DeepRT: A Hybrid Framework Combining Large Model Architectures and Ray Tracing Principles for 6G Digital Twin Channels

Abstract

With the growing demand for wireless communication, the sixth-generation (6G) wireless network will be more complex. The digital twin channel (DTC) is envisioned as a promising enabler for 6G, as it can create an online replica of the physical channel characteristics in the digital world, thereby supporting precise and adaptive communication decisions for 6G. In this article, we systematically review and summarize the existing efforts in realizing the DTC, providing a comprehensive analysis of ray tracing (RT), artificial intelligence (AI), and large model approaches. Based on this analysis, we further explore the potential of integrating large models with RT methods. By leveraging the strong generalization, multi-task processing capabilities, and multi-modal fusion capabilities of large models while incorporating physical priors from RT as expert knowledge to guide their training, there is a strong possibility of fulfilling the fast online inference and precise mapping requirements of the DTC. Therefore, we propose a novel DeepRT-enabled DTC (DRT-DTC) framework, which combines physical laws with large models like DeepSeek, offering a new vision for realizing the DTC. Two case studies are presented to demonstrate the possibility of this approach, which validate the effectiveness of physical law-based AI methods and large models in generating the DTC.

Full text

Academic Editor: Dimitra I.
Kaklamani
Received: 5 March 2025
Revised: 25 April 2025
Accepted: 30 April 2025
Published: 1 May 2025
Citation: Li, M.; Wu, T.; Dong, Z.; Liu,
X.; Lu, Y.; Zhang, S.; Wu, Z.; Zhang, Y.;
Yu, L.; Zhang, J. DeepRT: A Hybrid
Framework Combining Large Model
Architectures and Ray Tracing
Principles for 6G Digital Twin
Channels. Electronics 2025,14, 1849.
https://doi.org/10.3390/
electronics14091849
Copyright: © 2025 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license
(https://creativecommons.org/
licenses/by/4.0/).
Article
DeepRT: A Hybrid Framework Combining Large Model
Architectures and Ray Tracing Principles for 6G Digital
Twin Channels
Mingyue Li, Tao Wu, Zhirui Dong, Xiao Liu, Yiwen Lu, Shuo Zhang, Zerui Wu, Yuxiang Zhang * , Li Yu
and Jianhua Zhang *
State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and
Telecommunications, Beijing 100876, China; mingyueli@bupt.edu.cn (M.L.); vvt@bupt.edu.cn (T.W.);
dzr2002@bupt.edu.cn (Z.D.); 2021212529@bupt.cn (X.L.); lyw@bupt.edu.cn (Y.L.);
zhangshuosheldor@bupt.edu.cn (S.Z.); wuzerui@bupt.edu.cn (Z.W.); li.yu@bupt.edu.cn (L.Y.)
*Correspondence: zhangyx@bupt.edu.cn (Y.Z.); jhzhang@bupt.edu.cn (J.Z.)
Abstract: With the growing demand for wireless communication, the sixth-generation (6G)
wireless network will be more complex. The digital twin channel (DTC) is envisioned
as a promising enabler for 6G, as it can create an online replica of the physical channel
characteristics in the digital world, thereby supporting precise and adaptive communication
decisions for 6G. In this article, we systematically review and summarize the existing efforts
in realizing the DTC, providing a comprehensive analysis of ray tracing (RT), artificial
intelligence (AI), and large model approaches. Based on this analysis, we further explore
the potential of integrating large models with RT methods. By leveraging the strong
generalization, multi-task processing capabilities, and multi-modal fusion capabilities of
large models while incorporating physical priors from RT as expert knowledge to guide
their training, there is a strong possibility of fulfilling the fast online inference and precise
mapping requirements of the DTC. Therefore, we propose a novel DeepRT-enabled DTC
(DRT-DTC) framework, which combines physical laws with large models like DeepSeek,
offering a new vision for realizing the DTC. Two case studies are presented to demonstrate
the possibility of this approach, which validate the effectiveness of physical law-based AI
methods and large models in generating the DTC.
Keywords: digital twin channel; ray tracing; artificial intelligence; deep ray tracing; large
models
1. Introduction
The sixth-generation (6G) communication system is expected to integrate sensing and
artificial intelligence (AI) capabilities to enable intelligent, hyper-reliable, and ubiquitous
connectivity, thereby enabling the Internet of Everything (IoE) [
1
]. The channel, serving as
the medium between the transmitter and the receiver, determines the performance limits
of wireless communication systems [
2
]. To fulfill the objectives of 6G, realistic, high-fidelity,
and efficient channel characterization is essential, leading to the emergence of the digital
twin channel (DTC) [
3
,
4
]. The DTC is a novel technology that provides digital replicas
of the entire process of channel fading states and variations in the physical world. This
capability allows for proactive adaptation to the evolving technologies and requirements of
wireless communication systems.
Ray tracing (RT), as a powerful physics-driven method for modeling wireless channel
characteristics [
5
,
6
], has been widely applied in channel modeling. For example, in [
7
],
Electronics 2025,14, 1849 https://doi.org/10.3390/electronics14091849

Electronics 2025,14, 1849 2 of 17
a hybrid modeling approach that combines RT methods with channel measurement is
presented for low-terahertz indoor communication, enhancing the overall accuracy of
the channel model. An RT-based approach for deterministic radio channel modeling
is proposed in [
8
], focusing on simulating the channel characteristics of reconfigurable
intelligent surfaces (RISs) in urban environments. The study in [
9
] also highlights the key
role of RT in constructing dynamic wireless environment models. Although RT has the
potential to offer detailed environment mapping and accurate simulation results, its high
computational complexity remains a significant challenge for implementing the DTC in the
dynamic, large-scale, and ever-changing environments of 6G systems.
With the rapid advancement of technology, AI has gained great attention in the com-
munications field due to its powerful analytical and generalization capabilities, enabling
reliable channel prediction and real-time communication decisions. Several works have
contributed to advancing the DTC through AI [
10
–
14
]. For example, the study in [
11
]
proposes an artificial neural network model that incorporates environmental parameters
for high-precision path loss prediction. An end-to-end neural network is designed based
on uplink pilots to predict downlink channel state information (CSI), avoiding additional
errors from uplink channel estimation [
13
]. All these works have shown the great potential
of AI to enable the DTC. However, many existing AI models are relatively small scale, often
designed for single tasks or specific scenarios. Moreover, most are purely data-driven meth-
ods, relying on large data while lacking physical constraints, which limits their ability to
meet the diverse, multi-task demands of the DTC. Recently, large models, such as DeepSeek
and OpenAI’s GPT series models, have risen to prominence due to their multi-task pro-
cessing abilities, efficient inference, and excellent generalization. Many researchers are
now exploring the application of these large models in the field of communications [
15
–
19
].
For example, the study in [
15
] builds a neural network for channel prediction based on the
pre-trained GPT-2, achieving great prediction performance on generalization tests with
low training and inference costs. A framework called WirelessLLM is proposed, which
leverages large models to address various tasks in wireless networks, such as channel
prediction and resource management, thus improving the intelligence and efficiency of
wireless systems [
16
]. The study in [
17
] introduces WirelessGPT, a pioneering foundation
model designed for multi-task learning in wireless communication and sensing. These
large models exhibit strong modeling capabilities, exceptional generalization, and efficient
processing in diverse, multi-task scenarios, offering a promising pathway for realizing
the DTC.
In this article, we first provide a systematic review of existing efforts and challenges
across different levels of channel twins, analyzing the roles and characteristics of RT, AI,
and large models in enabling the DTC, particularly in terms of fast online inference and
accurate channel prediction. Motivated by these insights, we explore the potential of
combining large models with RT methods. By leveraging the strong generalization, multi-
task processing capabilities, and multi-modal fusion capabilities of large models while
incorporating RT-based physical priors as expert knowledge to guide their training, a
promising trade-off between accuracy and efficiency can be achieved, further paving the
way toward the vision of the online and precise mapping requirements of the DTC. Based
on this, we propose a novel DeepRT-enabled DTC (DRT-DTC) framework, which combines
the physics laws of RT with large models such as Deepseek, offering a promising approach
to realize the DTC and addressing the identified challenges.
This article is organized as follows: Section 2introduces the DTC and its framework;
Section 3summarizes the existing work on different levels of channel twins and introduces
the promising framework of DRT-DTC; Section 4offers two examples of how to help realize

Electronics 2025,14, 1849 3 of 17
the DTC; Section 5discusses some open issues and future opportunities; and Section 6
gives the conclusion.
2. Framework of DTC
In this section, we introduce the DTC and then provide a detailed description of
its framework.
2.1. What Is the DTC?
The DTC utilizes digital twin technology to model and reconstruct the wireless channel
by creating a virtual representation of the physical environment. It is capable of accurately
reflecting the entire process of channel fading states and dynamic variations in the physical
world, thereby enabling it to provide intelligent decision-making for wireless commu-
nication tasks. Meanwhile, the DTC interacts continuously with real-world feedback,
dynamically updating its digital model and engaging in online learning to improve accu-
racy and better approximate actual channel conditions. As a result, it can provide more
reliable and adaptive communications decisions. To support autonomous 6G networks,
the DTC is expected to meet several key requirements, including precise mapping, with
real-time, self-updating, and task-oriented characteristics [3].
2.2. Framework
Building on the capabilities of the DTC and the predictive 6G network [
4
], a novel DTC
implementation framework is proposed, as shown in Figure 1. This framework follows the
paradigm of “sensing–prediction–decision and interaction”, with each functional module
designed to support intelligent and adaptive communication processes. The detailed
operations of each module are elaborated below.
Figure 1. The framework of the DTC, including data acquisition, sensing and reconstruction, channel
fading prediction, and communication decision [4].
•
Sensing: As noted in [
3
,
4
], two main types of information are required: environ-
mental information and channel information. In this module, multi-modal data are
collected from the physical world using various sensing devices such as cameras, radar,
and GPS. These devices capture environmental information, including the location,
size, and material of scatterers, as well as environmental maps. At the same time,
channel information, another type of critical information for the DTC, is obtained from

Electronics 2025,14, 1849 4 of 17
wireless communication devices. Then, the collected multi-modal and heterogeneous
data undergo denoising, feature extraction, and dimensionality reduction, enabling
an accurate reconstruction of the physical entities in the digital world. A summary of
related work is presented in Table 1.
Table 1. Different methods for environmental sensing and reconstruction.
Ref. Source Method Ability
[20]Environment sensing
devices
A sensing-enhanced radio
environment prediction
platform
Swiftly reconstruct the
physical environment
[21]Communication signaling
with camera image
Multi-user selection and
Multi-modal fusion
Increases the accuracy and
robustness of
environment reconstruction
[22] LiDAR point cloud data Pre-processing and
post-processing techniques
Real-time mapping for
digital twin development
[23] Radio frequency data Deep learning
Paving the way towards
lightweight and
scalable reconstructions
[24] Channel knowledge map
Combines physical
environment maps with
deep learning techniques
Constructing detailed
spatiotemporal maps of
channel characteristics
•
Prediction: The collected multi-modal data are further processed to explore and es-
tablish the relationship between environmental information and channel data. AI
techniques are employed to enable accurate channel prediction. For example, the wire-
less environment knowledge pool (WEKP) is proposed to explore the relationship
between the channel and the environment, thereby providing prior knowledge for
channel prediction [
4
]. An environment feature-based model is presented for predict-
ing path loss using random forest methods [25]. Similarly, the study in [26] proposes
a framework for millimeter-wave communication systems based on environmental
semantics, using AI methods to extract semantic information in image form for beam
and blockage prediction.
•
Decision and interaction: Communication decisions are made in the digital world for
the served users based on predicted channel data and then transmitted to the physical
world. Subsequently, the DTC engages in continuous interaction with the physical
environment via a feedback mechanism, iteratively updating its models and data. This
closed-loop process enhances the accuracy of digital–physical mapping and ensures
more reliable and adaptive decision-making.
3. DeepRT Enabled DTC
The evolution of the channel twin can be defined as five different levels, as mentioned
in [
3
]. In this section, we first provide a systematic overview of recent research efforts and
key challenges in realizing the DTC for 6G, with a particular focus on the latter three stages
of channel twin evolution: intermediate, advanced, and autonomous twins. Furthermore,
the promising DRT-DTC framework is introduced, which integrates RT-based physics laws
with AI methods to achieve a level-5 (L5) autonomous twin.
3.1. Traditional RT-Enabled Level-3 Intermediate Twin
RT [
27
], as a deterministic modeling method, simulates radio wave propagation
based on the provided environmental information and transmitter–receiver configura-
tions, thereby generating parameters such as path loss (PL), channel impulse response

Electronics 2025,14, 1849 5 of 17
(CIR), and CSI. As shown in Figure 2, the RT process primarily consists of environmental
construction, path-finding, electromagnetic computation, and channel prediction. The
environmental model is typically constructed manually or imported from sources like
Google Maps or OpenStreetMap. Additionally, the material properties of scatterers within
the scene need to be defined to facilitate the subsequent electromagnetic computations.
Figure 2. The framework of RT, including environmental construction, path-finding, electromagnetic
computation, and channel prediction.
Path-finding utilizes geometric optics (GO) and the uniform theory of diffraction
(UTD) [
28
] to simulate electromagnetic wave propagation. It includes two types of al-
gorithms: a direct algorithm and an inverse algorithm. The direct algorithm, known as
shooting and bouncing rays (SBR), generates rays from the transmitter and traces them
in all directions until they reach the receiver or exit the scene. When the rays encounter
objects, their new directions are calculated based on reflection or diffraction. The inverse
algorithm, based on Fermat’s principle, calculates the path between the transmitter and
receiver using image-based methods.
Once the propagation paths of the electromagnetic waves are determined, parameters
such as field strength at the receiver can be computed, including reflection and diffraction.
The reflection electromagnetic calculation is based on the Fresnel equations:
R∥=
ˆ
ε2
ˆ
ε1cos θi−qˆ
ε2
ˆ
ε1−sin2θi
ˆ
ε2
ˆ
ε1cos θi+qˆ
ε2
ˆ
ε1−sin2θi
(1)
R⊥=cos θi−qˆ
ε2
ˆ
ε1−sin2θi
cos θi+qˆ
ε2
ˆ
ε1−sin2θi
(2)
where
R∥,⊥
is the reflection coefficient,
θi
is the angle of incidence, and
ˆ
ε1
and
ˆ
ε2
are
the complex dielectric constants of the materials on either side of the reflective surface.
Diffraction electromagnetic computation is based on UTD theory, and the diffraction
coefficient Ds,his given as

Electronics 2025,14, 1849 6 of 17
Ds,h=−e−jπ
4
2n√2πksin β0
×cotπ+(ϕ−ϕ′)
2nFkLa+ϕ−ϕ′
+cotπ−(ϕ−ϕ′)
2nFkLa−ϕ−ϕ′
∓cotπ+(ϕ+ϕ′)
2nFkLa+ϕ+ϕ′
+cotπ−(ϕ+ϕ′)
2nFkLa−ϕ+ϕ′
(3)
where
ϕ
and
ϕ′
are the angles of the diffracted and incident rays, respectively, and
β0
is
the angle between the incident ray and the tangent to the edge. The
k
is the propagation
constant of light, and
n=
2
−α
π
, where
α
is the angle of the interior wedge.
L
is a distance
parameter, given by the following expression, and the function
F(X)
involves a Fresnel
integral and is given by F(X)=2j√XejX R∞
√Xe−jτ2dτ:
L=











ssin2β0for plane-wave incidence
rr′
r+r′for cylindrical-wave incidence
ss′
s+s′sin2β0for conical- and spherical-wave incidences
(4)
where
s
and
s′
are the distances along the diffracted and incident ray paths, respectively.
For a cylindrical wave with radius
r′
incident on the edge,
r
is the perpendicular distance
of the field point from the edge.
Based on the derived propagation paths and electromagnetic computations, the chan-
nel characteristics within the environment can be generated, including delay, PL, and CIR.
The accuracy of these results mainly depends on the fidelity of environment modeling,
the electromagnetic parameters of the materials, and the selected mechanisms for simulat-
ing wave propagation.
Traditional RT methods achieve accurate channel modeling by analyzing all pos-
sible electromagnetic wave paths in complex environments. They offer high precision,
strong adaptability across diverse scenarios, and robustness to frequency variations. More-
over, RT methods are well suited to support new features emerging in 6G channels, such
as integrated sensing and communication (ISAC) [
29
], extra-large-scale multiple-input
multiple-output (XL-MIMO) [
30
], and RIS [
31
], making it a promising research direction for
future channel modeling. However, in complex scenarios, the large number of face inter-
section tests significantly increases computation costs. Additionally, manually constructed
or imported environmental models often lack sufficient geometric and material details [
32
],
and the assigned electromagnetic parameters may not accurately reflect the properties of
real materials, leading to decreased accuracy. Therefore, traditional RT methods are suitable
for tasks with less stringent online requirements, such as network planning, management,
and optimization, and they are typically used to support level-3 (L3) twins in [3].
3.2. RT and AI Methods Enabled Level-4 Advanced Twin
In response to the high computational cost of traditional RT, various acceleration
strategies have been proposed, as shown in Table 2. For example, the study in [
33
] leverages
a kD-tree structure for the efficient processing of environmental geometry and utilizes
GPU-based parallel computing to reduce the path-finding time. Moreover, methods such
as simulated annealing (SA) [
34
] are employed to calibrate the electromagnetic parameters
of materials, thus enhancing the accuracy of electromagnetic computations.

Electronics 2025,14, 1849 7 of 17
In recent years, with the rapid development of AI, machine learning (ML) has been
increasingly applied to channel prediction, leading to performance improvements. These
methods can be broadly categorized based on their level of AI integration: one enhances
traditional RT by incorporating AI to address its limitations, while the other adopts an
end-to-end paradigm, replacing physical modeling with neural networks. Table 2lists
some related works. For example, the study in [
35
] uses neural networks to simulate
the interaction between rays and surfaces, improving the path-finding speed of the SBR
model. AI can also be applied to the automatic calibration of electromagnetic parameters.
In [
36
], gradient-based methods are used to calibrate the material properties, resulting
in models that combine neural networks (NNs) with physics-informed characteristics.
In environmental reconstruction, deep neural networks can generate 3D scenes with elec-
trical information from point cloud data [
37
]. Furthermore, a trained super-resolution
(SR) model is proposed to predict clusters and CIR [
38
]. Furthermore, NN models trained
on large datasets can predict the path loss based on input environmental information
and transceiver
positions [25,39,40],
without relying on RT algorithms. The study in [
41
]
employs an end-to-end convolutional neural network (CNN) to rapidly and accurately
generate radio maps, given the known environmental structure and transmitter location.
Table 2. Related work combining RT and AI.
Ref. Method Application Efficiency Limitation
RT optimization [33] GPU and kD-tree Path-finding Acceleration
The bottleneck of
the physical model,
non-task-oriented,
non-real time
RT dominance
[35] Neural rendering Path-finding Reducing path-finding
time
[36] Model-based NN Material calibration Automatic calibration
[37] Semantic segmentation 3D mapping High precision
AI dominance
[38] SR algorithm Channel data
generation
Generating HR
propagation data
Requiring data,
weak interpretability,
small scale
[39] PMNet PL prediction Predicting PL fast and
accurately
[40]Transformer-based
model PL prediction Adaptation to various
maps
[25]Environment
feature-based model PL prediction
Environment
information acquisition
and accurate prediction
[41] UNet Simulating radio maps Reducing run time and
improving accuracy
NN: neural network; SR: super-resolution; HR: high resolution; PMNet: path loss map prediction-oriented NN;
PL: path loss; UNet: U-shaped convolutional network.
With the assistance of AI, the efficiency and environmental adaptability of RT have
been improved, enhancing the accuracy of channel modeling and supporting level-4 (
L4
)
twins. However, the two approaches both have limitations. The former is based on physical
models, using fixed physical formulas that attempt to simulate the real-world radio propa-
gation phenomenon, but discrepancies between simulation and reality still lead to accuracy
bottlenecks. Although various acceleration strategies have been employed, online perfor-
mance in complex environments remains unsatisfactory. Moreover, the non-differentiable
nature of traditional RT makes it unable to self-adapt or optimize according to environmen-

Electronics 2025,14, 1849 8 of 17
tal or task changes, limiting its flexibility. The latter approach, being purely data-driven,
often neglects physical laws, requiring large volumes of high-quality data for training. It
also suffers from weak interpretability and operates as a “black box”. Furthermore, many
current AI models are small scale, task-specific, and lack generalization capabilities.
3.3. DeepRT Enabled Level-5 DTC
The core advantages of the DTC lie in its fast online inference and self-updating
abilities, which define it as an
L5
twin, known as an autonomous twin [
3
]. As mentioned
earlier, RT methods, as physical law-based models, have high computational complexity,
making it challenging for them to meet real-time processing requirements. In contrast, AI-
driven approaches offer improved efficiency through data-driven learning, but they often
struggle with generalization in diverse environments and still face accuracy limitations.
Thus, in
L5
, to enable the DTC to support self-sustaining and proactive online learning
wireless systems and unlock novel capabilities and applications, it is essential to establish
a data-model dual-driven AI-based network. In this context, large models offer greater
potential than conventional small-scale AI models, owing to their strong generalization
abilities, multi-task processing capabilities, and aptitude for multi-modal data fusion. By
integrating RT-derived physical priors as expert knowledge into the training process of
large models, it becomes possible to strike an effective balance between accuracy and
efficiency, thereby unlocking new capabilities and applications for the DTC in diverse
6G scenarios.
To achieve this, we propose a novel framework called DRT-DTC, which is inspired
by large-scale models such as DeepSeek [
42
–
44
] and is combined with physics-informed
methods, providing a promising solution for the autonomous twin.
3.3.1. Framework
The system framework is shown in Figure 3, including data acquisition, the DeepRT
network, and action decision.
Figure 3. The framework of DRT-DTC, including data acquisition, DeepRT network, and action decision.
1.
Data acquisition: Multi-modal data are collected from the physical world, including
both environmental information and channel information. These data are sourced
from base stations, environmental sensors, and user terminals. For example, 2D

Electronics 2025,14, 1849 9 of 17
cameras capture RGB images, depth cameras provide depth maps, and LiDAR or
3D cameras extract point clouds. Additionally, real-time transceiver configurations
(e.g., the location, frequency, bandwidth, and antenna) and multi-dimensional chan-
nel feedback (e.g., CIR, angle, delay, and power) are collected to support dynamic
prediction and DRT-DTC adaptation. These multi-modal data sources provide a rich
and accurate informational foundation for environment reconstruction and subse-
quent channel prediction. Through a sensor network and edge computing platforms,
real-time data streams are transmitted to the DeepRT network for further processing
and analysis.
2.
DeepRT network: The collected multi-modal data are processed and analyzed in
this module, and communication decisions are made. The process consists of the
following key steps:
•
Data processing: Data collected from various sources undergo cleaning, normal-
ization, and temporal processing to ensure quality and consistency. This step is
essential to make the data usable for subsequent analysis. Using multi-modal
fusion techniques, the processed data are integrated into a unified representa-
tion. This ensures that information from different modalities can be effectively
handled in the same space.
•
Multi-modal feature extraction: Specialized neural networks, such as deep neural
networks, convolutional neural networks, or long short-term memory networks,
are used to extract features from each data modality. The model automatically
learns high-dimensional features from the data, while self-attention mechanisms
are applied to dynamically adjust the weights of each modality feature, enhancing
the network’s capacity to capture complex relationships across heterogeneous
data sources.
•
Wireless environment knowledge pool: Inspired by RT methods, wireless envi-
ronment knowledge (WEK) can be constructed based on fundamental electro-
magnetic propagation principles to explore the mapping relationship between
the environment and the channel. The WEKP represents the integration and
extension of the WEK. The architecture of the WEKP is introduced in [
4
]. Based
on physical laws, the WEKP serves as a physical prior and provides expert
knowledge to the network.
•
Knowledge embedding: The WEKP is designed to serve as a source of structured
domain knowledge and physical constraints within the deep learning frame-
work. Rather than treating knowledge as auxiliary features, this conceptual
module aims to embed physically meaningful relationships, such as propagation
principles and environmental semantics, into the model’s latent space to guide
learning in a physics-consistent manner. Potential implementation approaches
include relation-aware attention [
45
,
46
], symbolic embeddings [
47
], or multi-
level alignment mechanisms [
48
] that map knowledge representations to the
internal layers of the model. By incorporating the WEKP in this manner, the
model is better constrained during training, mitigating issues common in purely
data-driven approaches, such as limited generalization and poor adaptability in
unfamiliar environments.
•
Mixture of experts (MoE) network: The processed multi-modal data are fed into
the MoE network, which can dynamically select the most appropriate expert net-
work based on the characteristics of the input data. This ensures that the network
chooses the optimal processing path for different types of data. By facilitating
the division of labor among multiple sub-networks, the MoE network effectively
handles heterogeneous data and adapts to varying environmental conditions.

Electronics 2025,14, 1849 10 of 17
• Hybrid-driven prediction module: This module first performs wireless channel
prediction by combining data-driven learning with physics-informed knowledge.
The predicted channel characteristics, such as path loss, delay spread, and spatial
parameters, serve as inputs for subsequent decision-making tasks, including
spectrum allocation, beamforming, and power control. By integrating both data-
driven and physical insights, the module provides more accurate and reliable
communication decisions.
3.
Action decision: The generated communication decisions are transmitted to physical
devices and interact with the real-world wireless network. Simultaneously, feedback
from the physical world is sent back to the DTC system through a feedback mech-
anism. Based on this, the model performs online learning, enabling self-updating
and optimization. This closed-loop learning cycle ensures that the communication
system can consistently make optimal decisions under dynamic environmental and
network conditions.
Overall, this framework establishes a closed-loop intelligent communication system
that integrates multi-modal perception, physics-informed learning, and adaptive decision-
making. By fusing data-driven insights with electromagnetic priors, the system enables
accurate and low-latency channel prediction, as well as robust decision-making in diverse
environments. Through the action decision and feedback mechanism, the model contin-
uously learns from real-world interactions and self-updates in real time. This ensures
long-term adaptability and optimal performance under dynamic network and environmen-
tal conditions, paving the way for fully autonomous 6G wireless systems.
3.3.2. Advantages
•
Strong modeling capability: The framework builds a data-model dual-driven net-
work by embedding physical knowledge as expert priors. This design enables the
model to learn complex features from data while maintaining consistency with real-
world electromagnetic behaviors, thereby enhancing both prediction accuracy and
physical consistency.
•
Multi-modal fusion: By leveraging large models’ capabilities in handling multi-
modal data, the framework integrates heterogeneous sources such as base stations,
sensors, and user terminals. Through unified representation and self-attention mecha-
nisms, it enables efficient cross-modal collaboration and the fusion of information.
•
Multi-scenario adaptability: The MoE network dynamically selects the most ap-
propriate expert model based on input data characteristics. This dynamic selection
empowers the system to adapt to various network environments and data types,
ensuring accurate communication decisions across diverse scenarios.
•
Online self-learning and optimization: Through feedback mechanisms and online
learning, the system performs self-optimization using real-world interaction feedback,
ensuring that communication decisions remain optimal under dynamic environments
and network conditions.
4. Validation
In this section, to validate the vision of DRT-DTC, which combines physical priors
with AI, we introduce two simulation results.
4.1. WEK-Based AI Methods for Channel Prediction
In this subsection, we present a WEK-based AI method for channel prediction.
The study in [
49
] builds a mapping relationship between environmental information
and electromagnetic wave propagation based on stochastic geometry and electromagnetic

Electronics 2025,14, 1849 11 of 17
propagation theory. This method quantifies the contributions of reflection, diffraction,
and blockage to the propagation process by leveraging the coordinate information of the
scatterers, transmitters, and receivers within the environment. The WEK is then constructed
and can be used for channel prediction. In this work, all datasets are generated based on
RT simulations from the Beijing University of Posts and Telecommunications and China
Mobile Communications Group DataAI-6G Dataset (BUPTCMCC-DataAI-6G Dataset) [
50
].
The dataset covers an area of 646 m × 290 m. Within this area, channel data are simulated
for a reception area of 59.5 m × 30.0 m, containing 7320 receivers arranged in 61 columns
of 120 devices. Taking one column of receivers as an example, the computed wireless
environmental knowledge is visualized in the form of a WEK spectrum, as shown in
Figure 4a.
(a)
(b)
Figure 4. WEK-based AI methods for channel prediction. (a) Wireless environment knowledge
spectrum for reflection, diffraction, and blockage; (b) CDF plots of the proposed WEK-based method
and the contrast methods [49].
As an example, the study in [
49
] uses path loss prediction to compare the performance
of different methods, including the unprocessed location data-based method, the environ-
ment feature-based method, and the proposed WEK-based method, using the RT-simulated

Electronics 2025,14, 1849 12 of 17
datasets as true values. In this setup, a convolutional neural network is employed for path
loss prediction. There are a total of 7320 data samples, each associated with a propaga-
tion knowledge matrix that includes three types of propagation: reflection, diffraction,
and blockage. To ensure a robust evaluation, 75% of the data are allocated for training
and 25% for testing, with the training and testing samples selected from non-overlapping
RX locations. The effectiveness of the WEK-based method is validated by comparing the
cumulative distribution function (CDF) plots of the predicted and true values. These CDF
plots are shown in Figure 4b. These results demonstrate that the proposed WEK-based
method exhibits a much more consistent trend with the true values than the contrast meth-
ods. This indicates that the WEK spectrum, as a physical prior provided to the AI model,
is more capable of capturing the entire path loss process and leads to a more accurate
representation of channel characteristics.
In addition, Table 3compares the three methods in terms of the normalized root mean
square error (NRMSE), root mean square error (RMSE), training time, and testing time.
The proposed WEK-based method achieves a significantly higher prediction accuracy,
with NRMSE reductions of 52.04% and 18.5% compared to the location data-based and
environment feature-based methods, respectively. In particular, the RMSE of the WEK-
based method is reduced to 2.020 dB, representing improvements of 2.192 dB and 1.379
dB over the respective baselines. While RT-based parameter generation takes about 20
minutes, the WEK enables predictions within milliseconds, reducing the computational
complexity by four orders of magnitude and making real-time communication prediction
feasible. These results demonstrate the effectiveness of integrating physical priors with AI
techniques, as proposed in this work.
Table 3. Performance comparison of path loss prediction obtained by the proposed WEK, the location
data-based, and the environment feature-based methods [49].
Methods NRMSE RMSE (dB) Training
Time (s)
Testing Time
(s)
Location data-based 0.565 4.212 8.21 0.227
Environment feature-based 0.456 3.399 7.67 0.170
Proposed WEK-based 0.271 2.020 4.07 0.004
4.2. Large Models to Enable DTC
In this article, we propose a new vision for realizing the DTC by integrating large
models. ChannelGPT [
51
] is a large model-driven DTC generator, which helps to validate the
effectiveness of incorporating large models into the DTC framework. In this work, an outdoor
urban scenario is constructed for dataset generation, covering an area of 200 m × 200 m
with four building groups and four roads. RT simulations are conducted with receivers
placed along the streets, resulting in rich multi-modal datasets that combine geometric, visual,
and wireless information. ChannelGPT performs channel prediction based on multi-modal
information and compares the normalized mean square error (NMSE) results with those
of two other methods: random pilot sampling without wireless environment information
(RS-WOWEI) and random pilot sampling with wireless environment information (RS-WWEI).
As shown in Figure 5, with the increase in training epochs, ChannelGPT demonstrates
a faster initial convergence speed than the other methods. Furthermore, it outperforms
other methods in terms of the NMSE, achieving the lowest NMSE throughout the training
process. Large models like ChannelGPT often exhibit greater variability during training
due to factors such as batch-to-batch variations, complex gradient updates, and the in-
terplay among diverse data modalities. These fluctuations, although more pronounced,
eventually stabilize as the model converges to a more optimal solution, resulting in su-
perior overall performance. Additionally, to assess the generalization capability of the

Electronics 2025,14, 1849 13 of 17
proposed ChannelGPT, all compared methods are trained on the original scenario and
tested on a newly constructed scenario with different building and vehicle distributions.
After being trained on the original scenario, the optimal models are re-evaluated in the new
scenario using the NMSE, as presented in Table 4. While the performance of RS-WWEI and
RS-WOWEI degrades significantly under the domain shift, ChannelGPT maintains stable
performance across both scenarios, demonstrating a promising generalization ability in
diverse wireless environments. This highlights the significant potential of large models
in optimizing channel prediction performance. Moreover, this suggests that integrating
large models into the DTC generation process has strong prospects for future wireless
communication systems.
Figure 5. The performance of ChannelGPT and other methods in terms of testing loss versus the
number of epochs and the NMSE [51].
Table 4. Comparison of the NMSE between the original and new scenarios for ChannelGPT and
other methods [50].
Methods NMSE (Original Scenario) NMSE (New Scenario)
RS-WOWEI 0.0893 0.1030
RS-WWEI 0.0320 0.0414
ChannelGPT 0.0128 0.0129
5. Challenges and Future Opportunities
Although the DRT-DTC framework presents promising potential in enabling the DTC
for 6G networks, it is still in its early stages, and several key challenges must be addressed
for its further research.
Large-scale multi-modal data collection: The DRT-DTC framework, which uses large
models, relies on high-quality, large-scale, and rich multi-modal datasets, incorporating
both environmental sensing data and channel data. Collecting and storing such comprehen-
sive data efficiently and cost-effectively remain a significant challenge, particularly given
the need to capture the characteristics of complex and dynamic wireless environments.
Additionally, synchronizing distributed devices for coordinated data collection poses fur-
ther difficulties, as ensuring consistency and timeliness across multiple data sources is
not trivial.
Collaboration between WEKP and large models: The integration of the WEKP with
large models is still at an early stage. Although it has been demonstrated that incorporat-
ing the WEKP into AI models can effectively enhance communication tasks, embedding

Electronics 2025,14, 1849 14 of 17
communication-specific knowledge into large models for online decision-making remains
a significant challenge. Current research has provided promising results, yet more research
is needed to explore the seamless collaboration between the WEKP and large models in
dynamic environments.
Hardware and energy consumption optimization: The computational demands
of training and running large models present significant challenges, especially in terms
of hardware capabilities and energy efficiency. Large models powering DRT-DTC require
substantial computational resources, which may not be readily available in current wireless
network infrastructures. Moreover, the energy consumption associated with these processes
is a critical concern, particularly in edge and mobile environments, where power constraints
are prevalent.
Despite the current challenges, the advanced technological foundations of 6G, such as
ISAC and distributed AI architectures, offer promising avenues to address these issues and
ultimately realize the potential of the DRT-DTC framework.
6. Conclusions
In this article, we review and analyze the current efforts and key challenges in realizing
the DTC for 6G networks, focusing on the evolution from
L3
to
L5
channel twins. We
provide an analysis of the characteristics of RT, AI, and large models in enabling the
DTC. While RT offers detailed physical modeling and accurate simulation results, its high
computational complexity limits its applicability in dynamic environments. AI offers strong
learning and generalization capabilities for real-time prediction, yet it struggles with multi-
tasking and adaptation in diverse scenarios. Large models show great potential in enabling
the DTC due to their strong generalization, multi-task processing, and multi-modal fusion
capabilities. Inspired by the strengths of these three methods, we propose the DRT-DTC
framework, which integrates physical priors from RT with the learning capabilities of
large models. This hybrid approach offers a promising solution for realizing the DTC and
effectively addresses the challenges identified. In addition, two case studies are presented
to demonstrate the possibility of this approach, which validate the effectiveness of physical
law-based AI methods and large models in generating the DTC. Finally, some open issues
and future opportunities related to the implementation of DRT-DTC are discussed.
Author Contributions: M.L. and T.W., conceptualization, methodology, writing—original draft,
and writing—review and editing. Z.D., X.L., Y.L., S.Z. and Z.W., investigation. Y.Z., conceptualization,
investigation supervision, and project administration. L.Y. and J.Z., investigation supervision. All
authors have read and agreed to the published version of the manuscript.
Funding: This research is supported by the Young Scientists Fund of the National Natural Science
Foundation of China (62201087, 62101069), National Key R&D Program of China (2023YFB2904803),
National Natural Science Foundation of China (62341128), and Beijing University of Posts and
Telecommunications China Mobile Research Institute Joint innovation Center.
Data Availability Statement: The data presented in this study are available within the cited arti-
cles [49,51].
Conflicts of Interest: The authors declare no conflicts of interest.
Abbreviations
The following abbreviations are used in this manuscript:
6G sixth generation
DTC digital twin channel
RT ray tracing

Electronics 2025,14, 1849 15 of 17
MIMO multiple-input multiple-output
RIS reconfigurable intelligent surface
IoE Internet of Everything
CSI channel state information
WEKP wireless environment knowledge pool
WEK wireless environment knowledge
PL path loss
CIR channel impulse response
GO geometric optics
UTD uniform theory of diffraction
SBR shooting and bouncing rays
ISAC integrated sensing and communication
XL-MIMO extra-large-scale multiple-input multiple-output
SA simulated annealing
ML machine learning
NN neural networks
SR super-resolution
MoE mixture of experts
NRMSE normalized root mean square error
RMSE root mean square error
CDF cumulative distribution function
NMSE normalized mean square error
References
1.
ITU. Framework and Overall Objectives of the Future Development of IMT for 2030 and Beyond; International Telecommunication
Union: Geneva, Switzerland, 2023. Available online: https://www.itu.int/rec/R-REC-M.2160-0-202311-I/en (accessed on 29
April 2025).
2. Molisch, A.F. Wireless Communications; John Wiley & Sons: Hoboken, NJ, USA, 2012; p. 34.
3.
Wang, H.; Zhang, J.; Nie, G.; Yu, L.; Yuan, Z.; Li, T.; Wang, J.; Liu, G. Digital Twin Channel for 6G: Concepts, Architectures and
Potential Applications. IEEE Commun. Mag. 2024,63, 24–30. [CrossRef]
4.
Wang, J.; Zhang, J.; Zhang, Y.; Sun, Y.; Nie, G.; Shi, L.; Zhang, P.; Liu, G. Radio Environment Knowledge Pool for 6G Digital Twin
Channel. IEEE Commun. Mag. 2025,63, 158–164. [CrossRef]
5.
Naruniranat, S.; Huang, Y.; Parsons, D. A Three-Dimensional Image Ray Tracing (3D-IRT) Method for Indoor Wireless Channel
Characterisation. In Proceedings of the MTT/ED/AP/LEO Societies Joint Chapter, United Kingdom and Republic of Ireland
Section, 1999 High Frequency Postgraduate Student Colloquium (Cat. No. 99TH8409), Leeds, UK, 17 September 1999; pp. 62–67.
[CrossRef]
6.
Torres, R.P.; Valle, L.; Domingo, M.; Loredo, S. An Efficient Ray-Tracing Method for Radiopropagation based on the Modified BSP
Algorithm. In Proceedings of the Gateway to 21st Century Communications Village, VTC 1999-Fall, IEEE VTS 50th Vehicular
Technology Conference (Cat. No. 99CH36324), Amsterdam, The Netherlands, 19–22 September 1999; pp. 1967–1971. [CrossRef]
7.
Chen, Y.; Li, Y.; Han, C.; Yu, Z.; Wang, G. Channel Measurement and Ray-Tracing-Statistical Hybrid Modeling for Low-Terahertz
Indoor Communications. IEEE Trans. Wirel. Commun. 2021,20, 8163–8176. [CrossRef]
8.
Pyhtilae, J.; Kokkoniemi, J.; Sangi, P.; Vaara, N.; Juntti, M. Ray Tracing Based Radio Channel Modelling Applied to RIS.
In Proceedings of the WSA & SCC 2023; 26th International ITG Workshop on Smart Antennas and 13th Conference on Systems,
Communications, and Coding, Braunschweig, Germany, 27 February 2023; pp. 1–6.
9.
Alkhateeb, A.; Jiang, S.; Charan, G. Real-Time Digital Twins: Vision and Research Directions for 6G and Beyond. IEEE Commun.
Mag. 2023,61, 128–134. [CrossRef]
10.
Jiang, W.; Schotten, H.D. Deep Learning for Fading Channel Prediction. IEEE Open J. Commun. Soc. 2020,1, 320–332. [CrossRef]
11.
Wu, L.; He, D.; Ai, B.; Wang, J.; Qi, H.; Guan, K.; Zhong, Z. Artificial Neural Network Based Path Loss Prediction for Wireless
Communication Network. IEEE Access 2020,8, 199523–199538. [CrossRef]
12.
Zhou, T.; Zhang, H.; Ai, B.; Xue, C.; Liu, L. Deep-Learning-Based Spatial-Temporal Channel Prediction for Smart High-Speed
Railway Communication Networks. IEEE Trans. Wirel. Commun. 2022,21, 5333–5345. [CrossRef]
13.
Fan, S.; Li, H.; Liang, X.; Liu, Z.; Gu, X.; Zhang, L. E2ENet: An End-to-End Channel Prediction Neural Network Based on Uplink
Pilot for FDD Systems. IEEE Wirel. Commun. Lett. 2024,13, 1285–1289. [CrossRef]

Electronics 2025,14, 1849 16 of 17
14.
Simmons, N.; Gomes, S.B.F.; Yacoub, M.D.; Simeone, O.; Cotton, S.L.; Simmons, D.E. AI-Based Channel Prediction in D2D Links:
An Empirical Validation. IEEE Access 2022,10, 65459–65472. [CrossRef]
15.
Liu, B.; Liu, X.; Gao, S.; Cheng, X.; Yang, L. LLM4CP: Adapting Large Language Models for Channel Prediction. J. Commun. Inf.
Netw. 2024,9, 113–125. [CrossRef]
16.
Shao, J.; Tong, J.; Wu, Q.; Guo, W.; Li, Z.; Lin, Z.; Zhang, J. WirelessLLM: Empowering Large Language Models towards Wireless
Intelligence. J. Commun. Inf. Netw. 2024,9, 99–112. [CrossRef]
17.
Yang, T.; Zhang, P.; Zheng, M.; Shi, Y.; Jing, L.; Huang, J.; Li, N. WirelessGPT: A Generative Pre-trained Multi-task Learning
Framework for Wireless Communication. arXiv 2025, arXiv:2502.06877. [CrossRef]
18.
Qiu, K.; Bakirtzis, S.; Wassell, I.; Song, H.; Zhang, J.; Wang, K. Large Language Model-Based Wireless Network Design. IEEE
Wirel. Commun. Lett. 2024,13, 3340–3344. [CrossRef]
19.
Chen, Z.; Zhang, Z.; Yang, Z. Big AI Models for 6G Wireless Networks: Opportunities, Challenges, and Research Directions. IEEE
Wirel. Commun. 2024,31, 164–172. [CrossRef]
20.
Miao, Y.; Zhang, Y.; Zhang, J.; Sun, Y.; Tian, Y.; Yu, L.; Liu, G. Demo Abstract: Predictive Radio Environment for Digital Twin
Communication Platform via Enhanced Sensing. In Proceedings of the IEEE INFOCOM 2023-IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS), Hoboken, NJ, USA, 20 May 2023; pp. 1–2. [CrossRef]
21.
Lin, B.; Zhao, C.; Gao, F.; Ye Li, G.; Qian, J.; Wang, H. Environment Reconstruction Based on Multi-User Selection and Multi-Modal
Fusion in ISAC. IEEE Trans. Wirel. Commun. 2024,23, 15083–15095. [CrossRef]
22.
Brock, E.; Huang, C.; Wu, D.; Liang, Y. Lidar-Based Real-Time Mapping for Digital Twin Development. In Proceedings of the 2021
IEEE International Conference on Multimedia and Expo (ICME), Shenzhen, China, 5–9 July 2021; pp. 1–6. [CrossRef]
23.
Khachatrian, H.; Mkrtchyan, R.; Raptis, T.P. Outdoor Environment Reconstruction with Deep Learning on Radio Propagation
Paths. In Proceedings of the 2024 International Wireless Communications and Mobile Computing (IWCMC), Ayia Napa, Cyprus,
27–31 May 2024; pp. 1498–1503. [CrossRef]
24.
Zeng, Y.; Chen, J.; Xu, J.; Wu, D.; Xu, X.; Jin, S.; Gao, X.; Gesbert, D.; Cui, S.; Zhang, R. A Tutorial on Environment-Aware
Communications via Channel Knowledge Map for 6G. IEEE Commun. Surv. Tutor. 2024,26, 1478–1519. [CrossRef]
25.
Sun, Y.; Zhang, J.; Zhang, Y.; Yu, L.; Yuan, Z.; Liu, G.; Wang, Q. Environment Features-Based Model for Path Loss Prediction.
IEEE Wirel. Commun. Lett. 2022,11, 2010–2014. [CrossRef]
26.
Yang, Y.; Gao, F.; Tao, X.; Liu, G.; Pan, C. Environment Semantics Aided Wireless Communications: A Case Study of mmWave
Beam Prediction and Blockage Prediction. IEEE J. Sel. Areas Commun. 2023,41, 2025–2040. [CrossRef]
27.
Yun, Z.; Iskander, M.F. Ray Tracing for Radio Propagation Modeling: Principles and Applications. IEEE Access 2015,3, 1089–1100.
[CrossRef]
28.
Kouyoumjian, R.G.; Pathak, P.H. A Uniform Geometrical Theory of Diffraction for an Edge in a Perfectly Conducting Surface.
Proc. IEEE 1974,62, 1448–1461. [CrossRef]
29.
Arnold, M.; Bauhofer, M.; Mandelli, S.; Henninger, M.; Schaich, F.; Wild, T.; ten Brink, S. MaxRay: A Raytracing-based Integrated
Sensing and Communication Framework. In Proceedings of the 2022 2nd IEEE International Symposium on Joint Communications
& Sensing (JC&S), Seefeld, Austria, 9–10 March 2022; pp. 1–7. [CrossRef]
30.
Schüßler, C.; Hoffmann, M.; Bräunig, J.; Ullmann, I.; Ebelt, R.; Vossiek, M. A Realistic Radar Ray Tracing Simulator for Large
MIMO-Arrays in Automotive Environments. IEEE J. Micro 2021,1, 962–974. [CrossRef]
31.
Liu, Y.; Sarris, C.D. Propagation Modeling for RIS-Enabled Channels Based on Ray-Tracing and the Equivalence Principle.
In Proceedings of the 2023 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting
(USNC-URSI), Portland, OR, USA, 23–28 July 2023; pp. 19–20. [CrossRef]
32.
Yun, Z.; Iskander, M.F.; Lim, S.Y. The Effect of 3D Building Reconstruction Errors on Propagation Prediction Using Geospatial Data
in Cyberspace. In Proceedings of the 2009 IEEE Antennas and Propagation Society International Symposium, North Charleston,
SC, USA, 1–5 June 2009; pp. 1–4. [CrossRef]
33.
Tan, J.; Su, Z.; Long, Y. A Full 3-D GPU-Based Beam-Tracing Method for Complex Indoor Environments Propagation Modeling.
IEEE Trans. Antennas Propag. 2015,63, 2705–2718. [CrossRef]
34.
Jemai, J.; Eggers, P.C.; Pedersen, G.F.; Kurner, T. Calibration of a UWB Sub-Band Channel Model Using Simulated Annealing.
IEEE Trans. Antennas Propag. 2009,57, 3439–3443. [CrossRef]
35.
Orekondy, T.; Kumar, P.; Kadambi, S.; Ye, H.; Soriaga, J.; Behboodi, A. WiNeRT: Towards Neural Ray Tracing for Wireless Channel
Modelling and Differentiable Simulations. In Proceedings of the Eleventh International Conference on Learning Representations
(ICLR 2023), Kigali, Rwanda, 1–5 May 2023.
36.
Hoydis, J.; Aoudia, F.A.; Cammerer, S.; Euchner, F.; Nimier-David, M.; Brink, S.T.; Keller, A. Learning Radio Environments by
Differentiable Ray Tracing. IEEE Trans. Mach. Learn. Commun. Netw. 2024,2, 1527–1539. [CrossRef]
37.
Zhang, F.; Zhou, C.; Brennan, C.; Wang, R.; Li, Y.; Xia, G.; Zhao, Z.; Xiao, Y. A Radio Wave Propagation Modeling Method Based
on High-Precision 3-D Mapping in Urban Scenarios. IEEE Trans. Antennas Propag. 2024,72, 2712–2722. [CrossRef]

Electronics 2025,14, 1849 17 of 17
38.
He, D.; Guan, K.; Yan, D.; Yi, H.; Zhang, Z.; Wang, X.; Zhong, Z.; Zorba, N. Physics and AI-Based Digital Twin of Multi-Spectrum
Propagation Characteristics for Communication and Sensing in 6G and Beyond. IEEE J. Sel. Areas Commun. 2023,41, 3461–3473.
[CrossRef]
39.
Lee, J.H.; Molisch, A.F. A Scalable and Generalizable Pathloss Map Prediction. IEEE Trans. Wirel. Commun. 2024,23, 17793–17806.
[CrossRef]
40.
Hehn, T.M.; Orekondy, T.; Shental, O.; Behboodi, A.; Bucheli, J.; Doshi, A.; Namgoong, J.; Yoo, T.; Sampath, A.; Soriaga, J.B.
Transformer-Based Neural Surrogate for Link-Level Path Loss Prediction from Variable-Sized Maps. In Proceedings of the 2023
IEEE Global Communications Conference (GLOBECOM 2023), Kuala Lumpur, Malaysia, 4–8 December 2023; pp. 4804–4809.
[CrossRef]
41.
[CrossRef] Levie, R.; Yapar, Ç.; Kutyniok, G.; Caire, G. RadioUNet: Fast Radio Map Estimation with Convolutional Neural
Networks. IEEE Trans. Wirel. Commun. 2021,20, 4001–4015. [CrossRef]
42.
Guo, D.; Zhu, Q.; Yang, D.; Xie, Z.; Dong, K.; Zhang, W.; Chen, G.; Bi, X.; Wu, Y.; Li, Y.; et al. DeepSeek-Coder: When the Large
Language Model Meets Programming—The Rise of Code Intelligence. arXiv 2024, arXiv:2401.14196. [CrossRef]
43.
Bi, X.; Chen, D.; Chen, G.; Chen, S.; Dai, D.; Deng, C.; Ding, H.; Dong, K.; Du, Q.; Fu, Z.; et al. Deepseek LLM: Scaling Open-Source
Language Models with Longtermism. arXiv 2024, arXiv:2401.02954. [CrossRef]
44.
Lu, H.; Liu, W.; Zhang, B.; Wang, B.; Dong, K.; Liu, B.; Sun, J.; Ren, T.; Li, Z.; Yang, H.; et al. Deepseek-VL: Toward Real-World
Vision-Language Understanding. arXiv 2024, arXiv:2403.05525. [CrossRef]
45.
Cai, W.; Ma, W.; Zhan, J.; Jiang, Y. Entity Alignment with Reliable Path Reasoning and Relation-Aware Heterogeneous Graph
Transformer. arXiv 2022, arXiv:2205.08806. [CrossRef]
46.
Sheikh, N.; Qin, X.; Reinwald, B.; Miksovic, C.; Gschwind, T.; Scotton, P. Knowledge Graph Embedding Using Graph Convolu-
tional Networks with Relation-Aware Attention. arXiv 2021, arXiv:2102.07200. [CrossRef]
47.
Xie, Y.; Xu, Z.; Kankanhalli, M.S.; Meel, K.S.; Soh, H. Embedding Symbolic Knowledge into Deep Networks. In Proceedings of
the 33rd International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 8 December 2019; pp. 1–11.
48.
Sun, Z.; Wang, C.; Hu, W.; Chen, M.; Dai, J.; Zhang, W.; Qu, Y. Knowledge Graph Alignment Network with Gated Multi-Hop
Neighborhood Aggregation. AAAI Conf. Artif. Intell. 2020,34, 222–229. [CrossRef]
49.
Wang, J.; Zhang, J.; Sun, Y.; Zhang, Y.; Yu, L.; Zhang, Z. Electromagnetic Wave Property Inspired Radio Environment Knowledge
Construction and Artificial Intelligence Based Verification for 6G Digital Twin Channel. Front. Inform. Technol. Electron. Eng. 2025,
26, 260–277. [CrossRef]
50.
Shen, Z.; Yu, L.; Zhang, Y.; Zhang, J.; Zhang, Z.; Hu, X.; Han, S.; Jin, J.; Liu, G. DataAI-6G: A System Parameters Configurable
Channel Dataset for AI-6G Research. In Proceedings of the 2023 IEEE Globecom Workshops (GC Wkshps), Kuala Lumpur,
Malaysia, 4–8 December 2023; pp. 1910–1915. [CrossRef]
51.
[CrossRef] Yu, L.; Shi, L.; Zhang, J.; Wang, J.; Zhang, Z.; Zhang, Y.; Liu, G. ChannelGPT: A Large Model to Generate Digital Twin
Channel for 6G Environment Intelligence. arXiv 2024, arXiv:2410.13379. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.