Introduction

Communication networks have evolved significantly over the past decades, progressing from infrastructures focused solely on connectivity to intelligent systems that tightly integrate communication, sensing, and computation1,2,3. In current 6G visions, this integration is expected to deepen further, supported by enabling technologies such as space-air-ground architectures, intelligent reconfigurable surfaces, and large-scale edge-cloud collaboration4,5,6. These developments point toward a future where networks are not only connected and aware, but also adaptive, perceptive, and context-sensitive7,8.

However, this functional convergence also introduces fundamental structural challenges9,10,11,12. The integrated communication, sensing, and computation architecture can be recognized to follow a pyramid-shaped hierarchy, in which a vast number of heterogeneous devices with embedded sensing, communication, and computation capabilities populate the lower layers, while computing resources and coordination functions become increasingly concentrated in the higher layers13,14. Although such concentration is efficient for centralized management, it also creates structural bottlenecks because overload or disruption in the higher layers can propagate downward, increase decision latency, and restrict devices’ ability to adapt to local changes. As a result, conventional top-down control and static hierarchical management approaches struggle to provide the scalability and responsiveness required in large-scale, real-time environments. Furthermore, heavy reliance on human intervention further diminishes the overall autonomy of the system.

To overcome these limitations, we propose a conceptual shift: to model future communication systems through the lens of natural ecosystems within the digital domain15,16,17. Inspired by real-world ecological systems, we introduce the concept of digital populations at the edge infrastructure, where multi-functional terminal agents are managed within the edge’s computing capabilities and communication range. At the cloud level, multiple digital edge populations interact through distributed coordination, adaptive behavior, and feedback-driven evolution, collectively fostering the development of the whole digital ecosystem. Under this perspective, the network is no longer viewed as a mere collection of devices, but as an evolving society of intelligent agents governed through distributed mechanisms.

A key driver of this transformation is our introduction of a layered digital twin (DT) architecture that enables bidirectional interaction between the physical and digital domains. This architecture establishes the foundational abstraction of physical communication devices and environments into digital ecological replicas, encompassing local DTs that represent individual terminals, intermediate edge-level DTs that coordinate digital populations, and global DTs that manage strategy functions at the ecosystem scale.

Within the edge-level digital populations, communication, sensing, and computation are decoupled from the rigid boundaries of physical devices and reallocated as population-level functions. Local digital agents are dynamically assigned roles based on environmental conditions, device capabilities, and observed demands, with some agents specializing in data acquisition and upward transmission, and others focusing on local processing or coordination. Beyond the level of individual populations, we also propose an evolutionary approach at the cloud ecosystem scale. In this model, the behavioral rules and control structures of multi-edge populations are continuously reorganized and reconfigured in response to feedback signals, mobility dynamics, and environmental variability, while maintaining consistency under human-intervention functional goals.

In summary, this paper presents a layered DT framework that capitalizes on the inherent capabilities of communication, sensing, and computation to establish a self-organizing digital ecosystem. It describes how edge-level digital populations can be constructed on top of existing infrastructures, dynamically adapt to mobility and functional heterogeneity, and operate under the guidance of dual flows of information and authority to manage local environments autonomously, while remaining integrated with strategy, multi-population governance from the cloud. Viewed holistically, the network can thus be regarded as a living digital ecosystem, co-evolving with its environment through feedback, regulation, and succession.

To present this perspective in a coherent manner, the paper is structured to follow the progressive construction of a digital ecosystem, as illustrated in Fig. 1. The system evolves from physical communication infrastructure to a layered DT system, and ultimately into an autonomous and adaptive digital ecosystem. Table 1 summarizes the ecological analogies used throughout the paper to provide a unified narrative.

Fig. 1
Fig. 1
Full size image

Progressive construction of a digital ecosystem: from physical networks to a three-layer DT system that mirrors device states, fuses edge-level information, and performs cloud-level optimization42,43, forming a near real-time and evolvable digital ecosystem.

Table 1 Unified Ecological Analogies for the Multi-Layer DT Ecosystem
Full size table

Results

Step 1: constructing an interactive replica of the physical environment—a multi-layer DT framework

The first step in abstracting a digital ecosystem from physical device systems is to construct an efficient and interactive digital modeling mechanism that enables real-time mirroring and intelligent management of the physical communication environment.

Digital twins as intelligent agents

To realize this, DT technology is introduced as a core enabler. DTs maintain consistent and timely synchronization with their physical counterparts through continuous bidirectional data exchange18,19,20. While DTs have already been widely applied in industrial and engineering domains for equipment monitoring, behavioral simulation, and decision support, their traditional use as static, human-facing replicas does not meet the demands of future communication systems.

Targeted communication systems are characterized by large scale, mobility, and functional integration of communication, sensing, and computation. As such, DTs should support these features systematically. Furthermore, DTs should not remain passive reflections. Instead, each DT must act as an autonomous digital agent that can interact persistently with its physical entity, perceive status and environmental changes, report operational feedback, and respond to high-level digital strategies.

In this context, DT modeling becomes more than just a mechanism for mapping physical states. It serves as the starting point for constructing autonomous and cooperative structures. Through this foundation, previously isolated physical nodes are integrated into a coordinated system that supports the formation of edge-level digital populations and enables long-term governance and evolution across the entire digital ecosystem.

Architecting the twinning pyramid: local, edge, and cloud

Modern communication networks are generally organized following a pyramid-shaped structure. At the base, a large number of resource-constrained devices carry out functions related to communication, sensing, and computation. These devices are diverse in distribution, often mobile, and primarily rely on wireless connectivity. As a result, the local layer exhibits the highest management complexity because it contains the largest number of heterogeneous and mobile nodes with distinct resource constraints and frequent state transitions. This diversity leads to combinatorial growth in coordination overhead, making orchestration intrinsically more challenging than at the edge or cloud layers21,22. Moving upward, the edge layer provides more concentrated functionality, utilizing moderate computational resources to analyze and control local terminals within its communication domain. At the top, the cloud layer manages global services and policies, supported by powerful computing data centers and dedicated wired links.

To effectively map and utilize this hierarchical physical infrastructure, we introduce a multi-layer DT architecture composed of three closely connected levels: local, edge, and cloud. Each level maintains its own form of organizing DT representation, tailored to the functional and management needs in that layer. From an ecological perspective, this hierarchy resembles a food-web structure: local terminals act as primary producers, edge infrastructures function as intermediate niches that channel and redistribute energy, and the cloud serves as the apex regulator responsible for maintaining balance and preventing systemic collapse23,24.

Local DT: At the lowest level of the DT system, local DTs are embedded within various terminal devices such as sensors, smartphones, vehicles, and robots. Relying on each device’s limited computing and sensing capabilities, local DTs collect real-time information including internal states such as CPU usage and power consumption, external environmental conditions such as temperature, location, and channel quality, as well as user behavior data such as application usage patterns and mobility trajectories.

In our design, local DTs are implemented as lightweight digital modules. They primarily execute pre-deployed models and control policies provided by higher layers, performing minimal preprocessing and fast inference to deliver immediate responses. Considering the limited computation and energy resources of terminal devices, the frequency of sensing and reporting is adaptively adjusted according to context dynamics, so that responsiveness is preserved without producing redundant transmissions. Beyond inference, local DTs also support small-scale on-device learning to fine-tune parameters to local contexts such as mobility or energy dynamics. These incremental updates are aggregated at the edge infrastructure, where federated learning, knowledge distillation, or transfer learning constitute the main training process for the refined model. The cloud does not perform heavy training but provides long-term policy guidance and global consistency, enabling local DTs to remain lightweight and autonomous functions while continuously improving through this closed-loop process.

Much like biological individuals in natural ecosystems, each local DT operates autonomously within its environment, perceiving changes, reacting to local stimuli, and interacting with nearby entities. This autonomy enables local DTs to perform judgment and response directly at the terminal level, which helps reduce latency and alleviates the burden on upper layers. They also manage the upward transmission of processed data, task requests, and runtime status to edge DTs, while receiving policy feedback from higher layers. To enhance robustness in wireless environments, local DTs support horizontal communication with neighboring devices, facilitating basic collaboration and information sharing.

Local DTs represent the most numerous, widely distributed, and frequently updated components of the digital ecosystem. They carry out micro-level responsibilities such as environmental perception, data preprocessing, and immediate response. Acting as both the sensory interface to the physical world and the initial node for upward information flow, each local DT lays the foundation for a modular, agent-based digital ecosystem.

Edge DT: At the middle layer of the digital ecosystem, the edge DT is deployed at edge infrastructures such as base stations and roadside units (RSUs). Its core task is to construct corresponding digital agents based on the data uploaded by local DTs within its coverage area. These agents are considered as digital individuals, collectively forming a regional digital population. Serving as the organizational and control center of this population, edge DT undertakes responsibilities such as state monitoring, resource scheduling, and strategy execution.

The primary characteristic of the edge DT is its ability for regional coordination. While local DTs focus on individual sensing and reaction, the edge DT abstracts and manages group-level behaviors. It integrates and allocates functions across multiple terminals within its domain, allowing the edge DT to balance the need for responsiveness with the complexity of coordination.

Another key feature is its structural flexibility and mobility awareness. The edge DT maintains a dynamic population structure where digital individuals can enter or exit the environment based on mobility patterns, resource availability, or communication conditions. Through collaborative perception, distributed decision-making, and internal information exchange, the edge DT governs its local DT members.

In short, the edge DT functions as a mid-layer control entity within the digital ecosystem. It transforms distributed local DTs into a coordinated population, enabling efficient regional services and preparing the foundation for higher-level orchestration at the cloud layer.

Cloud DT: At the top layer, the cloud DT governs multi-population dynamics and acts as the global coordinator for long-term optimization, much like the invisible hand that guides balance in natural ecosystems. In contrast to the local and edge DTs that handle immediate sensing and regional coordination, the cloud DT functions as the strategy brain of the entire system. It continuously receives feedback from distributed edge DTs and integrates information across diverse geographic and functional domains.

The cloud DT enables the digital ecosystem to evolve over time. Through the analysis of historical records, behavioral trends, and inter-population interactions, it preserves an evolutionary memory that supports system-wide adaptation. This includes updating global models, refining coordination policies, and launching large-scale optimization strategies that reflect the long-term needs of the network.

From the ecosystem perspective, each edge DT is abstracted as a regional species embedded in a broader digital ecology. These populations may vary in density, mobility, service demands, and resource capabilities. The cloud DT is responsible for managing their coexistence and co-evolution, ensuring responsiveness, stability, and fair development across the system.

In summary, the cloud DT aggregates and coordinates multiple edge populations into a unified view and applies global policy adaptation and optimization mechanisms that enable the ecosystem to evolve over time25. By maintaining strategy foresight, adapting governance logic, and enabling scalable coordination, it constitutes the highest layer of intelligence within the multi-layer DT architecture.

As illustrated in Fig. 2, the multi-layer DT system will evolve from physical communication infrastructures into a digital ecosystem composed of digital individuals, edge populations, and the cloud ecosystem. Each layer fulfills distinct yet collaborative roles: local DTs in user terminals enable real-time responsiveness, edge infrastructures coordinate distributed intelligence across mobile populations, and the cloud performs long-term modeling and ecosystem-wide evolution. Together, they form a closed ecological loop where information flows upward as nutrients and authority flows downward as regulatory signals, sustaining balance and adaptability across the entire system.

Fig. 2
Fig. 2
Full size image

Digital ecosystem derived from the multi-layer DT system, illustrating the hierarchical coordination among local individuals, edge populations, and the cloud ecosystem.

Step 2: forming mobile-adaptive digital populations at the edge DT layer—information ascent and strategy descent

Based on the “local to edge to cloud” structure of modern networks, the first step proposed a layered DT system architecture that lays the foundation for developing a digital ecosystem. In this framework, each device is treated as an individual entity equipped with sensing, communication, or computation capabilities. By binding with its DT, it becomes a digital unit that can be perceived, managed, and responded to within the edge layer.

Building on this foundation, the next step is to construct a dynamically evolving digital population at the edge layer. This population must adapt to continuous changes in the state and location of underlying devices, while also supporting the upward aggregation of local information and the downward dissemination of higher-level strategies. Functionally, this structure resembles what ecologists describe as intermediate ecological niches, which are positions within an ecosystem that link basal producers and top consumers, enabling energy flow and stabilizing population dynamics26. In our context, the edge layer plays this intermediary role: it mediates interactions among individual devices, buffers fluctuations, and sustains system-level adaptability in response to dynamic environments.

Edge-level digital population organization

The primary task of the edge DT is to organize the originally scattered local DTs into a logically unified and dynamically adaptive digital population. Each local DT represents a local terminal and continuously reports its status and environmental data to the edge DT. Under the coordination of the edge DT, these local terminals and their digital counterparts collectively form an expandable and contractible population managed by the edge infrastructure.

Within this structure, the edge DT does not impose rigid centralized control over the digital individuals. Instead, it introduces a functional differentiation mechanism inspired by natural populations. Sensing, communication, and computation capabilities are decoupled from the fixed boundaries of physical devices and reallocated as population-level functions. The edge DT observes the state of local terminals and edge resources, including their compute capacity, communication quality, energy status, and task load. Based on these observations, it flexibly assigns roles such as data collection, local processing, and coordination to appropriate participants. These roles are not static: when task demand, mobility, or network conditions shift, the edge DT re-evaluates the allocation and reshapes the division of functions, enabling the population to adapt to the current environment and escalate tasks to neighboring edge infrastructures or the cloud when local capacity is saturated.

Then, the dynamic organization of the population is reflected in two key aspects. First, structural flexibility: the edge DT adjusts its management boundaries in real time according to terminal mobility, link quality, or resource availability, enabling smooth entry and exit of local terminals. Second, functional diversity: decoupled functions allow local terminals to form a distributed division of labor across sensing, communication, and computation, improving adaptability to complex environments and supporting cooperative behaviors beyond the constraints of individual hardware capabilities.

Bottom-up information ascent

In the digital ecosystem, bottom-up information ascent is a core mechanism to support regional intelligence, vertical coordination and adaptation. Ecologically, this process is analogous to nutrient and energy flow in trophic networks, where primary producers collectively channel energy upward to sustain higher-level organisms23,27. The edge DT, functioning as the population control center, continuously receives heterogeneous data streams from its associated local individuals and builds a cohesive understanding of the regional environment.

This process begins with the on-device preprocessing capabilities of local individuals assigned sensing and computation roles. These agents perform operations such as downsampling, denoising, and feature extraction. The preprocessed information is labeled with device identifiers and contextual metadata, then transmitted by communication-role agents to the edge DT either periodically or in response to specific triggers.

At the edge DT, incoming data streams are aggregated and modeled to construct a structured view of regional states. This includes the fusion of sensing data such as environmental conditions and mobility, communication metrics such as link quality and congestion, and computational information such as task load and resource availability. Through spatial and temporal analysis, the edge DT identifies patterns, correlations, and latent risks across the digital individuals. For example, in urban scenarios, traffic flow maps can be generated in real time by aggregating speed and location data from multiple vehicle DTs, allowing the system to detect congestion or anomalies.

Processed insights are encapsulated into structured uplink packages and transmitted to the cloud ecosystem. Each package contains event labels, structured descriptions of current states, and trend indicators that support higher-layer strategy modeling, anomaly prediction, and policy refinement.

Top-down strategy descent

In the digital ecosystem, the edge DT not only supports bottom-up information ascent but also plays a central role in top-down strategy descent. This process resembles top-down control in ecology, where apex predators, climatic drivers, or seasonal cues regulate population behaviors to maintain ecosystem stability28,29. The cloud ecosystem, based on cross-regional historical data and system-wide optimization goals, formulates strategies such as resource allocation policies, model update plans, or service-level adjustments. However, these strategies often require contextual refinement before they can be applied effectively to heterogeneous and dynamically changing edge populations. The edge DT acts as both a translator and scheduler of these strategies.

Specially, the edge DT interprets the strategy intent behind cloud-issued policies and adapts them to its current environmental context. Based on the available resources, it transforms high-level commands into executable actions suitable for each local individual. This includes dynamically reassigning sensing, communication, and computation roles to match the current operational priorities. For example, in scenarios with communication congestion, certain agents may be designated as high-bandwidth relays. In areas with degraded situational awareness, high-precision sensing nodes may be activated to improve environmental awareness.

As the edge DT advances in its ability to interpret and enact cloud strategies, it progressively attains higher levels of edge autonomy, thereby supporting decentralized coordination. It also forms the operational backbone for top-down guidance within the digital ecosystem, ensuring that the intentions of the cloud ecosystem are effectively realized across diverse and mobile edge populations.

Step 3: cloud-integrated multi-population ecosystem—evolutionary population governance

Building upon the local digital replicas and the edge-level digital populations, the third step introduces a long-term governance mechanism at the cloud layer, aiming to establish an evolutionary multi-population ecosystem. Rather than relying on centralized control, this ecosystem leverages the cloud DT’s capabilities in historical accumulation, global awareness, and strategy planning to coordinate the co-evolution of multiple edge populations.

Serving as the highest-level intelligent core, the cloud assumes three essential roles. First, it acts as an integrator of knowledge by continuously receiving summarized information from edge populations and constructing a global understanding of the system state. Second, it functions as a strategy generator, creating actionable policies for lower layers based on historical behavior trajectories, trend predictions, and optimization objectives. Third, it serves as the system’s evolutionary memory, transforming long-term operational experiences, adaptation feedback, and dynamic resource distributions into experiential rules.

Knowledge accumulation

The cloud continuously aggregates uplink data packets from multiple edge regions to construct a cross-regional and cross-modal state graph. This graph not only includes raw indicators such as environmental measurements, resource metrics, and task distributions, but also integrates higher-level abstractions like behavioral patterns, service request trends, and the evolution paths of abnormal events.

This knowledge accumulation process parallels the concept of cultural memory or genetic inheritance in natural ecosystems. In biological populations, complex behaviors such as migratory routes or cooperative hunting are rarely the product of isolated individuals. Instead, they emerge through collective experiences, reinforced across generations. Likewise, the cloud layer of a digital ecosystem structures and models accumulated operational memory, refining its understanding of system dynamics and encoding this learning into strategy templates and predictive models.

Strategy generation

With its global view and historical memory, the cloud DT formulates a multi-population and cross-regional strategy framework. These generated strategies span across domains such as resource reallocation, task offloading, energy management, security enforcement, and model evolution.

Crucially, the cloud does not directly command execution at lower layers. Instead, it generates abstract policies tailored to specific regions. For example, based on predicted communication bottlenecks, the cloud might derive adaptive threshold rules and send them to edge DTs as parameterized policy blueprints. Each edge DT further localizes these policies, translating them into scheduling decisions, device activation plans, or content caching instructions.

This process is analogous to seasonal migration, which in ecology refers to the periodic, large-scale movement of animal populations triggered by environmental cues such as temperature, photoperiod, or resource availability29. Similar to how migration signals provide a shared directional trend while allowing each flock or herd to adjust timing and route to local conditions, the cloud-level policies set the global direction, whereas each edge infrastructure adapts them to its own network state and device context.

Multi-population coordination

Edge DTs independently manage their digital populations within bounded geographic or functional domains. However, for a digital ecosystem to maintain long-term adaptability and resilience, cross-population coordination is essential. Ecologically, this process resembles networked interactions among distinct species populations, where mutualism, competition, and facilitation collectively regulate community structure and resource distribution. The cloud DT achieves this by establishing mapping relationships among populations, identifying complementary pairs such as overloaded versus underutilized zones, and enabling resource balancing across the network.

For instance, when a specific region experiences computation overload, the cloud may coordinate low-load edge DTs from neighboring areas to contribute processing support.

Evolutionary governance

The cloud DT’s primary goal is to enable a self-improving and continuously evolving twin ecosystem. Rather than relying on fixed rules, it follows a complete feedback cycle that starts with the collection of real-time operational data from lower layers, including task completion ratio, average response latency, resource utilization, and edge autonomy levels. These metrics are then evaluated to identify performance bottlenecks or imbalances. Updated strategies are generated through model retraining or parameter adjustment and are first deployed in a limited subset of the network for validation. During this pilot deployment, system behavior is continuously monitored, and only strategies that deliver verified improvements and stability are gradually rolled out to the entire network.

Through this iterative process, the cloud DT evolves from a passive controller to an active meta-coordinator. It maintains institutional memory, steers the long-term trajectory of the digital ecosystem, and aligns population-level behavior with system-wide objectives, all while preserving local autonomy and short-term responsiveness. This closed-loop governance ensures that the digital ecosystem remains stable, scalable, and adaptive in the face of changing network conditions.

Step 4: a healthy digital society—the synergistic coupling of social and natural sciences

In the first three steps, we constructed a multi-layer DT system from an engineering perspective, tailored to future network architectures. This system integrates real-time sensing at the local layer, dynamic organization of edge populations, and long-term governance at the cloud. Together, these elements form a self-adaptive and self-evolving digital ecosystem, as illustrated in Fig. 3. Throughout the design process, we drew from natural systems, introducing mechanisms such as population coordination, functional differentiation, information propagation, and feedback control to support dynamic stability similar to ecological systems.

Fig. 3
Fig. 3
Full size image

Digital ecosystem architecture composed of local individuals, edge populations, and the cloud ecosystem, reflecting roles in real-time interaction, distributed coordination, and long-term governance.

As the system moves into long-term operation, its internal behavior will begin to exhibit social patterns beyond traditional engineering control. For example, cooperation and role division among DTs within edge populations resemble organizational structures seen in social groups. Strategy delivery, resource scheduling, and task competition also show characteristics of incomplete-information games. At this stage, theories from governance studies, behavioral science, and trust modeling can offer tools to improve coordination, resilience, and fairness.

To sustain healthy evolution, we integrate further detailed examples for embedding social science principles into DT system operations. Reputation systems are deployed at the edge, where each DT maintains a reputation score updated according to its contribution history and compliance with cooperation rules; tasks from high-reputation DTs receive higher scheduling priority30. Cooperation incentives are implemented by translating collaborative behavior into token-like credits redeemable for preferred access to shared resources, encouraging collective action in multi-domain environments31. Institutional trust enforcement is supported by rule-based access control and transparent decision policies at the cloud. These mechanisms can be implemented as lightweight protocol extensions and northbound APIs, making them deployable without excessive overhead.

Moreover, the multi-layer DT environment itself provides a safe digital sandbox (e.g., cloud DT) where these mechanisms can be validated at scale. Large-scale simulations measure fairness, cooperation rates, and stability under different governance policies, enabling gradual tuning before live deployment32. The most effective configurations are then applied to the operational network in a phased manner, reducing risk and enhancing trust. In this way, our approach grows from reality, is refined through digital experimentation, and then returns to reality to guide network protocols and resource policies, completing a virtuous cycle for a healthy digital society.

Discussion

Building on the preceding steps and results, this section discusses the broader implementation pathways and comparative perspectives of the proposed digital ecosystem. The aim is to interpret how the observed system behaviors translate into practical deployment strategies, highlight its correspondence with existing decentralized paradigms and network standards, and examine its scalability and adaptability across heterogeneous environments.

Recent research has explored decentralized and distributed paradigms to mitigate the limitations of centralized networks. Representative efforts include Web 3.0 as a decentralized alternative to Web 2.0, peer-to-peer overlay frameworks such as libp2p that enable dynamic topology adaptation, and Decentralized Autonomous Organizations (DAOs) for distributed governance33. In parallel, federated learning and edge intelligence have been proposed for Industrial Internet of Things (IIoT) scenarios to reduce latency and preserve privacy34,35. These approaches, while valuable, remain largely confined to the application or protocol layer and present three key limitations: they are loosely integrated with the underlying communication infrastructure, provide only partial support for joint communication-sensing-computation coordination, and seldom enable dynamic role reassignment across layers. Our ecological multi-layer DT framework embeds autonomy directly into the network fabric. Local DTs act as autonomous producers sensing micro-environments, edge DTs form intermediate ecological niches that coordinate populations, and the cloud DT serves as a global regulator performing long-term optimization. This layered ecological perspective supports mobility, heterogeneity, and continuous adaptation, extending existing decentralized paradigms into a self-evolving communication-sensing-computation ecosystem.

For further practical deployment, the three-layer DT architecture can be mapped to current network standards. Local DTs correspond to user equipment (UE) and Internet of Things (IoT) terminal devices under the Third Generation Partnership Project (3GPP) specifications, edge DTs align with Multi-access Edge Computing (MEC) servers and next-generation NodeBs (gNBs) or RSUs defined by ETSI MEC, and cloud DTs are instantiated in Open Cloud (O-Cloud) and 5G Core (5GC) infrastructures36,37. Information ascent, such as state synchronization and sensing updates, is realized via Service-Based Interfaces (SBIs), N2/N3/N4 signaling paths, and MEC northbound APIs, while strategy descent, including policy enforcement and scheduling, leverages the Network Data Analytics Function (NWDAF), Policy Control Function (PCF), and Application Function (AF) over the N5 interface38. This mapping shows that the framework can interoperate with Wi-Fi/IEEE 802.11 backhaul and Open Radio Access Network (O-RAN) disaggregated architectures, while remaining forward-compatible with AI-native 6G networks39.

Scalability and adaptability are also essential for next-generation networks operating across heterogeneous and dynamic environments. Edge DTs can federate across multiple access points or RSUs to form regional populations for low-latency coordination40,41, and inter-MEC orchestration with network slicing enables cooperation across multiple domains. The framework is topology-agnostic, allowing deployment on cellular, Wi-Fi, or decentralized infrastructures without modifying its core design. Peer-to-peer overlays such as libp2p extend the edge-cloud plane laterally rather than introducing new vertical layers, enabling direct state exchange and decentralized strategy sharing to improve resilience and reduce dependence on centralized cloud control33. When trust guarantees are needed, blockchain and distributed ledgers can provide transparent and tamper-resistant state management39. Together, these mechanisms allow the framework to scale from local populations to large, multi-domain ecosystems while maintaining responsiveness and robustness35,38.

Methods

Multi-layer ecosystem showcase for smart city traffic

To validate the proposed digital ecosystem, we simulate a compact smart city scenario with six roadside units (RSUs) deployed at major intersections, each managing about 200 vehicles. Every vehicle maintains a lightweight local DT that periodically reports location, velocity, energy, and navigation intent. RSUs aggregate these updates to form edge-level digital populations, while the cloud layer fuses summaries for global optimization and disseminates top-down policies.

Compared scheme

We evaluate three representative architectures for systematic comparison:

  • Traditional Network: Centralized cloud processing with backhaul offloading and unified queueing, representing a baseline centralized system.

  • Non-Adaptive Edge: Tasks are pinned to their originating RSU without dynamic load balancing or adaptive evolving, representing a static edge deployment.

  • Layer Ecosystem (Proposed): Multi-layer coordination with adaptive edge populations. Each RSU receives the evolved polices from cloud, dynamically adjusting thresholds for execution and migration, while cooperating with neighboring RSUs to progressively enhance autonomy and coordination.

Evaluation metrics and methodology

To ensure clarity and reproducibility, we explicitly define the measured metrics:

  1. 1.

    95th-Percentile Latency: Measured from task generation at the vehicle to completion at the assigned layer (local, edge, or cloud). We use the 95th percentile across evaluation windows to capture responsiveness under both typical and tail conditions.

  2. 2.

    Edge Autonomy and Coordination Ratios: Autonomy is defined as the fraction of tasks completed within the RSU without cloud involvement. Coordination is defined as the fraction of tasks jointly executed across neighboring RSUs. Both are computed per evaluation window to reflect temporal evolution.

  3. 3.

    Relative Cost Profile: We normalize costs against the traditional baseline and decompose them into four categories: (i) deployment (infrastructure rollout), (ii) edge computation (CPU cycles consumed), (iii) signaling overhead (control-plane exchanges for migration and coordination), and (iv) terminal energy (per-vehicle communication cost).

Analysis

Figure 4 presents the comparative evaluation across responsiveness, autonomy, and cost dimensions.

Fig. 4: The comparative performance of the three architectures is illustrated in this figure across latency, autonomy–coordination evolution, and cost.
Fig. 4: The comparative performance of the three architectures is illustrated in this figure across latency, autonomy–coordination evolution, and cost.
Full size image

In a), the end-to-end latency distributions under moderate (blue) and heavy (orange) load conditions are shown. In b), the temporal evolution of autonomy and coordination is presented, where dotted blue and dashed green lines represent the layer-ecosystem architecture, dotted grey lines represent the non-adaptive edge architecture, and the black line denotes the traditional network architecture. In c), the normalized cost components, including deployment, edge computation, signaling overhead, and terminal energy, are compared using blue (traditional network), cyan (non-adaptive edge), and red (layer-ecosystem) bars.

(a) Responsiveness. The proposed Layer Ecosystem lowers 95th-percentile latency by about 40–50% compared to centralized baseline and 15–25% compared to static edge. It also narrows the interquartile range, indicating reduced variance. While the centralized system appears “stable but slow” under moderate load, its variance grows sharply under heavy load. By contrast, the layered scheme sustains both low latency and temporal stability.

(b) Autonomy and Coordination. The Layer Ecosystem exhibits a monotonic rise in both autonomy and coordination ratios, reflecting the evolutionary growth of edge populations reinforced by long-term strategy optimization from the cloud. In contrast, the centralized scheme remains flat at near-zero autonomy, while the static edge solution stays at low but fixed levels.

(c) Trade-offs and Costs. The layered system introduces measurable overheads: signaling increases by about 8% and edge CPU utilization rises by roughly 15% compared to static edge. However, these are offset by a reduction of more than 40% in backhaul traffic, significantly alleviating core congestion and enhancing scalability. Terminal-side energy consumption remains comparable or slightly lower, thanks to fewer retransmissions and faster completions. Deployment costs are moderately higher due to additional RSU intelligence modules, but the operational benefits justify this investment.

The results show that the proposed multi-layer DT ecosystem not only accelerates and stabilizes responsiveness but also fosters evolving autonomy and coordination at the edge. Importantly, the analysis makes trade-offs explicit: while deployment and operational costs rise moderately, they are outweighed by significant reductions in latency tails and backhaul load. This balance supports scalable and resilient governance in smart city traffic management, highlighting the broader potential of multi-layer DT ecosystems for future communication–sensing–computation networks.