceptual scope and map objects

Within the MBCM literature, the basic object is not a single scalar coverage map but a family of location-indexed channel descriptions. A Channel Knowledge Map (CKM) is defined as a database that provides a mapping from location $q$ to channel knowledge $r$,

$W: q \in Q \to r,$

for any transmitter–receiver geometry of interest. A Channel Gain Map (CGM) is a specific CKM instance focusing on large-scale channel gain in dB, whereas a channel map in the richer sense may include not only path loss but also small-scale multipath parameters, angles, Doppler, and full MIMO channel matrices (Li et al., 2021, Qi et al., 1 Apr 2026).

The literature has progressively specialized this general idea. A Link State Map (LSM) stores the probability of line-of-sight (LoS) between a ground base station and a UAV at each location in the flying plane, and thereby acts as a foundational layer for selecting LoS-conditioned or NLoS-conditioned submodels in a broader MBCM (Yang et al., 2024). A Spatial Correlation Map (SCM) stores a location-specific spatial correlation matrix $R(x)$ for multi-antenna systems, rather than a scalar gain value, and is intended to support beamforming, spatial multiplexing, and CSI reduction when instantaneous CSI is unavailable or costly (Chen et al., 22 Apr 2026). For movable-antenna systems, a small-scale channel map is the function $h(r)\in\mathbb{C}$, $r\in\mathbb{R}^3$, that maps each antenna position to the complex baseband channel coefficient (Huang et al., 27 May 2025).

Beam-aware extensions push the concept further. BeamCKM reformulates the channel map for multi-antenna systems by indexing it with a beamforming codeword $\mathbf f_j$,

$$\mathrm{CKM}(\mathcal P; \mathbf f_j) \triangleq \mathcal F(\mathcal P; \mathbf f=\mathbf f_j \mid \mathbf E),$$

so that the stored quantity remains path loss, but its beam dependence captures beam propagation characteristics such as main-lobe directionality, reflections, and shadowing by obstacles (Wang et al., 23 Nov 2025). This development makes explicit that MBCM is a family of map representations rather than a fixed data structure.

2. Physical and mathematical representation

A central mathematical form in MBCM is the path-sum representation of the channel. In a space–time continuous formulation, the frequency-domain and time-domain channels can be written as

$$H(f,\mathbf{x},t)=\sum_{\ell=1}^{L(\mathbf{x},t)} \alpha_\ell(\mathbf{x},t)\,e^{-j2\pi f\tau_\ell(\mathbf{x},t)}\,e^{j\phi_\ell(\mathbf{x},t)},$$

and

$$h(\tau;\mathbf{x},t)=\sum_{\ell}\alpha_\ell(\mathbf{x},t)\,e^{j\phi_\ell(\mathbf{x},t)}\,\delta(\tau-\tau_\ell(\mathbf{x},t)).$$

Each path $r$0 carries delay, angles of arrival and departure, Doppler, and complex gain. Dynamic scatterers induce time-varying delays and Dopplers, with

$r$1

Large-scale attenuation is commonly represented through the log-distance model

$r$2

(Qi et al., 1 Apr 2026).

The same structure appears in earlier map-based mmWave formulations. A generic SISO channel impulse response is written as

$r$3

while a narrowband or per-subcarrier MIMO response takes the form

$r$4

Because MBCM provides path delays and angles directly from the mapped environment, it naturally supports derived quantities such as delay spread, angular spread, polarization behavior, and beam-selection objectives (Lim et al., 2017).

For multi-antenna second-order descriptions, the SCM literature adopts

$r$5

Under amplitude–phase independence, uniformly distributed independent path phases, and a path-based channel decomposition, this reduces to

$r$6

and at a location $r$7,

$r$8

This decomposition motivates the Path Gain Map (PGM) and Path Angle Map (PAM) used to reconstruct SCMs from sparse samples (Chen et al., 22 Apr 2026).

Location dependence is especially severe at the wavelength scale. In multi-frequency location-to-channel mapping, a path contribution contains the factor $r$9, so a displacement of the receiver by a fraction of a wavelength causes an $W: q \in Q \to r,$0 phase change. The literature therefore treats the location-to-channel function as highly oscillatory and not well matched to generic low-frequency implicit representations (Chatelier et al., 2024).

3. Main methodological families

The MBCM literature distinguishes several methodological families, each defined by how site information, propagation physics, and measurements are combined.

Paradigm	Core mechanism	Characteristic trade-off
Pure RT	Deterministic paths from CAD/GIS geometry and materials	Location-specific and dense, but computationally heavy and weak on diffuse components and fast dynamics
Pure GBSM	Scenario-specific stochastic clusters and rays	Efficient and adaptive, but not anchored to exact geometry
Statistical CKM/CGM	Finite measurements plus expert likelihood models	Compact and efficient, but typically focused on limited channel quantities
Hybrid map-based models	Deterministic static paths plus stochastic or learned residual components	Higher fidelity and dynamics, but require calibration and environment reconstruction
Data-driven learned maps	Neural reconstruction from sparse observations	Fast inference and strong interpolation, but vulnerable to domain shift or retraining demands

A recurring clarification in the literature is that MBCM is not synonymous with pure ray tracing. The foundational mmWave overview categorizes map-based modeling into Category I map-based deterministic parameters, Category II map-based stochastic parameters, and Category III map-based hybrid parameters, the last including hybrid cluster, hybrid parameter, and hybrid with hardware measurements (Lim et al., 2017). Later work makes the distinction sharper: pure RT provides location-specific specular paths but misses diffuse multipath components and many dynamic effects, while pure GBSM preserves statistical consistency and adaptability but sacrifices point accuracy at specific coordinates (Qi et al., 1 Apr 2026, Qi et al., 16 Apr 2026).

The CKM/CGM line broadens the notion further. In that view, an MBCM need not require detailed physical maps or ray tracing at all; it can instead be realized as a compact mixture of parametric channel models whose parameters vary across latent spatial groups and are inferred from finite measurements by maximum-likelihood estimation with latent variables (Li et al., 2021). This usage differs from classical deterministic map-based modeling but remains location-specific and environment-aware.

A second common misconception is that channel maps only record gain. That is true for CGM, but later work explicitly extends CKM/MBCM to LoS probability maps, spatial correlation maps, small-scale 3D channel maps, and beam-indexed maps (Yang et al., 2024, Chen et al., 22 Apr 2026, Huang et al., 27 May 2025, Wang et al., 23 Nov 2025).

4. Construction and inference procedures

Classical deterministic construction begins from a digital map, a CAD/GIS model, or a user-defined digital terrain model. Ray-tracing MBCMs assign materials, mesh the scene, configure antenna patterns and interaction orders, and compute valid LoS, reflection, diffraction, or scattering paths. In UAV mmWave A2G work, the original digital map is reconstructed into simplified scene databases; in the reported campus scenario, the simplified database can save up to 50% time consumption while the difference of statistical properties is slight (Zhu et al., 2020). For U2V mmWave communication, the 3D propagation space is reconstructed from a digital map into triangular facets, deterministic inter-path quantities are computed by reduced RT, and intra-path rays are generated stochastically from Gaussian mixture models trained on massive RT data (Zhu et al., 2021).

Statistical CKM/CGM construction starts from finite measurement data and an expert channel model. In the EM-based CKM formulation, each sample $W: q \in Q \to r,$1 contains a location and the measured channel knowledge; the mixed-model likelihood is

$W: q \in Q \to r,$2

and the log-likelihood is maximized by an EM procedure with responsibilities

$W: q \in Q \to r,$3

For CGM, the M-step admits closed-form weighted least-squares updates, the per-iteration complexity is $W: q \in Q \to r,$4, and the method is compact because only a small number of modeling parameters need to be determined and stored (Li et al., 2021).

Bayesian sequential map updating appears in the LSM literature. There, the LoS field is modeled as a binary spatial random field on the UAV flying plane, the prior LoS probability is initialized by an empirical elevation-angle model, and measurements are assimilated by a binary Bayesian filter in log-odds form. Spatial propagation to unmeasured locations uses hard distance-ray constraints along the same azimuth and a parametric angular correlation model. In the reported simulation, with $W: q \in Q \to r,$5 and $W: q \in Q \to r,$6 m, the proposed method reduces MAE from approximately $W: q \in Q \to r,$7 for the KNN baseline to approximately $W: q \in Q \to r,$8 (Yang et al., 2024).

Neural construction methods address dense map recovery from sparse or partial observations. In movable-antenna systems, a 3D CNN with residual blocks reconstructs a small-scale channel map over a cubic movement region from a subset of measured antenna positions. With $W: q \in Q \to r,$9 and $R(x)$0, only $R(x)$1 measurements, approximately $R(x)$2 of exhaustive probing, are used at runtime; under this setting the proposed CNN achieves channel-gain-map MSE $R(x)$3 versus $R(x)$4 for 3D trilinear interpolation, a $R(x)$5 reduction (Huang et al., 27 May 2025). For SCM construction, E-SRResNet decomposes the target into PGM and PAM, augments SRResNet with multi-head attention and multi-scale feature fusion, and uses LoS, building, and BS priors; on CKMImageNet, the cosine similarity between reconstructed and ground-truth SCM exceeds $R(x)$6 in most regions (Chen et al., 22 Apr 2026).

Model-based learning addresses the wavelength-scale oscillatory nature of location-to-channel maps by deriving a physics-informed neural architecture from a propagation model. The resulting decomposition writes the channel as a sparse sum of plane-wave spatial atoms, antenna steering-vector atoms, and delay-based frequency-response atoms. In the reported experiments, MB-$R(x)$7 achieves NMSE $R(x)$8 dB on D1, $R(x)$9 dB on D2, and $h(r)\in\mathbb{C}$0 dB on D3, while near-perfect reconstructions are obtained with approximately $h(r)\in\mathbb{C}$1 locations per $h(r)\in\mathbb{C}$2, below the $h(r)\in\mathbb{C}$3 locations per $h(r)\in\mathbb{C}$4 2D Shannon–Nyquist density (Chatelier et al., 2024).

5. Dynamic, multi-antenna, and adaptive extensions

A major contemporary direction is the dynamic, queryable, full-CSI channel map. The Data–Model co-Driven (DMcD) framework constructs a space–time continuous MBCM by two-stage interpolation: a model-driven prior via hybrid RT/GBSM and an online data-driven fusion stage via an inductive edge-conditioned GNN. The prior combines a static RT component, a dynamic GBSM component, and a diffuse multipath component (DMC) calibrated to measurements; the second stage uses a Wasserstein-informed graph, GraphSAGE aggregation, ECC refinement, reconstruction and regularization losses, and online measurement injection without retraining. Reported results show that the framework significantly outperforms data-only and model-only baselines, achieves NMSE below $h(r)\in\mathbb{C}$5 at approximately $h(r)\in\mathbb{C}$6 m sampling interval with about $h(r)\in\mathbb{C}$7 observed nodes, and refreshes the map in approximately $h(r)\in\mathbb{C}$8 s per update, decomposed into $h(r)\in\mathbb{C}$9 s HCM generation and $r\in\mathbb{R}^3$0 s GNN inference on an i9-14900/32 GB RAM desktop (Qi et al., 1 Apr 2026).

The Dynamic Channel Map (DCM) literature proposes a related hybrid design. RT-GSHCM preconstructs the static map offline by RT and updates it online through a geometry-conditioned GBSM, with a Rician split

$r\in\mathbb{R}^3$1

The static specular part is anchored to map geometry, while the dynamic part models moving interaction objects. In the reported 5.5 GHz urban deployment with 320 MHz bandwidth, updating a CKM by RT took $r\in\mathbb{R}^3$2 s for $r\in\mathbb{R}^3$3 rays with up to sixth-order reflections, whereas DCM updates via RT-GSHCM took $r\in\mathbb{R}^3$4 s; measurements further show that delay PSDs and CDFs of RMS angular spread and RMS delay spread align well with the hybrid model (Qi et al., 16 Apr 2026).

Beam-aware multi-antenna extensions address the mismatch between scalar pathloss maps and coherent beamforming. BeamCKM uses CKMTransUNet, a hybrid ResNet–Transformer–UNet trained with a composite loss consisting of $r\in\mathbb{R}^3$5, Laplacian pyramid, edge-aware, and SSIM terms, to predict beam-dependent maps from environmental contours. On the reported dataset, CKMTransUNet attains MSE $r\in\mathbb{R}^3$6 versus $r\in\mathbb{R}^3$7 for RadioWNet, RMSE $r\in\mathbb{R}^3$8, NMSE $r\in\mathbb{R}^3$9, PSNR $\mathbf f_j$0 dB, and SSIM $\mathbf f_j$1, with inference time approximately $\mathbf f_j$2 s per $\mathbf f_j$3 map, about eight times faster than accelerated Sionna ray tracing at approximately $\mathbf f_j$4 s (Wang et al., 23 Nov 2025).

Domain adaptation has also entered MBCM-enabled estimation. In a MIMO-OFDM setting, quasi-static source-domain channels are generated by DeepMIMO, while target-domain MBCM data are produced by importing an OpenStreetMap file into MATLAB and using RayTracing with a 3GPP CDL model. The paper reports a Wasserstein-1 distance of approximately $\mathbf f_j$5 between source and target magnitude distributions, fine-tunes late network layers with only $\mathbf f_j$6 labeled target samples, and shows that fine-tuned LS-LI-CNN and LS-LI-GAN variants outperform LS+LI at low SNR on the MBCM domain, whereas models trained only on QSCM can underperform LS+LI because of domain mismatch (Hoang et al., 11 Jul 2025).

6. Applications, limitations, and open directions

MBCM is used wherever location-specific channel awareness matters. In mmWave system design, it supports realistic evaluation of irregular layouts, penetration, blockage, indoor realism, beam selection, mobility, and multi-link interference (Lim et al., 2017). In UAV communications, MBCM-compatible LSMs support trajectory planning, base-station association, handover anticipation, and predictive beam steering through location-specific LoS probabilities and mixed LoS/NLoS channel submodels (Yang et al., 2024). In movable-antenna systems, a reconstructed small-scale channel map enables position optimization through

$\mathbf f_j$7

and thereby supports real-time antenna reconfiguration (Huang et al., 27 May 2025). For 6G embodied intelligent agents, the channel map is explicitly framed as a perceptive world model that allows agents to plan motion, select beams and links, and close perception–action loops using queryable full CSI along trajectories (Qi et al., 1 Apr 2026).

The literature also uses MBCM as training infrastructure. The open map-based mmWave database supports DNN-based beam selection; in the reported comparison, training on GSCM data yields $\mathbf f_j$8 DNN accuracy, whereas training on map-based data yields $\mathbf f_j$9 accuracy and significantly higher beamforming gain (Lim et al., 2017). This suggests that user-specific, spatially consistent map-based data expose dependencies that are obscured in broad scenario-level stochastic models.

Limitations are equally explicit. Pure RT is site-specific and high fidelity, but computationally heavy, storage-heavy, and slow to update; pure GBSM is efficient and adaptive, but lacks location-specific accuracy; data-only methods may be locally accurate but often require retraining under distribution shift and can produce patchy artifacts under sparsity (Li et al., 2021, Qi et al., 1 Apr 2026). Map fidelity, material assignment, and calibration remain central sources of error in RT-based MBCM, especially at mmWave and above 24 GHz, where rough-surface scattering and diffuse effects become more consequential (Zhu et al., 2020, Qi et al., 16 Apr 2026). Learned MBCMs generalize within the same site more readily than across unseen environments, and domain mismatch can be substantial enough that supervised adaptation becomes necessary (Hoang et al., 11 Jul 2025).

Current research directions are correspondingly hybrid. Reported extensions include multimodal sensing with vision or LiDAR to refine geometry and detect moving scatterers, Bayesian or ensemble map fusion for uncertainty-aware decisions, active exploration policies for measurement selection, cross-frequency and cross-domain transfer, network digital twins, temporal CKMs, multi-frequency generalization, and explicit incorporation of time-varying dynamics into beam-aware maps (Qi et al., 1 Apr 2026, Wang et al., 23 Nov 2025). A plausible implication is that future MBCM systems will be judged less by whether they are deterministic, stochastic, or learned in isolation, and more by how effectively they combine environment structure, measurement assimilation, and computational tractability into a continuously queryable radio-environment model.