Content area
The anticipated transition to sixth-generation (6G) wireless systems is set to redefine how network resources are managed in environments characterized by vast device heterogeneity, stringent latency requirements, and increased autonomy at the network edge. As centralized control paradigms struggle to keep pace with these demands, there is a growing need for adaptive, decentralized solutions that can make intelligent decisions in real time. In this study, we propose a new architecture that integrates federated learning (FL) with digital twin (DT) technologies to improve the responsiveness and efficiency of resource management in edge-enabled 6G networks. Our approach enables edge nodes to collaboratively train machine learning models without the need to share raw data, thereby preserving privacy and reducing communication overhead. These local models contribute to a central digital twin—a virtual replica of the network environment—that continuously evolves to reflect real-time operational states and predict system behavior. Within this framework, the digital twin enables dynamic optimization across multiple domains, including spectrum distribution, computation task offloading, and energy balancing, by leveraging insights generated from the distributed FL models. Simulation results across varied 6G scenarios reveal that the proposed system offers considerable improvements in network performance metrics, such as reduced latency, higher resource utilization, and enhanced scalability under high device density. The hybridization of FL and DT within a unified architecture demonstrates a viable pathway toward autonomous, self-optimizing network infrastructures, aligning with the envisioned capabilities and challenges of future telecommunications ecosystems.
Introduction
The telecommunications industry is undergoing a transformative evolution with the advent of sixth-generation (6G) wireless networks. Envisioned as the successor to 5G, 6G is expected to enable unprecedented levels of connectivity, intelligence, and autonomy, driven by a confluence of technologies including artificial intelligence (AI), edge computing, terahertz (THz) communications, blockchain, and quantum-safe cryptography. The main goals of 6G are not only to enhance the data rates and reduce latency but also to provide dynamic, context-aware, and adaptive services that support emerging applications such as holographic communication, remote surgery, immersive virtual reality, and autonomous systems [1, 2].
A critical challenge in achieving these ambitious goals lies in resource management—specifically, the real-time orchestration of communication, computation, and energy resources across increasingly complex and heterogeneous networks. With billions of connected devices anticipated, traditional centralized approaches are no longer sustainable due to scalability bottlenecks, privacy concerns, and latency constraints [3]. Moreover, the network edge, populated by IoT devices, sensors, autonomous vehicles, and unmanned aerial vehicles (UAVs), demands distributed intelligence capable of responding to localized events while cooperating within a global network context [4].
Motivation
To address these emerging demands, this paper explores a novel integration of federated learning (FL) and digital twin (DT) technologies. FL has gained traction as a decentralized machine learning paradigm where models are trained collaboratively across multiple devices or nodes, without transmitting raw data to a central server. This approach inherently supports data privacy, lowers communication costs, and enables edge intelligence [5, 6]. However, FL alone lacks the environmental awareness and system-wide visibility needed to coordinate resources across diverse network segments.
This is where the digital twin concept becomes crucial. A digital twin is a continuously updated digital representation of physical systems that can simulate, predict, and optimize their behavior in real time. Originally developed in industrial engineering, DTs are increasingly being applied in smart cities, healthcare, and manufacturing [7]. In the context of 6G networks, a DT can act as a centralized, virtual model that consolidates updates from distributed edge nodes, providing a holistic view of network status and facilitating predictive decision-making.
The integration of FL and DT into a unified framework is a novel and underexplored approach. While prior works have discussed FL for edge learning and DTs for simulation or system modeling, few have proposed a tightly coupled architecture that combines the predictive power of DTs with the decentralized intelligence of FL for adaptive resource management in 6G edge networks. This paper aims to fill that gap.
Problem statement
Modern 6G edge networks are expected to dynamically manage vast resources across multiple domains—bandwidth allocation, task offloading, caching, and energy use—while meeting the ultra-reliable low-latency communication (URLLC) requirements of mission-critical applications [8, 9]. The inherent challenges include:
Data privacy and heterogeneity: Devices generate sensitive data with varying formats, which cannot be directly centralized without privacy risks.
Model generalization: Traditional AI models often fail to adapt to varying local contexts in distributed settings.
Real-time coordination: High-speed decisions are needed for spectrum usage, energy efficiency, and user mobility, which traditional methods struggle to support.
System scalability and resilience: Networks must remain functional and efficient under high traffic volumes, node failures, and cyber threats.
In this paper, we propose a framework that uses federated learning to locally adapt to edge environments, while a digital twin synchronizes and guides global resource orchestration based on insights from those distributed models.
Contributions
This paper makes the following key contributions:
Framework design: We propose a novel architectural framework that tightly integrates federated learning with a dynamic digital twin model tailored for 6G edge networks.
Resource optimization mechanisms: The framework introduces multi-level optimization strategies for spectrum allocation, computation offloading, and energy balancing using predictive analytics from the digital twin.
Mathematical modeling and simulation: We formulate the optimization problem mathematically and simulate its performance across 6G-like scenarios using NS-3 and MATLAB-based environments.
Performance evaluation: Our results demonstrate significant improvements in average network latency, model convergence time, and resource utilization compared to traditional centralized and standalone FL-based systems.
Scalability and resilience: The proposed system is evaluated for robustness under device churn and communication failures, showing effective adaptation and system recovery.
Relevance to 6G research and practice
The convergence of federated learning and digital twin technologies within telecommunications aligns closely with national and international initiatives to build secure, intelligent, and decentralized communication infrastructures. Entities such as the U.S. National Science Foundation (NSF), National Telecommunications and Information Administration (NTIA), and the European 6G flagship program have all emphasized the need for AI-driven, resilient, and context-aware networks [10, 11]. Our proposed framework contributes directly to these goals by enhancing local autonomy without compromising system-wide efficiency.
Furthermore, the research supports ongoing industry efforts to enable real-time closed-loop control systems, where insights from the digital twin can inform immediate changes in edge network behavior. These capabilities are crucial for future services such as autonomous mobility, remote diagnostics, and critical infrastructure management.
Related work
The integration of federated learning and digital twin technologies into next-generation telecommunications is an emerging field, but foundational components have been explored extensively in related domains. This section reviews the most relevant work across four intersecting themes: (1) federated learning in wireless and edge networks, (2) digital twin applications in networked systems, (3) adaptive resource management in 5G and emerging 6G systems, and (4) the convergence of AI, edge computing, and network virtualization.
Federated learning in wireless and edge networks
Federated learning (FL) was originally introduced to enable collaborative training of machine learning models on edge devices without centralized data aggregation, thereby preserving data privacy and reducing communication overhead [12]. Its relevance to wireless networks was soon recognized, especially in scenarios where data is generated at distributed nodes—such as smartphones, sensors, and vehicular networks.
Bonawitz et al. [13] proposed a secure FL protocol for mobile devices, highlighting issues such as model drift, synchronization, and communication cost. In the context of wireless systems, Niknam et al. [14] studied FL in mobile edge computing (MEC) environments, optimizing bandwidth and computation resources while maintaining model accuracy. Their work revealed that FL could outperform centralized models in latency-sensitive scenarios, though scalability and client reliability remained challenges.
Recent work by Wang et al. [15] extended FL to vehicular networks (V2X), demonstrating its capability to support decentralized learning with dynamic topology changes. However, they noted a decline in model performance in highly mobile or adversarial settings, leading researchers to consider hybrid models that incorporate real-time feedback from the network environment—a precursor to digital twin integration.
Although FL addresses privacy and bandwidth constraints effectively, it does not inherently capture the broader network state or enable predictive analytics beyond local environments. Thus, the concept of coupling FL with an overarching network model, such as a digital twin becomes essential for adaptive global optimization.
Digital twin applications in communication networks
Digital twin (DT) technology—originally developed for manufacturing and aerospace—has gained traction in smart city planning, infrastructure monitoring, and more recently, communication systems. A digital twin is a virtual replica of a physical asset, process, or system that mirrors real-time changes and can be used for predictive analysis, fault diagnosis, and optimization [16].
In the context of telecommunications, Li et al. [17] introduced a preliminary DT architecture for 5G networks, demonstrating how real-time network telemetry could be mapped onto a virtual simulation layer for system diagnosis and proactive maintenance. Similarly, Zhang et al. [18] proposed using DTs for end-to-end network visibility and fault prediction in cloud-native telecom environments.
More advanced work by Xu et al. [19] suggested integrating DTs into the 6G design phase to enable joint modeling of spectrum, traffic patterns, and device mobility. However, these approaches typically rely on centralized data processing and lack the distributed intelligence required at the network edge, where rapid local decisions are necessary.
A major limitation in existing DT-based telecom systems is their static nature and dependence on complete, often delayed, datasets. Our work addresses this by combining DTs with FL, allowing the DT to evolve in real time based on distributed insights, rather than relying solely on central telemetry.
Adaptive resource management in 5G/6G networks
Resource management in wireless networks involves dynamically allocating radio spectrum, computational capacity, and energy to maximize system performance and meet service-level agreements (SLAs). In 5G networks, these functions are typically handled by centralized controllers or software-defined networking (SDN) architectures [20].
With the move toward 6G, more decentralized and AI-driven models are being considered. Saad et al. [21] outlined a vision for 6G that included intelligent agents at the edge capable of real-time learning and adaptation. Their work emphasized the need for decentralized orchestration mechanisms to support massive device connectivity and ultra-low-latency requirements.
Moreover, Chen et al. [22] presented a deep reinforcement learning (DRL) model for joint task offloading and energy management in MEC systems. While DRL showed promise in handling dynamic environments, the training overhead and instability of convergence limited its real-world applicability in federated settings.
Challita et al. [23] highlighted that edge intelligence could benefit from collaborative learning among devices to balance trade-offs between resource use and service quality. However, they noted that existing models often assume static or predictable user behavior, which limits their generalizability in real-time mobile networks. This limitation motivates the use of predictive DT models that are continuously updated from real-world data, as proposed in our framework.
Convergence of edge AI, FL, and network virtualization
The convergence of AI, edge computing, and network virtualization is paving the way for intelligent, autonomous network management. Network slicing, for instance, allows operators to allocate virtual resources to different applications based on service requirements. When combined with edge AI and federated learning, slicing can be enhanced to become adaptive and self-optimizing [24].
Park et al. [25] demonstrated that FL could be used to manage resource slices in a multi-access edge computing (MEC) environment by learning user behavior patterns and network load dynamics. Nevertheless, their approach lacked a mechanism for global coordination, leading to potential inefficiencies during network congestion or node failures.
Research by Hassan et al. [26] proposed a framework combining software-defined networking (SDN) and FL to orchestrate virtualized network functions. They showed improvements in service latency and resource allocation efficiency, but did not incorporate a predictive or modeling layer to anticipate system behavior—something a DT could provide.
The integration of digital twins into such AI-based orchestration frameworks is still in its infancy. By aligning FL and DT technologies within a 6G-specific context, our work offers a novel, scalable, and intelligent solution to the resource management problem.
Methodology
This section introduces the architecture and mathematical foundation for the proposed Federated Learning-Driven Digital Twin (FL-DT) Framework, designed for real-time, intelligent resource management in 6G edge networks. The system combines decentralized federated learning at edge nodes with a predictive digital twin model to optimize computation offloading, bandwidth allocation, and energy efficiency. Figure 1 illustrates the overall operational pipeline of the proposed FL-DT framework. The architecture integrates decentralized model training at edge nodes with a central digital twin that performs forecasting and global optimization. Edge devices locally train FL models using their private datasets, transmitting only model updates to the twin. The twin aggregates global states, predicts system behavior, and issues resource control commands. This closed feedback loop ensures adaptive orchestration of bandwidth, computation, and energy across heterogeneous nodes.
[See PDF for image]
Fig. 1
Conceptual architecture of the Federated Learning–Driven Digital Twin (FL-DT) Framework for adaptive resource management in 6G edge networks
We begin with an overview of the architectural layers, followed by detailed formulations for the federated learning process, digital twin modeling, spectrum management, task scheduling, and energy optimization.
System model and assumptions
We consider a 6G-enabled heterogeneous network composed of:
A set of edge nodes .
A global digital twin orchestrator hosted in a cloud or core server.
A set of user equipment (UE) devices .
A total system bandwidth budget and computation power budget .
Each edge node operates in three layers:
Learning layer: Trains a model using local data through FL.
Digital twin layer: Maintains a synchronized replica of local state .
Optimization layer: Executes decisions based on control commands from the global digital twin.
The network operates in discrete time slots with synchronization intervals . We assume each UE can offload tasks to nearby edge nodes depending on wireless conditions and load.
The system follows a hierarchical model in which local intelligence is embedded within edge nodes, while global coordination is handled by the digital twin. Each node maintains local estimators of CPU load, queue occupancy, and link quality. The digital twin, residing in the core, aggregates these statistics to maintain a holistic representation of the network. The synchronization interval controls how frequently updates are exchanged. This multi-layered structure allows near-real-time adaptation while keeping communication overhead manageable.
Federated learning modeling
The global learning objective is to minimize a distributed empirical loss function:
1
Each edge node performs local gradient updates:
2
and periodically transmits weights to a global aggregator that applies Federated Averaging:
3
.We also model convergence behavior under communication constraints. Define as the model drift. The aggregated drift across all nodes affects convergence:
4
.To stabilize convergence under noisy updates, we include adaptive learning rates:
5
. where is a variance penalty factor.The convergence analysis assumes bounded gradient variance and Lipschitz continuity of local objectives. Under these assumptions, the expected deviation of the global model satisfies:
6
where is the adaptive learning rate and is the strong convexity parameter.Digital twin modeling and synchronization
Each edge node maintains a local state vector:
7
where:
= bandwidth,
= CPU frequency,
= power level,
= incoming service rate,
= task queue length.
The global digital twin aggregates:
8
and performs forecasting of future system states:
9
Each forecast is used in the resource optimization problem as a predicted constraint. The twin also performs state normalization:
10
.to ensure consistent cross-node comparison. Here, and are historical mean and standard deviation vectors.
The DT’s forecasting engine employs a recurrent model (e.g., LSTM-based predictor) with input features and outputs next-slot demand . The prediction error is iteratively minimized using:
11
Spectrum allocation optimization
We allocate a limited bandwidth among edge nodes. Let denote the bandwidth assigned to .
Objective:
Maximize utility based on Shannon capacity approximation:
12
Subject to:
13
We solve this convex optimization using the Karush-Kuhn-Tucker (KKT) conditions. The Lagrangian is:
14
Optimal is:
15
Computation offloading model
Each task has data size and computation requirement be a binary variable indicating whether task is offloaded to node .
Delay model:
16
.We minimize system delay:
17
.Subject to:
18
This forms an integer linear program (ILP), solved using branch-and-bound or Lagrangian relaxation.
Energy optimization
We define power consumption at node as:
19
. where:= hardware energy coefficient,
= transmission energy cost.
We minimize total energy:
20
.Subject to delay and resource constraints:
21
This is a quadratically constrained mixed-integer program (QC-MIP), typically solved using heuristic or dual decomposition methods.
Joint optimization and coordination
The complete control cycle at intervals includes:
Edge nodes send and to the digital twin.
DT forecasts and solves:
22
.subject to:
23
.Output control vector:
24
sent to edge node .
Model and state differences:
25
.used to decide whether to trigger re-training or control updates.
FL–DT coordination workflow
The overall workflow operates cyclically in three main phases:
(1) Local training and state observation
Each edge node trains its FL model on private data while measuring local state .
(2) Global forecasting and optimization
The digital twin collects state updates, forecasts future demand , and solves an optimization problem minimizing the joint cost of delay and energy consumption.
(3) Feedback and actuation
Control decisions are transmitted back to edge nodes to adjust their configurations. This cyclical process repeats every time slots, forming a self-learning feedback loop.
This workflow ensures closed-loop coordination between predictive global intelligence and localized adaptation.
Algorithm overview
Algorithm 1
Federated learning-driven digital twin optimization process.
Initialization:
For each edge node , initialize the following variables:
: Initial local model weights.
: Initial local state vector .
: Initial allocated bandwidth.
: Initial CPU frequency.
For each time slotto:
At each edge node:
Train the local federated learning model using local dataset to obtain updated weights:
26
Measure current local state variables (bandwidth usage, queue length, task load) and update the state vector:
27
Send and to the digital twin orchestrator.
At the digital twin (DT) orchestrator:
Forecast the expected service load for each edge node using historical traffic data:
28
Solve the joint optimization problem to determine optimal resource allocation:
29
where is the multi-objective cost function described in Sect. Joint optimization and coordination.Send control decisions to the corresponding edge nodes.
At each edge node:
Receive control signals from the digital twin.
Update local configurations accordingly:
30
.Convergence Checks (every slots):
Compute model weight change:
31
.Compute system state deviation:
32
If both fall below convergence thresholds and , halt or reduce update frequency.
Results
In this section, we present the simulation results and analytical evaluations of our proposed Federated Learning-Driven Digital Twin (FL-DT) Framework for adaptive resource management in 6G edge networks. The experiments are designed to validate the effectiveness of the framework in terms of bandwidth utilization, offloading delay, energy efficiency, CPU frequency adaptation, and overall latency reduction.
The simulations were performed using Python-based emulation, incorporating simplified but realistic models of wireless communication, task scheduling, and federated optimization. Key performance metrics are visualized using five primary figures.
Simulation setup and parameters
To ensure reproducibility, we provide the full details of our simulation environment. The experiments were implemented in Python 3.10 with TensorFlow 2.10 for federated learning, and NS-3 (v3.35) for network emulation. The simulation topology consisted of 10 edge nodes and 20 user equipment (UEs) randomly distributed in a 500 m × 500 m area. Each edge node had a maximum CPU frequency of 2.5 GHz and power budget of 15 W. UEs generated computation tasks following a Poisson arrival process with mean arrival rate λ = 5 tasks/s. Task sizes were exponentially distributed with mean 1 MB, and computational demands followed a truncated normal distribution with mean 0.5 Giga-cycles and variance 0.1. Wireless links used a standard Rayleigh fading model with path loss exponent 3.5, noise power density − 174 dBm/Hz, and system bandwidth B = 20 MHz.
The federated learning model used a two-layer neural network with 128 hidden units per layer, ReLU activation, and cross-entropy loss. Hyperparameters were: learning rate = 0.01, batch size = 32, local epochs per round = 5, optimizer = Adam. Synchronization interval Tsync was set to 5 slots. Random seeds (123, 321, 999, 2025, 777) were used for robustness.
A summary of key parameters is provided in Table 1.
Table 1. Simulation parameters for reproducibility
Parameter | Value |
|---|---|
# Edge nodes (N) | 10 |
# User Equipments (M) | 20 |
Area size | 500 m × 500 m |
Bandwidth B | 20 MHz |
CPU max | 2.5 GHz |
Power budget | 15 W |
Channel model | Rayleigh fading, path loss 3.5 |
Noise density | −174 dBm/Hz |
Task arrival | Poisson (λ = 5 tasks/s) |
Task size | Exponential(mean 1 MB) |
Computation demand | Normal(0.5 Gcycles, var = 0.1) |
Learning rate | 0.01 |
Batch size | 32 |
Local epochs | 5 |
Optimizer | Adam |
Synchronization interval Tsync | 5 slots |
Random seeds | {123, 321, 999, 2025, 777} |
Bandwidth allocation efficiency
To evaluate dynamic spectrum distribution under the FL-DT framework, we tracked the allocated bandwidth to each edge node across 20 consecutive time slots. Figure 2 illustrates the temporal-spatial variation in bandwidth allocation across edge nodes, using a heatmap format. Nodes experiencing high traffic are clearly identifiable by the warmer color bands during specific time slots.
[See PDF for image]
Fig. 2
Heatmap showing dynamic bandwidth allocation across 10 edge nodes over 20 time slots
The x-axis represents time slots (unit: discrete time units, each approximately 10 ms), the y-axis represents edge node IDs (1 to 10), and the color scale indicates allocated bandwidth (unit: MBps, ranging from 0 to 50 MBps).
The system demonstrates a consistent pattern of adjustment, with higher bandwidth allocations to nodes experiencing peak traffic and minimal reallocation delays. Node 5 and Node 8, for instance, receive higher bandwidth during peak loads, supporting timely task processing. This validates the effectiveness of digital twin-based forecasting in anticipating edge node demands and reallocating bandwidth accordingly.
This adaptive allocation is driven by the digital twin’s forecasting of traffic patterns, ensuring that resources are proactively shifted to high-demand nodes, which minimizes bottlenecks and improves overall network throughput by up to 25% in peak periods.
Offloading delay minimization
One of the critical challenges in 6G networks is minimizing the task offloading latency. Under our framework, task scheduling is dynamically adjusted based on the twin’s forecast and federated insights. Figure 3 summarizes the average delay observed per edge node across all simulation time slots.
[See PDF for image]
Fig. 3
Average Offloading Delay Per Edge Node
The x-axis represents edge node IDs (1 to 10), and the y-axis represents average delay (unit: ms).
Most nodes maintained an average offloading delay below 50 ms, which is well within acceptable thresholds for URLLC applications. Notably, Node 3 experienced a higher average delay, which was attributable to intermittent congestion and lower available CPU cycles—a condition automatically mitigated in later cycles via FL-based reallocation.
The low delays across most nodes highlight the effectiveness of FL’s local adaptations combined with DT’s global coordination, allowing for efficient task distribution that aligns with URLLC requirements and reduces processing queues by dynamically offloading to underutilized nodes.
CPU frequency adaptation
Dynamic CPU frequency tuning at edge nodes plays a pivotal role in managing energy efficiency and computation latency. Figure 4 presents a box plot that summarizes CPU frequency variation across nodes. Nodes 3 and 7 exhibit greater frequency spread, indicating fluctuating computational demands during peak service intervals.
The FL-DT system was able to throttle CPU frequencies between 1.0 GHz and 2.5 GHz depending on task loads and energy constraints. Peak frequencies were observed during high-intensity load intervals (e.g., time slots 7–12), while low-demand periods triggered down-throttling to conserve power. This dynamic adaptability illustrates the twin’s role in predictive energy balancing.
This variation demonstrates how the framework’s energy optimization mechanisms, informed by DT predictions, enable real-time throttling to balance computational demands with power constraints, resulting in an average 15% reduction in unnecessary high-frequency operations.
Energy consumption trends
Figure 5 illustrates the smoothed average energy consumption across edge nodes over time, with a shaded variability band representing ± 1 standard deviation. The gradual decline reflects the system’s adaptive energy optimization through predictive scheduling and learning convergence.
The observed decline underscores the learning convergence of FL models, which refines predictions over time, coupled with DT’s workload consolidation, leading to more efficient energy use and fewer wasteful retransmissions.
[See PDF for image]
Fig. 4
Box plot summarizing CPU frequency distribution for each edge node across all time slots
[See PDF for image]
Fig. 5
Smoothed average energy consumption across all edge nodes over 20 time slots
In Fig. 4, the x-axis represents edge node IDs (1 to 10), and the y-axis represents CPU frequency (unit: GHz, ranging from 1.0 to 2.5 GHz).
In Fig. 5, the x-axis represents time slots (unit: discrete time units, each approximately 10 ms), the y-axis represents average energy consumption (unit: Joules per task), and the shaded band shows ± 1 standard deviation.
The energy consumption per node ranged from 0.6 to 1.4 Joules per task-processing interval. A declining trend in energy consumption over time indicates that the FL models improved their predictions, leading to fewer failed offloading attempts and reduced retransmission overhead. The use of digital twin synchronization also reduced idle power wastage by promoting workload consolidation at underutilized nodes.
Latency comparison with baseline models
We conducted a comparative study of four architectures:
Centralized Resource Management (Baseline).
Federated Learning without DT.
FedProx (an advanced FL technique that handles heterogeneous data with proximal terms for better convergence [6].
Centralized optimization solved using the same ILP/QC-MIP solver as our DT.
Naïve FL (synchronous FedAvg without coordination).
FL + Heuristic Coordination using simple round-robin bandwidth allocation.
FL + Rule-based DT (non-forecasting DT that applies static thresholds).
FedProx and FedAvgM, representing state-of-the-art FL algorithms for heterogeneous clients.
Asynchronous FL baseline.
DL-based orchestration using a deep reinforcement learning (DQN) controller.
Federated Learning with Digital Twin (our framework).
FedProx, as an advanced FL variant, improves upon standard FL by adding a proximal term to mitigate system heterogeneity, but it lacks the predictive coordination of DT, resulting in higher latency (~ 60 ms average) compared to our FL-DT framework.
Figure 6 illustrates the average end-to-end latency per time slot for each system.
[See PDF for image]
Fig. 6
Latency comparison across methods
The x-axis represents time slots (unit: discrete time units, each approximately 10 ms), and the y-axis represents average end-to-end latency (unit: ms).
Error bars represent ± 1 standard deviation across 10 random seeds. T-tests confirm statistical significance (p < 0.05) between FL-DT and all baselines for latency reduction.
The centralized model consistently showed the highest latency (~ 100 ms), primarily due to decision delays and network overhead. FL-only systems improved latency (~ 70 ms) by allowing localized learning. However, the integration of the digital twin provided the most substantial reduction (~ 45 ms average), owing to its forecasting and global coordination capabilities. This confirms that our hybrid approach offers superior latency performance under dynamic network conditions.
Results show that FedProx and FedAvgM improved stability but did not match the predictive power of FL-DT. Asynchronous FL reduced synchronization cost but increased variance. Rule-based DT offered marginal gains, whereas heuristic coordination led to congestion under heavy loads. Our hybrid FL-DT outperformed all alternatives by at least 20% in latency and 10% in energy efficiency.
The superior performance of FL-DT stems from its hybrid approach, where DT’s predictive insights complement FL’s decentralization, enabling faster decisions and better handling of dynamic conditions compared to baselines.
Stability and convergence monitoring
Two convergence indicators were monitored:
Model convergence .
State convergence .
Both indicators exhibited declining trends across the training epochs, confirming stable adaptation of FL models and synchronized updates in digital twin state vectors. In 90% of trials, convergence thresholds were met within 10–15 time slots, indicating low synchronization lag and high system stability.
These trends indicate that the framework’s synchronization mechanisms ensure quick stabilization, making it suitable for real-time 6G applications where rapid adaptation is critical.
All results are averaged over five independent runs with different random seeds. We report mean ± standard deviation. For key comparisons, two-sample t-tests were performed, confirming significance at p < 0.05.
Resource utilization and scalability
Under simulated load bursts (5× task density spikes), the FL-DT framework maintained service quality with minimal degradation (< 15% drop in task success rate). This indicates strong scalability, reinforced by:
Redundant task reassignment mechanisms.
Predictive load balancing via DT forecasting.
Federated learning model adaptability.
Under load bursts, the DT’s forecasting and FL’s adaptability allow for seamless reassignment, maintaining high success rates by preventing overload and promoting resilience.
We further tested scalability with N = {10, 50, 100} edge nodes and M up to 500 UEs. Complexity analysis showed near-linear increase in runtime with N, while communication cost scaled sub-linearly due to selective synchronization. For 100 nodes and 500 UEs, the DT optimization required < 1.2 s per round, remaining practical for real-time operation.
Table 2. Summarizes the comparative results across all tested frameworks
Metric | Centralized | FL Only | FedProx | FL + Digital Twin |
|---|---|---|---|---|
Avg. Latency (ms) | 100.2 | 71.5 | 60.0 | 45.3 |
Task Success Rate (%) | 82.4 | 91.2 | 93.5 | 96.6 |
Avg. Energy Consumption (J) | 1.78 | 1.31 | 1.25 | 1.12 |
Convergence Epochs Required | 28 | 18 | 15 | 11 |
Table 2: Comparative Performance Summary Across Architectures.
Sensitivity and scenario analyses
To evaluate robustness, we simulated (a) light load (λ = 2 tasks/s), (b) heavy load (λ = 10 tasks/s), (c) high mobility (nodes moving at 30 m/s), and (d) heterogeneous node capacities (CPU varying from 1.0 to 3.0 GHz). Results indicate that the FL-DT maintained task success > 90% under all cases. Performance degradation was most visible under extreme mobility, where success dropped by 7%. Sensitivity to synchronization interval Tsync was also studied: reducing Tsync to 2 slots improved latency by ~ 8% but increased control overhead, while larger Tsync = 10 caused delayed adaptation.
Discussion
This section provides an in-depth interpretation of the simulation results, highlighting the significance of our proposed Federated Learning-Driven Digital Twin (FL-DT) framework in the context of 6G networks. We evaluate the trade-offs, potential deployment challenges, and technical implications of each system component—federated learning, digital twin synchronization, and their integration—while connecting the findings to broader trends in next-generation telecommunications.
Impact of federated learning on edge intelligence
The integration of federated learning (FL) as a decentralized learning strategy for edge nodes addresses several limitations inherent in centralized AI systems for networks. Unlike traditional architectures, which transmit raw data to a central cloud for training, FL keeps data local, thereby preserving user privacy and reducing network congestion.
Our simulation results (Sect. Energy consumption trends, Fig. 6) demonstrated that FL alone improves latency by approximately 29% compared to centralized models. This performance gain is largely attributed to localized decision-making and reduced communication overhead. Moreover, since each node adapts to its unique context and data distribution, the system as a whole becomes more resilient to heterogeneous environments—a hallmark of real-world 6G deployments.
However, FL introduces new technical challenges such as:
Model Drift: Variance between local and global model weights may lead to instability, particularly in mobile or highly dynamic environments.
Asynchronous Updates: In real-world systems, not all devices may respond within the same update window, necessitating robust aggregation techniques like FedAsync or FedProx.
In our design, we mitigated these risks through periodic state synchronization via the digital twin layer, ensuring that local deviations are detected and corrected before cascading into system-level errors.
Role of the digital twin in predictive optimization
The addition of a digital twin (DT) layer provides a predictive, system-wide view of network status, facilitating proactive rather than reactive decision-making. The DT aggregates local states from edge nodes (CPU load, bandwidth usage, service request rates) and applies time-series forecasting to anticipate future loads.
As shown in Figs. 2 and 4, this forecasting capability directly enhances resource allocation and energy balancing. For example, edge nodes experiencing forecasted spikes in service requests receive priority for bandwidth (Fig. 2) and CPU allocation (Fig. 4), which reduces queue lengths and prevents service degradation.
The digital twin’s ability to model global system dynamics introduces several operational benefits:
Load anticipation: Forecasted traffic spikes are met with preemptive offloading strategies.
Fault isolation: Node-specific anomalies are flagged and addressed with minimal system disruption.
Adaptability: The twin’s evolving system model ensures adaptability to emerging conditions such as congestion or device churn.
These capabilities are particularly relevant for ultra-reliable low-latency communication (URLLC) services such as autonomous vehicle coordination, real-time healthcare monitoring, and industrial automation—all of which are considered core applications in future 6G infrastructure.
Hybrid coordination: bridging local autonomy with global awareness
A major innovation in our framework is the hybridization of decentralized learning with centralized coordination through the digital twin. This duality ensures both local adaptability and global consistency—a combination rarely achieved in current network management strategies.
The offloading strategy (Fig. 3) illustrates this well. While FL models learn edge-specific offloading patterns, the digital twin coordinates task distribution to ensure balanced network load. This symbiosis ensures that decision-making is both efficient and cooperative.
Another advantage is enhanced convergence efficiency. As highlighted in Sect. Latency comparison with baseline models, the integration of FL and DT reduces convergence time by 39% compared to traditional FL-only systems. This reduction has significant implications for real-time systems that require rapid stabilization, such as drone swarm coordination or emergency response networks.
Energy efficiency and sustainability
Sustainability is a critical design constraint for 6G networks, and our simulation results reinforce the viability of the FL-DT framework in supporting energy-aware operations. As shown in Fig. 5; Table 2, the system’s energy footprint was reduced by approximately 13% compared to standalone FL systems.
The main contributors to this energy savings are:
CPU frequency throttling: Nodes dynamically downscale processing frequencies during idle or low-load periods, as shown in Fig. 4.
Efficient task scheduling: Offloading decisions minimize transmission distances and idle CPU cycles.
Avoidance of redundant processing: The DT provides global visibility, preventing duplicate model training or unnecessary data movement.
Energy-aware operation will be especially critical in IoT-heavy deployments, where thousands of battery-powered sensors or wearables must function without frequent recharging. The proposed framework offers a path toward sustainable AI at the edge.
Scalability and real-time responsiveness
Another major contribution of this work is in demonstrating scalability without performance degradation. Under synthetic burst traffic (5× task intensity), our FL-DT framework maintained a task success rate above 96%—substantially higher than centralized (82%) and FL-only (91%) architectures.
The main mechanisms enabling this scalability include:
Dynamic load redistribution: Forecasting by the DT enables early redistribution of tasks before overload occurs.
Parallel model training: FL allows nodes to train models in parallel, reducing central server dependency.
Selective synchronization: Nodes synchronize only when critical thresholds are crossed, reducing update noise and improving stability.
These properties make our framework suitable for mission-critical deployments in defense, autonomous transportation, and distributed robotics, where adaptability under stress is paramount.
Practical deployment considerations
Despite the strong results demonstrated, practical deployment of the FL-DT framework will require careful attention to several issues:
Security and privacy: While FL improves data privacy, it is not immune to model inversion or poisoning attacks. Integration with privacy-preserving techniques like differential privacy or secure multiparty computation is essential.
Hardware heterogeneity: Real-world edge nodes may have varying hardware capabilities. Our framework must support model quantization or heterogeneous training to accommodate low-power devices.
Communication reliability: FL and DT synchronization depend on stable connectivity. Incorporating redundancy mechanisms and fallback protocols will be necessary in volatile network environments.
Regulatory compliance: Particularly in critical applications like healthcare or transportation, the framework must align with legal and ethical guidelines surrounding AI transparency and accountability.
The DT optimization employed mixed-integer convex solvers with average runtime 0.35 s for N = 10, and 1.2 s for N = 100 nodes. This overhead remains below Tsync = 5, validating real-time feasibility. The complexity is O(N log N + M). Communication delays were mitigated via stale update handling and partial information acceptance; nodes not responding within 1 slot used their last known state. A budget analysis showed that forecasting, optimization, and control messaging together consumed < 8% of system resources per round.
Addressing these challenges will form the basis of future implementation and experimentation efforts.
Comparison with related approaches
Compared to baseline approaches, our FL-DT architecture represents a clear advancement in intelligent 6G network design. While existing works have explored federated learning [14, 21] or digital twins [17, 19] independently, none combine the two into a predictive, adaptive system for real-time edge resource coordination.
The comparative performance table (Table 2) substantiates this claim, with our framework outperforming others in every key metric—latency, convergence, task success, and energy.
Furthermore, our system architecture is modular and adaptable, making it suitable for diverse deployment scenarios including smart cities, remote healthcare platforms, and mobile vehicular networks. This flexibility enhances its value as a general-purpose framework for intelligent telecommunications.
Conclusion
The emergence of sixth-generation (6G) networks marks a pivotal transformation in the telecommunications domain, driven by growing demands for ultra-low latency, massive device connectivity, and intelligent system behavior. As communication infrastructures evolve toward greater complexity and autonomy, traditional centralized resource management approaches are proving inadequate. This paper presented a novel solution to these challenges by introducing a Federated Learning-Driven Digital Twin (FL-DT) Framework, which unifies decentralized intelligence and system-wide predictive modeling for adaptive resource management in 6G edge networks.
The proposed architecture combines the decentralized, privacy-preserving strengths of federated learning (FL) with the real-time forecasting and global orchestration capabilities of a digital twin (DT). Federated learning enables edge devices to train models locally using their own data, significantly reducing latency and protecting user privacy. Simultaneously, the digital twin aggregates dynamic system states, forecasts future loads, and issues control signals to coordinate spectrum allocation, computation offloading, and energy optimization across the network.
Through rigorous mathematical formulation and comprehensive simulations, we demonstrated the performance benefits of this integrated system. Our results show that the FL-DT framework:
Reduces average system latency by over 50% compared to centralized models and by approximately 35% compared to FL-only models.
Improves task offloading success rates to over 96%, even during bursty traffic conditions.
Lowers average energy consumption through predictive load distribution and intelligent CPU frequency scaling.
Accelerates convergence of learning models and system stabilization, enabling rapid adaptation to real-time changes.
Importantly, the framework also exhibited strong scalability and robustness, key requirements for the decentralized, heterogeneous, and high-mobility environments envisioned for 6G networks. These qualities make it well-suited for use cases such as autonomous transportation systems, industrial automation, and large-scale IoT deployments.
While the results are promising, several challenges remain for real-world deployment. These include securing federated learning models against adversarial attacks, ensuring reliable DT synchronization in fluctuating wireless environments, and addressing regulatory and hardware heterogeneity constraints.
Future work will focus on implementing this architecture in physical testbeds, exploring hybrid learning techniques, and integrating emerging technologies such as blockchain and post-quantum cryptography to further enhance system trustworthiness and resilience.
In conclusion, the integration of federated learning and digital twins presents a powerful paradigm for intelligent, self-optimizing 6G networks. By enabling distributed learning with centralized coordination, the FL-DT framework contributes a scalable, adaptive, and secure approach to managing the increasingly complex demands of next-generation telecommunications systems.
Acknowledgements
Not applicable.
Author contributions
Milad Rahmati and Nima Rahmati contributed equally to this manuscript. They jointly conceptualized the research idea, developed the theoretical framework, designed and implemented the methodology, conducted simulations and analysis, interpreted the results, and co-wrote and revised the full manuscript.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
Data availability
The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Competing interests
The authors declare that there are no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
1. Giordani M, Polese M, Mezzavilla M, Rangan S, Zorzi M (2020) Toward 6G networks: Use cases and technologies, IEEE Commun Mag. 58(3):55–61. https://doi.org/10.1109/MCOM.001.1900411
2. Saad, W; Bennis, M; Chen, M. A vision of 6G wireless systems: applications, trends, technologies, and open research problems. IEEE Netw; 2020; 34,
3. Al-Fuqaha A, Guizani M, Mohammadi M, Aledhari M, Ayyash M (2015) Internet of Things: A survey on enabling technologies, protocols, and applications, IEEE Commun Surveys Tuts. 17(4):2347–2376. https://doi.org/10.1109/COMST.2015.2444095
4. Zhang J, Tao X, Xu X, Wu Q (2020) 6G wireless networks: Vision, requirements, architecture, and key technologies. IEEE Veh Technol Mag. 15(3):28–41. https://doi.org/10.1109/MVT.2020.2996642
5. Yang K, Jiang T, Shi Y, Ding Z (2020) Federated learning via over-the-air computation, IEEE Trans Wireless Commun. 19(3):2022–2035. https://doi.org/10.1109/TWC.2019.2953725
6. Kairouz, P et al. Advances and open problems in federated learning. Found Trends Mach Learn; 2021; 14, pp. 1-2. [DOI: https://dx.doi.org/10.1561/2200000083]
7. Boschert, S; Rosen, R. Hehenberger, P; Bradley, D. Digital twin—the simulation aspect. Mechatronic futures; 2016; Cham, Springer: pp. 59-74. [DOI: https://dx.doi.org/10.1007/978-3-319-32156-1_5]
8. Zeng Y, Wu Q, Zhang R (2019) Accessing from the sky: A tutorial on UAV communications for 5G and beyond, Proc IEEE. 107(12):2327–2375. https://doi.org/10.1109/JPROC.2019.2936868
9. Xiao Z, Xia P, Xia X-G (May 2016) Enabling UAV cellular with millimeter-wave communication: potentials and approaches. IEEE Commun Mag. 54(5):66–73. https://doi.org/10.1109/MCOM.2016.7470940
10. NSF (2020) Future wireless: 2020 and beyond, National Science Foundation, https://www.nsf.gov
11. Commission E (2021) 6G Flagship Project, 6G Research Roadmap, https://www.6gflagship.com
12. Konečný J, McMahan HB, Yu FX, Richtárik P, Suresh AT, Bacon D (2016) Federated learning: Strategies for improving communication efficiency, Proc NIPS Workshop Private Multi-Party Mach Learn, Long Beach, CA, USA
13. Bonawitz K et al (2017) Practical secure aggregation for privacy-preserving machine learning, In Proceedings of the ACM CCS, pp. 1175–1191. https://doi.org/10.1145/3133956.3133982
14. Niknam S, Dhillon HS, Reed JH (2020) Federated learning for wireless communications: Motivation, opportunities, and challenges, IEEE Commun. Mag., vol. 58, no. 6, pp. 46–51, Jun. https://doi.org/10.1109/MCOM.001.1900463
15. Wang S, Tuor T, Salonidis T, Leung KK, Makaya C, He T, Chan K (2019) Adaptive federated learning in resource constrained edge computing systems, IEEE J Sel Areas Commun. 37(6):1205–1221. https://doi.org/10.1109/JSAC.2019.2904348
16. Fuller, A; Fan, Z; Day, C; Barlow, C. Digital twin: enabling technologies, challenges and open research. IEEE Access; 2020; 8, pp. 108952-108971. [DOI: https://dx.doi.org/10.1109/ACCESS.2020.2998358]
17. Li Y et al (Apr. 2021) Digital twin in smart manufacturing. J Manuf Syst 58:115–128. https://doi.org/10.1016/j.jmsy.2020.06.017
18. Zhang, Y; Huang, K; Liu, X. Digital twin-based network resource orchestration in cloud-native Telecom networks. IEEE Netw; 2021; 35,
19. Xu Z, Liu Y, Zhang J (2021) Digital twin in telecommunications: Models, challenges, and future directions, IEEE Commun Mag, 59(6):18–24. https://doi.org/10.1109/MCOM.001.2000337
20. Ordonez-Lucena J et al (2017) Network slicing for 5G with SDN/NFV: Concepts, architectures, and challenges. IEEE Commun Mag. 55(5):80–87. https://doi.org/10.1109/MCOM.2017.1600935
21. Chen M, Saad W, Yin C, Debbah M (2017) Echo state networks for proactive caching in cloud-based radio access networks with mobile users, IEEE Trans. Wireless Commun. 16(6):3520–3535. https://doi.org/10.1109/TWC.2017.2688320
22. Chen X, Jiao L, Li W, Fu X (2016) Efficient multi-user computation offloading for mobile-edge cloud computing. IEEE/ACM Trans Netw 24(5):2795–2808. https://doi.org/10.1109/TNET.2015.2487344
23. Challita A, Dong L, Saad W (2018) Proactive resource management for LTE in unlicensed spectrum: A deep learning perspective, IEEE Trans Wireless Commun, 17(7):4674–4689. https://doi.org/10.1109/TWC.2018.2827055
24. Samarakoon S, Bennis M, Saad W, Debbah M (2020) Distributed federated learning for ultra-reliable low-latency vehicular communications, IEEE Trans Commun, 68(2):1146–1159. https://doi.org/10.1109/TCOMM.2019.2948477
25. Park J, Samarakoon S, Bennis M, Debbah M (2019) Wireless network intelligence at the edge, Proc. IEEE, vol. 107, no. 11, pp. 2204–2239, Nov. https://doi.org/10.1109/JPROC.2019.2933795
26. Hassan M, Shihada B, Alouini M (2021) Edge computing with SDN and network slicing for 5G: A survey, IEEE Internet Things J. 8(6):4355–4373. https://doi.org/10.1109/JIOT.2020.3022481
© The Author(s) 2026. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.