Site-Specific Radio Channel Representation for 5G and 6G (2025)

Organization of the paper

In Section II we will motivate the need for an SSCR that is able to cover research and development work for upcoming new physical layer technologies as described in Sec. III and new application scenarios as explained in Sec. IV. Three basic methods for an SSCR are introduced in Sec. V, which use the scenario geometry as natural parameterization and avoids additional calibration steps. In Sec. VI we summarize a step-by-step description for the setup, validation, and use of an SSCR. Open research directions for model validation techniques are introduced in Sec. VII and unexplored channel measurement scenarios are outlined in Sec. VIII. In Sec. IX real-time radio channel emulation method for an SSCR are explained. Final conclusions are drawn and open research directions are summarized in Sec. X.

III-A Joint communication and sensing

Sensing capabilities will create a disruptive and explosive growth of new 5G and 6G applications and services. These can be economically implemented at minimal or no additional cost in existing base stations and/or user equipment heralding a new 5G and 6G era of JCAS (also known as integrated sensing and communication). This sensing capability will span a very wide range of applications, including but not limited to: Intrusion detection, health monitoring, environmental monitoring, road traffic, daily life, industrial automation, etc. For example, a base station detects and alerts a pedestrian entering a pedestrian crossing while a car is about to enter, as shown in Fig. 1.

Mono-static and bi-static network topologies are the simplest configuration for sensing. In mono-static sensing, a single base station implements the TX and RX functions at the same site, performing the sensing function within its radio coverage. In the bi-static case, two sites work together to achieve enhanced performance. The communication waveform may also be used opportunistically for the sensing function.The reliable and robust design of such solutions will require new propagation models that enable deployment in a variety of environments for different applications with a high probability of user satisfaction. Current radio channel models are excellent in the context of a TX-RX link. However, at a minimum, the return power ratios from the object being sensed and its surrounding background are critical enhancements to be added to the existing propagation modes. Other improvements required are to include spatial consistency and Doppler statistics.

III-B Antenna element group with wide aperture

With emerging technologies such as RIS and D-MIMO, antenna elements are distributed over a large area. For RISs potentially the whole facade of a building can be covered by $\lambda/2$ spaced antenna elements with adjustable reflection coefficients. For D-MIMO, distributed remote radio heads can be deployed over an area with the size of about one $\text{km}^{2}$, being connected to a central processing unit (preferably with fibre optic links) that processes their signals jointly and coherently. In both cases an SSCR is required that models the near field characterized by spherical wave fronts.

III-C Multi-band communication

With decreasing wave-length of the TX the received power decreases quadratically. Hence, directive antennas are required for frequency range (FR) 2 ($24.25$ GHz to $52.6$ GHz) and FR3 ($7$ GHz to $24$ GHz) in 5G and 6G systems. The required beam search to establish a stable link between TX and RX is a challenging and time-consuming task in outdoor mobile scenarios. Hence, multi-band communication aims to use a radio signal, e.g. in FR1 ($0$ to $7$ GHz), to identify a suitable direction of arrival and direction of departure. These directions are used to establish a high-data rate broadband link in FR2 or FR3. For this technology a frequency-consistent SSCR is needed.

IV-A Railways

Railway communications must evolve from 2G (GSM rail, GSM-R) to the next generation railway communication system with improved performance and intelligence. 5G and related technologies [2] are seen as a solution to support the increasing data traffic, various new services, and high safety requirements for future intelligent railways. The next generation railway communication system is expected to support speeds of up to 500 km/h, which poses significant challenges for dynamic channel modeling, including dynamic estimation, tracking, and characterization of MPCs, which are not well studied for 2G railway communication systems and need to be addressed in the future.Channel dispersion in the delay, Doppler, and angle domain should also be accurately characterized by an SSCR for future intelligent railways for a wide range of scenarios such as viaducts, tunnels, cuttings, stations, etc. For characteristic environments encountered in railway scenarios, an SSCR demonstrates distinct responses tailored to each specific scene. For instance, in constrained environment like tunnels, it needs to account for the confined space and multiple reflections. In contrast, in open areas like viaducts, the focus might be on line-of-sight and diffraction over edges. Well-designed large and small scale models are essential for site-specific coverage prediction and link budget calculation. By tailoring specific methods and parameters to specific environments, an SSCR can provide highly accurate and realistic simulations of radio wave propagation, enabling better design and optimization of wireless communication systems.

IV-B Street transport

Vehicular communication on streets is defined by a wireless channel where potentially both, TX and RX, are moving. Hence, time-variant Doppler shifts play a major role for an SSCR and scattering objects on both sides of the road, such as buildings, foliage, uneven terrain, etc. must be considered. Vehicular communication breaks down into different scenarios [3], ranging from static intra-vehicle or indoor parking to dynamic outdoor links between two vehicles or between a vehicle and the infrastructure. Vehicular radio channels are inherently location-specific resulting in a non-stationary fading process defined by the vehicle movement, and the movement of the surrounding vehicles in closer proximity. The building structure, street width and vegetation along the street have a strong influence on the radio channel properties as summarized in [3].

IV-C Aircraft

As the number of manned and unmanned aircraft continues to grow, there is a need to manage airspace more efficiently. This includes all aspects of communication, navigation and surveillance technologies. For all three aspects, it is important to understand the propagation channel between aircraft (manned or unmanned) and ground infrastructure. There are several key differences between the air-ground channel and traditional terrestrial communication systems. Aircraft can travel faster than terrestrial vehicles, both within and over areas where terrestrial communications take place. This poses challenges for accurate modeling of the air-ground channel, as the Doppler spectral density (DSD) [4] and power delay profile (PDP) can be rapidly time-varying.

Site-specific air-ground channel models are required, e.g., for small aircraft conducting fly-by inspections of utilities such as large wind farms, advanced air mobility aircraft maneuvering near rooftops in urban areas, and air-ground communications with a large swarm of aircraft. In addition, aircraft-specific effects such as airframe shadowing are not addressed in typical terrestrial channel modeling.

In the common case where aircraft antennas are omnidirectional, the primary components of the channel impulse response are the line of sight and the earth surface reflection. The strength of the received line-of-sight and surface reflection also depends on the altitude of the aircraft, its bank angle or pitch during take-off and landing. These components must be supplemented by additional MPCs, which may be intermittent MPCs as the aircraft moves through a volume of space. In environments where the geometric and electrical properties of obstacles (or interacting objects) can be accurately quantified, these intermittent MPCs can be estimated and tracked in the channel model. In the case of propeller aircraft, the radio signal is additionally Doppler-shifted by the frequency of the revolutions of the rotating propeller.

V-A Ray tracing

Ray tracing is a high-frequency approximation used to solve Maxwell’s equations. It has long been used successfully for deterministic channel prediction [6], particularly for deployment planning, both indoor and outdoor. As it computes the parameters of the MPCs, including amplitude, direction of arrival, and direction of departure, it provides a true double-directional channel characterization that can be combined with (almost) arbitrary antenna patterns and array patterns in post-processing. In addition, ray tracing results are inherently spatially consistent, i.e., the change in power, angles, and delays follows the environment’s geometry, and - in a static environment - the same user equipment location is associated with the same channel.

Ray tracing is typically implemented either as (i) image-theory-based ray tracing, where potential locations of image sources according to different orders of reflection are computed and rays between these locations and the RX determined, or (ii) ray launching, where a TX sends rays into different directions and follows their interactions with environmental objects until they arrive at the RX, leave the area of interest, or become too weak to be of interest, see Fig. 2. The former method is more efficient for point-to-point simulations, while the latter has advantages for many RX locations and/or area coverage investigations.

As the size of antenna arrays increases and wavelength decreases, wavefront curvature must be considered. Wavefront curvature implies that the RX is within the Rayleigh distance of the TX; as an important consequence, spatial multiplexing or mode multiplexing (for orbital angular momentum) becomes feasible. At the same time, operating under the assumption of homogeneous plane waves, when in reality, wavefront curvature occurs, can lead to a model mismatch in channel estimation, e.g., based on the sparsity assumption. Since the rays used in ray tracing are generally homogeneous plane waves, new ways of incorporating wavefront curvature are being explored. One way is to increase the number of rays representing a small angular range to satisfy the plane-wave assumption locally; however, this significantly increases the computational effort.D-MIMO systems and ultra-massive MIMO arrays are susceptible to another effect: different parts of the antenna elements/distributed radio units experience different amounts of shadowing. This effect, observed in several measurement campaigns, significantly impacts both channel capacity and transceiver implementations, such as sub-array selection in hybrid beamforming. Addressing this effect requires finer angular resolution, but this comes at the expense of efficiency.

Evaluation at higher frequencies first requires more accurate and higher-resolution databases. These can be obtained, for example, from light detection and ranging (LIDAR) scanning of the environment or artificial intelligence/machine learning (AI/ML)-based mapping of photographs onto full maps. In either case, the result is a point cloud, often with a resolution in the centimeter range. This leads to a dramatic increase in computation time for both image-based ray tracers and ray launchers.

Another research topic for high-frequency ray tracing is accurate models for diffuse scattering, as this phenomenon becomes more important with increasing frequency. Both, phenomenological approaches and models derived from first principles, have been investigated [7].

Finally, a ray tracer needs a calibration step. While the geometrical positions of environmental objects can be accurately determined, this is not the case for the electromagnetic properties of materials. The permittivity and conductivity of walls, windows, etc., can vary widely and are often determined by comparing the results of measurements at sample locations with the output of the ray tracer; the resulting coefficients are then used for all buildings in the area. While trial and error were typical for this task, more recently, differentiable ray tracers have been introduced, which allow more systematic optimization [8].

V-B Quasi-deterministic method

Quasi-deterministic SSCRs, also called geometry based stochastic channel models (GSCMs) in the literature, combine deterministic predictions, e.g., from a ray tracer, with stochastic but still geometry based channel modeling.

Simplified ray tracing (a small number of reflections considered) is performed on a low-resolution environmental map to deterministically find the dominant MPC (the map can be actual or synthetically generated). Smaller MPCs associated with the dominant contributions are then generated stochastically either (i) in the geometric space by placing point scatterers according to a given distribution close to surfaces of geometric objects to represent surface roughness and point clusters to represent other objects such as vegetation, moving vehicles, and pedestrian or (ii) in the delay and direction of departure/direction of arrival domain. In both cases birth/death statistics of the associated MPCs and intra-cluster statistics are usually obtained from measurements.

Quasi-deterministic models were first proposed by Kunisch and Pamp and by COST 259 in the early 2000s, and their concepts were then re-used by the European METIS and MiWeBa projects and adopted as an option for 3GPP simulations [9]. Compared to purely stochastic models (such as 3GPP), quasi-deterministic models have the advantages of spatial consistency (since the evolution of angles and delays of clusters or dominant paths follows from the geometry) and easier incorporation of specific geometries (e.g., the impact of street width can be easily incorporated). The advantages compared to ray tracing are the significantly reduced computational effort, easier exchange of environmental databases between entities that want to compare their results, and a straightforward way of incorporating mobile scatterers.

V-C AI/ML based method

A promising approach of SSCR is data-driven AI/ML-based channel modeling [10]. AI/ML-based channel models are trained to learn the mapping relationship between complex environmental features and channel characteristics. It can achieve a balance between generalization and prediction accuracy compared to traditional channel models. An SSCR requires the prediction of channel characteristics of a specific site according to environmental conditions, and for AI/ML-based modeling, the ability to extract features from environmental information determines the prediction accuracy of the model.

For the extraction of environmental features, different data sources can be used, such as images, satellite maps, point cloud data, etc. The impact of data resolution on the accuracy of environmental features should be carefully considered. Methods like neural networks are usually used to extract propagation-related features from the data, which are further employed to predict channel characteristics. In the phase of model selection and network training, supervised learning is the most common approach, where models are trained to minimize prediction errors using labeled data, unsupervised learning methods are also employed in certain scenarios. For example, clustering algorithms can classify different channel conditions, and auto-encoders can extract meaningful features from raw data without labels. Additionally, generative adversarial networks can generate synthetic channel data, which is useful in scenarios with limited real-world data.

The developed AI network should take the environmental features as input and provide the corresponding site-specific channel parameters as output. Of course, physical radio propagation mechanisms and empirical channel models can be used and incorporated into the designed AI model to improve the site-specific channel prediction. Methods such as a regularization strategy, an early stop strategy and ensemble learning can also be used to avoid model over-fitting for a specific scenario. In addition, the ability of the model to self-evolve for different sites should be used as an evaluation index, and its essence lies in strengthening the existing model with additional data and features. Sufficient multi-source data and prior knowledge can support AI/ML-based channel modeling for SSCR needs with low complexity.

The combined use of the three methods presented above is a promising future research direction to achieve the best trade-off between complexity and accuracy.

VII-1 Non-stationary environment evaluation

VII-2 Frequency domain continuity

Another validation requirement for SSCR is frequency domain continuity, i.e. ensuring the applicability of channel parameters over the entire frequency range of operation. 5G NR pushes the traditional sub-$6$ GHz boundary to $7.125$ GHz (FR$1$) and introduces millimeter wave (mmWave) bands in FR$2$. FR$3$ is called the upper mid-band and is being considered for 5G-Advanced and 6G. Experiments in the sub-THz range from $52.6$-$114.25$ GHz (FR$4$) and $114.25$-$275$ GHz (FR$5$), as well as in the THz range ($>300$ GHz) have attracted interest. The validation of parameters over such different frequency ranges prompted a series of simultaneous multi-band measurements. The blocking and deviation characteristics of a scatterer become more pronounced as the carrier frequency increases. This means that small scatterers, which could be safely ignored for FR$1$ or FR$3$ bands, may need to be included in the SSCR when applied to bands for FR$2$ or above.

VII-3 Backscatter radio channel properties

JCAS requires new radio channel properties to be represented by an SSCR. Hence, radar cross-section parameters need to be included for path loss calculation in mono-/bi-static sensing. If the link lengths for TX-target and target-RX are given by $d_{1}$ and $d_{2}$, then the path loss is given by $\gamma_{\mathrm{dB}}(d_{1},d_{2},\sigma,f)=\gamma_{\mathrm{dB}}(d_{1})+\gamma_$, where $\sigma$ is the normalized reflectivity of the target. Measuring the radar cross section of different targets and classifying them according to their radar cross section range are important objectives for JCAS field test campaigns. Additional SSCR validation parameters include clutter/scattering patterns and object mobility, depending on the accuracy, resolution, and refresh rate requirements of the JCAS use case. The return power from natural or man made backgrounds is also critical for defining the target to noise power ratio.

VIII-1 Vehicular scenarios

For rail and road applications, new measurement campaigns are aimed at capturing the interplay between higher centre frequency and highly mobile site-specific features. Challenges for such campaigns include ensuring repeatable and fault-free operation of the channel sounder under extreme environmental conditions (weather, vibration, etc.) and maintaining proper synchronization between different subsystems. The node mobility causes fluctuations in the beamwidth, and as a result, the number of MPCs varies proportionally with the beamwidth. The channel becomes highly non-stationary, and a local scattering function, as described in subsection VII-1 and [3], can be used to characterize and compare the measured and modelled channel.

VIII-2 Airborne scenarios

For airborne scenarios, the mobile-to-mobile (M2M) channel model validation requires at least two additional facets: first, consideration of the full three-dimensional site geometry (unlike vehicular scenarios where a 2D model is generally sufficient), and second, consideration of scattering components (along with line-of-sight and specular reflections). The validation is done in the delay-Doppler plane to account for the fast mobility patterns of TX and RX. Classical site-general parameters, such as shadowing variance, are augmented with new features like shadowing correlation distances to ensure the validity of an airborne mobile-to-mobile channel model at different TX/RX heights.

VIII-3 Urban cellular high frequency scenarios

For example, [13] builds on the quasi deterministic channel model type described in Sec. V-B. The authors present a recipe for mmWave MPC generation for a specific urban cellular environment in Yokohama, Japan. Here, the measured data is used to estimate an exponential decay model to calibrate the ray tracer. Random clusters are generated with the measured site-specific statistical parameters of large and small scale fading parameters. This step improves the accuracy of the quasi-deterministic channel model by incorporating real-world measurement data that captures the specific characteristics of the environment under consideration.