1
OpenTwin: Digital Twin–Driven Closed-Loop KPM Inference and Control for Open RAN
arXiv:2605.24662v1 [cs.NI] 23 May 2026
Md Sharif Hossen*, Senior Member, IEEE, Zifan Zhang*, Dara Ron, Yuchen Liu, Member, IEEE, Vijay K. Shah, Senior Member, IEEE
Abstract—The open radio access network (O-RAN) RAN intelligent controller (RIC) hosts data-driven xApps and rApps to optimize network performance. However, two challenges hinder ML-driven xApp/rApp development: (i) key performance metric (KPM) data scarcity caused by interface latency, and (ii) network disruption risks when testing and validating AI models directly on live networks. We develop OpenTwin, a digital twin framework built on an open-source O-RAN simulator (ns-O-RAN-flexRIC) and KPM streaming via the O1 interface, deployed within the non-RT RIC. OpenTwin uses a two-step ML approach: an XGBoost model that learns time-varying network behavior to generate simulator configuration parameters, followed by a time-aware recursive least squares (RLS) tuner that continuously corrects KPM deviations between the twin and real-world measurements. A deviation-aware scoring mechanism monitors twin fidelity and automatically triggers resynchronization upon detecting network drift. We demonstrate OpenTwin with an energy-saving xApp that validates control policies in the virtual space before applying reconfigurations to the physical network. Experimental results show that OpenTwin mirrors real-world KPMs with up to 96% accuracy and enables the xApp to significantly reduce energy consumption without disrupting live operations. Index Terms—Open RAN, Digital Twin, Closed-loop Control, KPM Modeling, XGBoost, Recursive Least Squares, Network Synchronization, Energy-efficient RAN
I. I NTRODUCTION The next-generation open radio access network (O-RAN) is transforming the cellular ecosystem and is already being trialed and deployed for 5G networks in the United States and worldwide [1], [2]. Unlike traditional, vertically integrated cellular architectures, O-RAN promotes openness, programmability, interoperability, and virtualization, enabling operators to mix and match hardware and software components from different vendors [3]. This paradigm shift enhances flexibility, scalability, and cost efficiency for operators while fostering competition and innovation among vendors, ultimately benefiting consumers. The momentum is reflected in market projections: the O-RAN sector, valued at $1.69 billion in 2023, is expected to surpass $455 billion by 2036, with a 53.8% compound annual growth rate (CAGR) [4]. Md Sharif Hossen, Dara Ron, and Vijay K. Shah are with the Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC, 27695, USA (Email: {mhossen, dron, vijay.shah}@ncsu.edu). (Corresponding author: Vijay K. Shah.) Zifan Zhang and Yuchen Liu are with the Department of Computer Science, North Carolina State University, Raleigh, NC, 27695, USA (Email: {zzhang66, yuchen.liu}@ncsu.edu). *These authors contributed equally to this work.
O-RAN has defined two levels of RAN intelligent controllers (RIC), which enable software-defined, data-driven control of RAN [5], [6]. RIC collects telemetry, hosts modular control applications, and coordinates decision-making across distributed network elements. It operates across two complementary control loops: the near-real-time RIC (near-RT RIC) manages fast control loops (10 ms–1 s) via xApps, while the non-real-time RIC (non-RT RIC) leverages rApps operating at slower timescales (≥1 s). Both RICs also host third-party artificial intelligence (AI)-driven software microservices, which perform various RAN functionalities, including optimizing operational networks and costs and addressing various O-RAN use cases to enhance network performance. O-RAN Work Group 1 has released a technical report that demonstrates the critical role of AI-driven xApp/rApp in advancing ORAN capabilities across various use cases, including seamless handovers, radio resource optimization, quality of experience (QoE) enhancement, traffic steering, RAN slice service level agreement (SLA) assurance, dynamic spectrum sharing, and congestion prediction [7]. While the RIC’s programmable control plane marks a major advance over legacy wireless systems, we identify two categories of systemic challenges: (i) telemetry data scarcity and (ii) the safety of network testing and validation for AI. (i) Telemetry data scarcity: KPMs and I/Q samples collected from physical networks through open interfaces (O-Fronthaul, E2, and O1) provide critical data for effective network control and orchestration. However, data scarcity caused by latency constraints in near-RT and non-RT RICs limits the frequency of KPM updates, compromising timely data collection and network adaptation. (ii) Lack of validation, safety, and scalability: AI-driven control xApps/rApps require continuous KPM feedback over time (e.g., at sub-millisecond intervals) to validate and refine their models. However, when AI agents learn directly through interactions with physical networks via the E2 and O1 interfaces, the limited availability of real-time KPMs and the trialand-error nature of learning can lead to network instability or service disruption. Despite the advances enabled by the RF scenario server, a massive channel emulator (MCHEM), high-fidelity ray-tracing, and 3D mapping, several limitations persist. Synchronization between virtual and real networks under dynamic conditions, configuration changes, and/or network drift remains challenging due to high computational complexity. Furthermore, the high computational cost of continuous digital twin (DT) updates limits scalability for large-scale deployments.
2
DT technology has emerged as a complementary tool for addressing the above challenges [8], [9]. It represents an evolution of next-generation wireless networks, offering a virtual representation of physical networks that serves as a promising framework to facilitate network design, analysis, operation, automation, control, optimization, monitoring, and orchestration, particularly for AI training, testing, and validation [10], [11], [12], [13]. Telecom unions, network operators, and university labs, including ITU-R, O-RAN Alliance, and NVIDIA’s Sionna, have actively designed, developed, and outlined roadmaps for future DT development. ITU-T has defined the requirements and architecture of DT technology [14], while ITU-R has identified DT as a key 6G enabler in its report on future international mobile telecommunications (IMT) systems toward 2030 and beyond [15]. ITU-R recommends prioritizing a top-level DT design for the RAN before extending it beyond the RAN scope. NVIDIA has actively developed a high-fidelity DT platform called Sionna Omniverse [16], designed as an advanced simulator for next-generation wireless network development and deployment. It incorporates site-specific 3D maps (e.g., buildings and other structures) and employs raytracing techniques to model signal reflection, refraction, and RF propagation channels. The Sionna Omniverse provides a virtual environment for deploying base stations and users within the simulator to perform network analysis, modeling, and optimization offline before real network deployment. However, the NVIDIA DT operates entirely independently of physical networks, raising concerns regarding calibration errors, missing configuration parameters, and limited ability to mirror real network dynamics. The O-RAN Next Generation Research Group has released a research report on DT RAN use cases, specifying DT development directly connected to physical networks for AI/ML workflows such as model training, testing, and continuous optimization [17]. The Colosseum DT described in [18] relies on the CaST workflow to identify node locations and trajectories, generate 3D maps using OpenStreetMap, and extract RF propagation channels from MATLAB or Sionna ray-tracer outputs. It is coupled with an RF scenario server and an MCHEM to achieve better alignment of RF channels between the DT and the real network. SAJD [19], [20] introduces a self-adaptive jamming detection framework leveraging closed-loop learning in O-RAN. However, such approaches rely solely on real-world data and lack integration with digital twin environments. Recent efforts have also investigated security and privacy challenges in O-RAN. For example, ZT-RIC [21] proposes a zero-trust framework that enables privacy-preserving xApp inference over encrypted KPM data. They do not provide mechanisms for predictive network modeling, digital twin synchronization, or safe pre-deployment validation of control policies. We identify three key challenges when applying the existing DT tool within O-RAN system, which can be highlighted as follows. • Lack of full attention to all metrics: Beyond high-fidelity RF channels, a comprehensive set of metrics is essential for efficient network prediction, control, and orchestration. The metrics requiring careful attention include time/frequency
offset, bit rate, bit error rate (BER), modulation and coding scheme, received signal strength indicator, transmitted packet size, latency, signal-to-noise ratio (SNR), physical resource blocks (PRBs), buffer size, and more. They enable a deep understanding of the network demands and allow the system to intelligently manage and cross-layer control the network (NET), medium access control (MAC), and physical (PHY) layers to effectively meet performance objectives. • Limitations in flexibility under real network drift: Despite the advances enabled by the RF scenario server, MCHEM, high-fidelity ray-tracing, and 3D mapping, several limitations persist. Synchronization between virtual and real networks under dynamic conditions, configuration changes, and/or network drift remains challenging. The high computational complexity of the Sionna DT and the absence of predictive network configuration in Colosseum limit their flexibility and scalability in mimicking real network changes. • Absence of predictive capability: Previous studies have primarily focused on constructing high-fidelity DT but lack the predictive capability to forecast near-future network states. In this paper, we develop an OpenTwin, an intelligent DT framework for O-RAN that centers around addressing three challenges outlined above by providing All the Metrics You Need, coupled with high flexibility, scalability, and predictive capability for efficient RAN network control and orchestration and ever-enhancing DT mirroring. The contributions of our proposed OpenTwin are clearly highlighted as follows. • OpenTwin leverages a well-known network simulator (ns-O-RAN-flexRIC [22]) and KPM streaming via O1 to develop an intelligent and reliable DT within non-RT RIC, enabling communication between the virtual twin networks and the physical networks through open interfaces. This bidirectional exchange allows DT to collect KPMs, continuously adjust simulator configurations to adapt to the dynamic environment, and enable xApps/rApps to feed well-suited configurations back to the physical network so as to maximize network operational efficiency. • Building OpenTwin requires a predictive network configuration file (JSON) for the network simulator, along with tuned KPMs generated by the simulator. Therefore, we incorporate a two-step AI-driven approach into the proposed OpenTwin framework to build a robust virtualized network. First, an extreme gradient boosting (XGBoost) algorithm capable of fast learning and overfitting prevention is employed to capture historical patterns of the upcoming KPMs and generate wellsuited configurations for the simulator. Second, a time-series tuning algorithm, recursive least squares (RLS), is applied on top of XGBoost to validate the KPMs generated by the simulator and minimize deviations from those collected in the physical networks. Beyond KPM error mitigation, both XGBoost and RLS work together to capture network variations and predict near-future network states. • An energy-saving xApp is developed to demonstrate the performance improvements achieved by our OpenTwin. This xApp interacts with OpenTwin to manage and validate the networks within virtual space, aiming to save energy while maintaining service reliability, with a focus on throughput and latency that adhere to application requirements, without any
3
Fig. 1: O-RAN network architecture: Network elements, RIC microservices, and control loop time scale.
physical network disruption. • The experiments demonstrate that the ML models incorporated into OpenTwin, such as XGBoost and RLS, outperform other time-series learning schemes, including Random Forest, LSTM, and Transformer. Working together, these models achieve remarkable performance, enabling the OpenTwin to accurately mirror real-world networks with high fidelity and flexibility under network drift, approaching an accuracy level of up to 96%. Moreover, we demonstrate that the energysaving xApp can manage the network within these virtualized networks, significantly reducing energy consumption. II. BACKGROUND : O-RAN Fig. 1 illustrates the O-RAN architecture, including disaggregated components, functionalities, open interfaces, and closed-loop control time scales. O-RAN is a disaggregated RAN architecture that builds upon and complements 3rd Generation Partnership Project (3GPP) standards and specifications while extending them through open interfaces for third-party integration, network virtualization, and intelligent functions for control and optimization of the RAN [23], [24], [3]. The ORAN architecture further extends 3GPP’s New Radio (NR) by splitting the traditional monolithic base station (BS) into disaggregated components, including the open central unit (OCU), distributed unit (O-DU), and radio unit (O-RU) [25]. The O-RU handles radio front-end operations, digital frontend processing, lower physical-layer functions, and fronthaul transport. The O-DU is responsible for real-time RAN control, handling the high physical layer along with the medium access control (MAC) and radio link control (RLC) layers. It executes tasks such as scrambling, modulation, layer mapping, and resource block allocation and is connected to the O-CU via the F1 interface. The O-CU is split into two components: one for the control plane (CP) and another for the user plane (UP), which hosts the RAN functionalities, including (i) the radio resource control (RRC) layer, which manages connection lifecycles; (ii) the service data adaptation protocol (SDAP) layer, which handles quality of service (QoS) for traffic flows; and (iii) the packet data convergence protocol (PDCP) layer, which performs reordering, duplication, and encryption to ensure secure and reliable over-the-air communication. The O-RAN Alliance has also standardized other open interfaces, including E2, A1, and O1, which enable connectivity between the O-CU and/or O-DU and the RAN intelligent controller (RIC) [1], [19]. The near-RT RIC hosts multiple microservices, known as xApps, which perform
various RAN control and optimization functions, such as latency and throughput management [26], traffic steering and load balancing [27], dynamic spectrum sharing [28], attack detection [29], quality of service (QoS) enhancement [30], and intelligent network slicing [31]. Multiple near-RT RICs can stream telemetry data to a meta server (non-RT RIC) via the A1 interface. The non-RT RIC, as a component of the service management and orchestration (SMO) framework, hosts RAN applications, known as rApps, which govern all ORAN components (O-DU, O-CU, and near-RT RIC) connected to the SMO, enabling high-level decision-making and policy control to enhance over-the-air communications between cellular networks and thousands of users across large geographical regions. We leverage key features of the non-RT RIC, such as its high computational capability and large-scale control, to develop a high-fidelity OpenTwin, creating a virtual space for network testing, validation, and training.
III. P ROPOSED O PEN T WIN S YSTEM In this study, we develop OpenTwin, which incorporates data-driven ML models into scenario-driven network simulators (e.g., ns-O-RAN-flexric [22]) to construct a reliable and flexible OpenTwin within the non-RT RIC. As illustrated in Fig. 2, our framework establishes a DT-driven feedback loop, with assistance from key performance measurements (KPMs), between configuration inference, reconstruction, and network optimization. Unlike existing simulators that operate in an open-loop manner, this closed-loop method transforms the conventional static network simulator into an adaptive twin with high fidelity that can replicate key KPM distributions and monitor deviations as the O-RAN network evolves. ML models are responsible for observing variations and capturing patterns in physical O-RAN network states to automatically generate optimal network configurations, tune the synthetic KPMs from simulators to align with the distribution of real KPMs from physical networks, and predict near-future network conditions to provide a lookahead window for xApps to control and orchestrate the network before events occur. In the next few subsections, we will first briefly give a highlevel walkthrough of our proposed OpenTwin in Sec. III-A, introduce several key modules in Sec. III-B-III-C, and provide a use case to demonstrate the effectiveness of OpenTwin in Sec. IV.
4
Fig. 2: OpenTwin framework with closed-loop feedback between KPM generation, ML inference, and simulation updates. A. High-level Framework Walkthrough The proposed OpenTwin framework encompasses three main components: configuration file generation, KPM tuning and prediction, and network testing and validation within virtual networks. We first generate a custom configuration file regarding how many user equipments (UEs), next generation nodes (gNBs), bandwidth, packet size, buffer size, inter-site distance between cells, inter-site distance between UEs, mobility models, handover mode, outage threshold, and many other parameters as well. We then modify the network simulator to parse this configuration file and apply the specified parameters. ① Once the setup is ready and UEs communicate with the cells, we extract the raw KPMs generated by the simulator via O1 interface and store them in an InfluxDB time-series database. The RIC testing as a platform (RIC-TaaP) studio is used to monitor the network status and KPMs of the gNBs’ cells and UEs. ② OpenTwin initiates its first ML component to leverage eXtreme gradient boosting (XGBoost) to learn network traffic patterns from historical raw KPMs and generate appropriate network parameters for configuring the network simulator. However, the static and discrete snapshot configuration parameters generated by XGBoost may not be sufficient to achieve precise time-varying pattern alignment between the real network KPMs and OpenTwin due to network variations such as UE mobility, unknown interference, noise, traffic load fluctuations, and changing channel conditions. ③ To address this challenge, OpenTwin activates its second ML component, called the time-aware recursive least squares (RLS) tuner, a synchronization engine that tunes the simulator-generated KPMs to minimize their deviation from the raw KPMs collected from real networks. Depending on the deviation of the generated OpenTwin and the historical KPMs, we calculate an average score using the exponentially weighted moving average (EWMA). ④ A Supervisor Block, as shown in Fig. 2, consumes this score that contains several microservices. Network configuration regeneration (ConFileReg) monitors the EWMA score; if the score exceeds a threshold, it regenerates a new configuration file. An initial configuration file is always generated to start the simulator. When the microservice program ConFileReg produces a new configuration and signals that the OpenTwin needs to be retuned. Once the deviation between raw and DT-predicted
KPMs is minimal, another microservice will be waiting for sufficient samples of KPMs from OpenTwin to initiate the xApp for network performance validations, such as turning off underutilized cells. When OpenTwin concludes that applying a given xApp policy improves performance in simulation, it sends a deployment decision to the A1 termination node. The non-RT RIC then coordinates with the near-RT RIC to deploy the xApp on the physical network. B. XGBoost-Based Configuration Generation 1) KPM Preprocessing: In this initial stage, KPMs are collected through an O1 interface and processed into a unified, time-aligned feature map. The extracted feature map is essential for subsequent OpenTwin construction and evaluation. Let M denote the set of collected KPMs. We fix a sampling interval ∆ and index samples by t, where both predefined parameters are selected to capture relevant system dynamics while controlling signal variance. Note that we use (m) xt = [xt ]m∈M ∈ R|M| for the KPM vector at time t. For each KPM m ∈ M (e.g., SNR, PRBs, buffer size, modulation and coding scheme (MCS), or others), we construct a feature (m) sequence {xt }Tt=1 as input for the twinning process. A primary challenge is handling raw telemetry that arrives at nonuniform intervals (with a fixed time interval ≥ 1 s for the nonRT RIC). As a first step, we apply a low-pass smoothing filter and resample all data to the common interval ∆. This ensures downstream twin models operate on standardized inputs. We further establish a consistent timeline by aligning timestamps from all sources to a common clock and reconciling unit conventions. Minor temporal offsets are corrected by crosschecking shared change points and minimizing discrepancies among co-varying KPMs, by aligning prominent change points across KPMs, and by mitigating spurious shifts that could be misread as drift. This preprocessing step standardizes the temporal base but does not resolve all data inconsistencies. Deeper misalignments, such as missing values or data from sources with fundamentally heterogeneous sampling policies, are explicitly deferred. These issues will be addressed by the fully constructed DT, as discussed in later subsections. The OpenTwin itself is designed to reconcile these inconsistencies and input missing values, thereby providing a complete and robust system representation.
5
To represent both bursty behavior and long-term trends, we preprocess the raw KPMs to extract short-lag and smoothed features. In accordance with the tuner formulation introduced later, we define lagged terms {xt−1 , . . . , xt−τ }, along with windowed average, which can be defined as xt = PaW −1 1 i=0 xt−i . Note that, for a generic scalar KPM xt , we W drop the superscript m for simplicity. This averaging reduces noise variance by a factor of approximately 1/W at the cost of an average lag of (W − 1)/2 samples. This trade-off stabilizes the learning process while retaining responsiveness to system changes. We complement these features with rateof-change metrics (e.g., xt − xt−1 ), calendar indicators (e.g., time-of-day and day-of-week), and cross-KPM ratios where domain knowledge suggests stable inter-metric relationships. This ensures that both OpenTwin construction stage and the tuner stage operate on a consistent format. All features are standardized per KPM using statistics derived from the training set. Furthermore, data quality masks are retained to enable principled weighting during configuration inference and calibration. 2) Configuration File Generation: Given the preprocessed feature maps from Sec. III-B1, our goal is to infer a plausible O-RAN configuration that explains the observed KPMs and serves as an initial OpenTwin state. We denote by xt ∈ Rd the KPM feature vector at time t produced by a KPM generator, where the generator can be either a real O -RAN testbed via O1 or a historical KPM dataset. Each xt is constructed using the same lag and smoothing operators (e.g., window length W and windowed average xt ) as in Sec. III-B1. Similarly, we denote by yt ∈ Rd the KPM feature vector generated by the network simulator under a known O -RAN configuration, after applying the same preprocessing pipeline. The mapping of configuration file xt 7→ c is learned with gradient-boosted trees (XGBoost), chosen for their ability to capture nonlinear interactions, handle heterogeneous features, and support structure-imposing constraints. We use hundreds of feature pairs (yt , ct ) solely from network simulators to learn an inverse mapping from KPM features to configuration parameters during the training process, and then apply it to KPM generator features xt , which share a similar feature schema. We denote the configuration parameters as a mixed vector cont cat at time t ct = ccont , c , c ∈ Rp includes p bounded t t t continuous parameters (like transmit power, bandwidth, buffer sizes), and ccat t includes categorical parameters (like scheduler type or handover policy). Each parameter has box constraints [ℓj , uj ] for dimension j, while each categorical parameter takes values from a set {Vk }, with one set Vk for each category k. We jointly train regressors {fj }pj=1 for continuous parameters and classifiers {gk } for categorical ones by minimizing a composite loss L, and let T be the index set of simulatorgenerated training samples:
L=
X t
wt
" p X
αj LHuber fj (xt ), ccont j,t
X
j=1
# +
k
βk LCE gk (xt ), ccat k,t
(1)
where αj , βk > 0 weight parameters by operational criticality. The sample weight wt incorporates both generated KPM quality and temporal recency, which is given by: wt = qt λ T −t ,
λ ∈ (0, 1),
(2)
with qt ∈ {0, 1} derived from the quality masks in Section III-B1. The Huber loss provides robustness against outliers, while the cross-entropy loss for ccat utilizes class weights to mitigate imbalance. At the inference phase, we take a KPM feature vector xt from the KPM generator. We pass it through the trained XGBoost models to get raw configuration parameters. For the continuous parameters, we concatenate the outputs as: ⊤ (3) ĉcont = f1 (xt ), . . . , fp (xt ) ∈ Rp . t In addition, for each categorical parameter k, we get a probability vector: (k)
π̂t
= gk (xt ) ∈ [0, 1]|Vk | .
(4)
The raw model outputs ĉ are then projected to a feasible configuration file to ensure schema validity. For continuous parameters, we solve a small quadratic program (QP): c̃cont = arg min ∥z − ĉcont ∥22 z
s.t. ℓj ≤ zj ≤ uj , Az ≤ b.
This enforces both box constraints (ℓj , uj ) and linear coupling constraints (Az ≤ b). For categorical parameters, we take the calibrated Maximum A Posteriori (MAP) label, c̃cat k = arg maxv∈Vk Pr gk (x) = v . The result c̃ is a schema-valid configuration file ready to initialize OpenTwin in the O-RANcompliant manner. This approach provides a fast, synthetic data-driven estimate of the configuration state via supervised inversion, which is then refined using simulation. The use of monotone and feasibility constraints ensures that predictions remain plausible. Because the inference features yt , although collected in the network simulator, reuse the same KPM format defined in Sec. III-B1, the inferred configuration c̃ is immediately consumable by the time-aware calibration module (Section III-C). This design yields an initialization that is feasible and wellconditioned for continuous synchronization against the real testbed. C. Time-Aware Recursive Least Square Tuner KPMs produced by the simulator with a generated configuration file usually differ from ground-truth measurements obtained from KPM generators due to modeling simplifications, stochastic network dynamics, measurement noise, and unpredictable user activities. To synchronize OpenTwin with the ground-truth measurements, we design an online correction tuner that adjusts DT-produced raw KPMs toward the ground-truth while tracking nonstationarity, e.g., slow drifts and regime changes. Note that the number of ground-truth KPM samples is small due to the high collection expense, so we adopt recursive least squares (RLS) as a data-efficient online tuning method. Let xt denote the ground-truth KPM preprocessed in Sec. III-B1 at time t and yt the corresponding
6
OpenTwin output. The instantaneous residual is rt = xt − yt , where our goal is to approximate the time-varying relation. W −1
xt ≈ g yt , yt−1 , . . . , yt−τ , y t ,
yt =
1 X yt−i , (5) W i=0
where the windowed average y t reduces high-frequency noise variance by approximately 1/W at the cost of an average lag of (W − 1)/2 samples. This feature set captures shortterm dynamics and longer-term trends that are predictive of systematic bias. We implement g(·) as a linear, adaptive filtering model trained with RLS with exponential forgetting. We define the feature vector ut as: T (6) ut = 1, yt , yt−1 , . . . , yt−τ , y t ∈ Rτ +3 , and the one-step prediction is given by: x̂t = uTt θt−1 ,
(7)
where θt ∈ Rτ +3 is the time-varying vector of correction coefficients. With forgetting factor λ ∈ (0, 1) and inverse covariance matrix Pt ∈ R(τ +3)×(τ +3) , the RLS updates are: Pt−1 ut , λ + uTt Pt−1 ut et = xt − uTt θt−1 ,
kt =
(8) (9)
θt = θt−1 + kt et , (10) 1 Pt = Pt−1 − kt uTt Pt−1 , (11) λ which is initialized with a small ridge term γ > 0 via P0 = γ −1 I to stabilize early updates under collinearity. Equivalently, at each time t the RLS solution θ̂t minimizes the exponentially weighted, ℓ2 -regularized objective " t # X 2 t−s T 2 θ̂t = arg min λ xs − us θ + ρ∥θ∥2 , (12) θ
s=1
where ρ > 0 is a ridge regularization coefficient. The corrected KPM estimate is then x̂t ≡ ŷt = uTt θ̂t , applied perKPM in parallel (each KPM maintains its own θt and Pt ). In practical settings, we optionally trigger the correction to enhance robustness under abrupt regime changes, letting dt denote a smoothed deviation defined in Section III-C1 and τm an acceptance threshold, we use ( x̂t , if |dt | ≤ τm , ŷt = (13) yt , otherwise, so that large, sudden discrepancies trigger conservative behavior until the tuner adapts. The forgetting factor λ controls the speed–stability trade-off: smaller values react quickly to shifts but increase variance; larger values favor smooth tracking of slowly varying biases. The choice of window W plays a complementary role through the bias–variance balance of y t . Together, the RLS updates and feature design provide a lightweight, online mechanism to continuously reduce DTreference gaps while respecting the nonstationary nature of the network.
1) Deviation Scoring and Continuous Synchronization: We continuously monitor the fidelity of OpenTwin with a single deviation signal that averages smoothed, relative errors across all KPMs, and then uses this signal to decide whether to retune the tuner, adjust cadence, or regenerate the configuration. Let (m) xt be the ground-truth value for KPM m at time t, and (m) (m) let yt and ŷt be the raw and RLS–corrected OpenTwin outputs. To make deviations comparable across heterogeneous scales, we work with relative percentage errors: (m)
∆raw m,t = 100
yt
(m)
(m)
− xt
, (m) max |xt |, ϵ
∆corr m,t = 100
ŷt
(m)
− xt
, |, ϵ (14) where ϵ > 0 is a small constant to avoid division by very small denominators. These instantaneous gaps can fluctuate rapidly, so each KPM stream is smoothed with an exponentially weighted moving average, dm,t = αm ∆corr m,t + (1 − αm ) dm,t−1 ,
(m)
max |xt
αm ∈ (0, 1), (15)
where larger αm favors recency and smaller αm favors stability. In practice, αm is set from a target half-life Hm at the working cadence ∆ via αm = 1 − 2−∆/Hm so that smoothing has a clear temporal interpretation. To prevent positive and negative errors from canceling out and to reflect KPM criticality, we aggregate absolute EWMAs with nonnegative weights wm , P wm |dm,t | d¯t = mP , (16) m wm Weights can encode operational importance and also compensate for noise scale differences. A practical choice is wm ∝ ωm /b σm , where ωm is a priority coefficient and σ bm is a robust scale estimate for ∆corr (for example, median absolute m,t deviation), which reduces domination by inherently highvariance KPMs. When some KPMs are missing or flagged as low quality by the masks from Section III-B1, we compute (16) over the available subset with renormalized weights and defer updates if coverage falls below a minimum threshold. For a single, interpretable decision variable we map d¯t to a normalized score St ∈ [0, 1] using global warning and alarm thresholds 0 ≤ τ warn < τ alarm , 0, d¯t ≤ τ warn , d¯t ≥ τ alarm , (17) St = 1, warn ¯ dt − τ , otherwise. τ alarm − τ warn In our implementation, we use a fixed choice of τ warn and τ alarm for regeneration alarms, initialized from offline analysis of representative traces with expert input. This single-score design naturally focuses on downstream measurements because the KPM set already includes metrics tied to the downstream tasks, such as energy and load-related measurements. Improvements or regressions in those outcomes directly influence d¯t and therefore St , so the same synchronization mechanism governs both OpenTwin fidelity and task-relevant behavior. Algorithm 1 summarizes the complete closed-loop operation. At each time step, the framework preprocesses incoming
7
Algorithm 1 OpenTwin Closed-Loop Framework 1: Input: Real-world KPM stream {xt }, window W , thresh-
Fig. 3: Comparison of ML model accuracy. KPMs and evaluates the deviation score St from (17). If St−1 exceeds the alarm threshold τ alarm , XGBoost regenerates the configuration file and restarts the simulator (lines 4-8); otherwise, the RLS tuner corrects the simulator-produced KPMs toward the real-world measurements (lines 11-13). When St falls below the warning threshold τ warn , the twin is considered faithful and the xApp is permitted to execute control actions (lines 17-19). The deviation tracking in (15)(17) runs continuously throughout the operation, keeping the twin aligned with the real testbed as load, mobility, and radio conditions evolve. We also evaluated sequence-based models, such as LSTMs and Transformers, using historical KPMs to predict OpenTwin KPMs. However, as illustrated in Fig. 3, these models struggle to track rapid PRB fluctuations and require fixed sampling intervals and large training sets, which is not realistic for KPM streams delivered by the O1 interface and is very time-consuming in practice. By contrast, RLS is lightweight, data-efficient, and responsive, making it more suitable for continuous OpenTwin synchronization. IV. U SE C ASE : E NERGY-S AVING X A PP In this study, we implement the energy-saving xApp to demonstrate how this control xApp employs our proposed OpenTwin to observe network status for efficient control and orchestration to save energy consumption. As illustrated in Fig. 2, the control xApp continuously monitors per-UE serving and neighbor SINR reports via A1 interface. These reports help as UE–cell attachments to determine when the xApp will trigger handovers and cell sleep transitions through RIC indication to control the RAN. When the UE’s serving-cell SINR drops below a smoothed neighbor SINR, the xApp triggers a controlled handover to the neighbor with the highest averaged SINR cell. After all associated UEs are handed over, if a cell does not serve any UE, xApp will send an RIC indication to change the cell status to Sleep state to reduce energy consumption. If a cell has at least one UE, xApp will treat that cell as active and will not change its state. Along with this, sleeping cells are also periodically probewoken for a short interval to detect new user activity. If there are no changes in SINR, the cell goes back into the Sleep state again. This ensures energy savings without compromising connectivity. V. R ESULTS AND D ISCUSSION In this section, we discuss the evaluation of fidelity, adaptability, and the effectiveness of control in the proposed
olds τ warn , τ alarm , RLS factor λ 2: for t = 1 to T do 3: Preprocess xt to form inference features vt (for XGBoost) 4: if St−1 ≥ τ alarm then 5: ĉ ← XGBoost(vt ) 6: c̃ ← arg minz ∥z− ĉ∥22 s.t. ℓ ≤ z ≤ u, Az ≤ b 7: Restart the network simulator with c̃ and reset Pt , θt 8: end if 9: ytsim ← network simulator(c̃) 10: Construct tuning feature ut ← [1, ytsim , . . . ]T based on W 11: kt ← Pt−1 ut (λ + uTt Pt−1 ut )−1 12: θt ← θt−1 + kt (xt − uTt θt−1 ) 13: Pt ← λ−1 (Pt−1 − kt uTt Pt−1 ) 14: ŷt ← uTt θt ← |(ŷt − xt )/xt |; dt ← αdt−1 + (1 − α)∆raw 15: ∆raw t t 16: Update global deviation score St based on dt 17: if St ≤ τ warn then 18: Trigger xApp logic control based on valid twin state 19: end if 20: end for 21: Output: Optimized network control actions
OpenTwin framework with the four steps: (i) DT fidelity, (ii) adaptation under network drift, (iii) configuration regeneration via deviation score, and (iv) the energy-saving xApp-based RAN control. A. Experimental Setup 1) Network Simulator: Network Simulator 3 (ns-3) is an open-source, packet-level, discrete-event simulation platform developed for the scientific community, industry, and educational purposes. This simulator supports high-fidelity virtualized experiments that are compliant with 3GPP standards and the Wi-Fi/LTE/5G protocol stack [32]. Its source code can be reused on real-world testbeds, demonstrating the minimal gap between simulation and prototype implementations. This opensource application can be deployed and run on RICs (near-RT RIC or non-RT RIC) to bridge the connection between 5G simulations in ns-3 and physical networks. This integration allows KPMs, including realistic channel data, to be streamed to the RICs via open interfaces (O-FH, F1, O1, E2, and A1), supporting the control actions of AI-driven software microservices called xApps and rApps. An ns-O-RAN simulator has been introduced as an extension of OpenRAN Gym [33] that integrates the ns-3 simulator with an E2-compliant RIC [34], [35]. Orange has leveraged the FlexRIC from EURECOM, the ns-O-RAN simulator from WIoT, the 5G-LENA module from CTTC, and the Sionna Ray Tracing tool from Nvidia to develop ns-O-RAN-flexric, an open-source for testing ecosystem initiative designed to streamline xApp/rApp functional and
8
operational testing, foster innovation in xApp/rApp design, and provide proven digital-twin networks [22]. Therefore, the network simulator block in the proposed OpenTwin framework uses this ns-O-RAN-flexRIC simulator, running within the nonRT RIC, to conduct the experiments. 2) Digital Twin Scenario: Our developed OpenTwin runs on top of ns-O-RAN-FlexRIC and utilizes the mmWave module. Note that the mmWave module only provides related attributes to our system, but OpenTwin does not directly run on mmWave band. The simulated network includes one LTE eNB and N number of gNB mmWave sites. The LTE eNB and one mmWave gNB are placed at the center of a 4 km by 4 km area, while the other mmWave cells are set up on a ring whose intersite distance is controlled by a parameter named IntersideDistanceCells. We apply the 3GPP UMi street canyon model inside ns-3 to emulate both path loss and channel conditions. In our scenarios, all antenna elements are modeled as isotropic, while the carrier frequency and bandwidth are set through the DT configuration parameters. On the other hand, the number of UEs, their spacing, and mobility patterns are set up from the configuration file that is generated by the inference step in Sections III-B and V-A5. UEs are uniformly distributed within a disc of radius given by IntersideDistanceUEs centered on the deployment, and their movements are triggered following the mobility model specified by the parameter MobilityModel. For example, if RandomDirection2dMobilityModel is set up, UEs move towards the random direction within a 2D plane, with speeds uniformly distributed within [Speed Min, Speed Max]. The bounds of the mobility area span the whole simulation area. Finally, handover behavior is controlled by the hoSinrDifference, handoverMode, and outageThreshold parameters, which map to the radio resource control (RRC) configuration used for the co-located gNBs. Each radio link control (RLC) entity uses a transmit buffer size determined by the bufferSize parameter. 3) KPM Collection: We need to collect the KPMs properly to synchronize them in the DT and the historical KPM generator, also referred to as real-world (RW) KPMs. We collect both cell-level and UE-level KPMs from the baseline ns-3 simulation, which serves as a synthetic ground truth. For the cell level, we collect KPMs such as downlink PRB usage, the number of active UEs, PDCP traffic volume, and detailed transport block statistics, which include the number of transmitted transport blocks (TB), counts of 16-quadrature amplitude modulation (16QAM), 64QAM, and quadrature phase shift keying (QPSK) modulation formats, and TB error events. While for the UE level, we log each UE’s PRB allocation, downlink PDCP throughput, serving and neighbor SINR, buffer status metrics, and PDCP volume based on the QoS filters. These KPMs are exported through the O1 interface, which imitates the behavior and periodicity of O1 reporting. These are then stored with timestamps and cell or UE identifiers in InfluxDB. These time series datasets emulate scheduler actions, traffic dynamics, mobility patterns, and link quality changes, which are then used as an input for the XGBoost-based configuration generation and time-aware RLS tuner.
4) Traffic and E2 Integration: To drive user traffic, we connect one remote host to the packet gateway (PGW) via a high-capacity point-to-point link. Each UE is served by one such saturated downlink user datagram protocol (UDP) flow from this remote host. Packets are sent every 500 microseconds by default, as the packet size is taken from the inferred PacketSize parameter. This sets up a constant load that stresses the scheduler and emphasizes how radio and mobility parameters impact PRB usage, PDCP volume, and TB error events. The ns-3 deployment is connected to FlexRIC, so it can expose E2 KPMs that match the real-world O1 telemetry. E2 reporting for DU, CU-CP, and CU-UP is turned on in accordance with configuration flags and the chosen indication period. The DT exports its KPMs according to the same data schema as the real network, which enables a deviation score computation. 5) Configuration File Prediction: Using a logical or inference approach, we can extract several configuration fields from the RW KPMs, such as the number of gNBs and UEs, node positions, E2 function flags, and the reporting interval. For the remaining radio, mobility, and buffering parameters that cannot be logically inferred from the KPM structure, we only apply ML to predict them. To generate those remaining fields from the RW KPMs, we compare two approaches, random forest (RF) and XGBoost, and we evaluate which one more accurately predicts the configuration file. Table I illustrates that RF significantly struggles with several key fields, such as bandwidth, packet size, and speed, often with large errors. In contrast, XGBoost recovers almost all continuous parameters with very high accuracy, which are rounded to match the corresponding RW KPM-derived configurations. XGBoost also shows good accuracy in terms of predicting categorical parameters, e.g., handover mode and mobility model. As illustrated in Table II, XGBoost matches closely with the ground truth configuration, while RF shows noticeable deviations. Given the importance of XGBoost’s model capability to reproduce the continuous radio and mobility settings, we consider XGBoost models to predict the configuration file in our DT. TABLE I: Example of continuous parameter inference error using RF and XGBoost. Parameter OutageThreshold Bandwidth CenterFrequency IntersideDistanceCells IntersideDistanceUEs Speed Min Speed Max PacketSize BufferSize HoSinrDifference IndicationPeriodicity
Random Forest MAE MAPE 0.61 12.2% 4.35e6 17.4% 4.768e7 1.20% 11 1.83% 6 6.0% 0.865 43.25% 2.83 56.6% 134 104% 3.1 6.2% 0.52 10.4% 0.032 3.2%
XGBoost MAE MAPE 0.0007 0.014% 0 0% 896 0.00002% 0 0% 0.000259 0.000259% 0.000034 0.0017% 0.000078 0.00156% 0.00043 0.00034% 0.00086 0.0017% 0.000427 0.0085% 0.000023 0.0023%
B. Experimental Results 1) OpenTwin Fidelity: Fig. 4 demonstrates the high fidelity of the proposed OpenTwin using three complementary results: (a) a comparison of KPMs collected from real networks and those generated by the DT over time, (b) a scatter
30 20 10
20 10 0
0 0
100
200
300
Time (s)
400
0
500
(a) RW serving-cell SINR over time for UE2.
10
20
RW SINR (dB)
30
96 .8%
97 .3%
96 .7%
97 .1%
100
UE4 UE5 UE6 Ideal match
95 .5%
40
30
DT accuracy [%]
RW DT
DT SINR (dB)
ue_2_l3 serving sinr
50
96 .6%
9
80 60 40 20 0
UE1 UE2 UE3 UE4 UE5 UE6
UE
(b) Scatter plot of DT-generated SINR versus (c) DT accuracy for serving–cell SINR for 6 UEs. real-world SINR.
Fig. 4: Serving-cell SINR comparison between real-world (RW) and Digital Twin (DT). (a) Time-series comparison for UE2. (b) Scatter plot of DT versus RW SINR for representative UEs, where the diagonal line indicates ideal agreement. (c) DT accuracy across multiple UEs. TABLE II: Example configuration reconstruction from RW KPMs. Parameter OutageThreshold Bandwidth (Hz) CenterFrequency (Hz) IntersideDistanceCells (m) IntersideDistanceUEs (m) Speed Min (m/s) Speed Max (m/s) PacketSize (bytes) BufferSize HoSinrDifference IndicationPeriodicity (s) MobilityModel HandoverMode
Ground True -5 25000000 3970000000 600 100 2 5 128 50 5 1.0 RD2MM FixedTtt
Random Forest -4.39 29350000 3922320000 611 106 2.865 7.83 262 46.9 4.48 0.968 RD2MM FixedTtt
XGBoost -4.9993 25000000 3969999104 600 100.000259 2.00003387 5.00007821 127.99957 50.00086 4.99957 1.000023 RD2MM FixedTtt
plot illustrating pointwise agreement between DT-generated and real-world SINR, and (c) the accuracy of user-specific KPMs generated by the DT compared to those from the real networks. These detailed results demonstrate the high fidelity of the proposed OpenTwin. The experimental setup is described in Subsection V-A, where the number of active users is set to 6, and the RandomDirection2dMobilityModel is set up to simulate UE mobility within a two-dimensional plane. The OpenTwin activates its ML models to learn the dynamics of the network changes in order to generate a well-suited configuration file and tune the KPMs to closely align with those from real networks. The experimental results demonstrate that (a) the SINR generated by the OpenTwin (· black dots) almost perfectly coincides with the real-world SINR (· red dots), (b) the scatter plot of DT-generated SINRs versus real-world SINRs shows strong pointwise agreement, with most data points closely aligned along the ideal diagonal line, and (c) the SINR accuracy of the OpenTwin reaches approximately 96% compared to the real-world measurements. These results showcase the robustness of our OpenTwin in generating network representations within a virtual space for control xApps/rApps. 2) Adaptability and Flexibility of OpenTwin: This experiment aims to demonstrate the flexibility of the OpenTwin in maintaining synchronization with real networks under network drift. The network is configured and scheduled to activate 10
UEs over a time interval of [0, 480] seconds, and 4 of them are scheduled to become inactive at the predetermined time of 480 seconds. As illustrated in Fig. 5(a), the PRB usage drops sharply at this point due to those UEs leaving the network. The results demonstrate that the OpenTwin adapts effectively to network dynamics, with its KPM (PRB usage) almost perfectly overlapping the measurements obtained from the real-world network over time. This achievement highlights the key features of the OpenTwin models: overfitting prevention, fast convergence, and the ability of XGBoost and RLS to adapt to time-varying networks. Fig. 5(b) summarizes the mean absolute percentage error (MAPE) across a broad set of KPMs to quantify the overall fidelity of the DT. The KPMs include transport block statistics, modulation formats, buffer status, PRB usage, and downlink PDCP volume and altogether span physical, MAC, and PDCP layer characteristics. Results show that the DT maintains low error across all metrics, as most KPMs show MAPE values in the range of approximately 3 to 6 percent. Of these, lower-level metrics such as QPSK and error-related TB counters show particularly small deviations, while more dynamic metrics like PRB usage and PDCP volume have slightly higher MAPE values because those KPMs have larger absolute magnitudes. Across these KPMs, the DT maintains roughly 93% to 98% accuracy, confirming that it stays tightly synchronized with the real network even as the number of active UEs changes. 3) Regeneration Triggering via Deviation Score: To instruct the regeneration of the configuration file in scenarios when the DT cannot capture the RW KPMs, we compare the deviation score in (17) over time as shown in Fig. 6 (a). It shows that the deviation score S changes over iterations with respect to a threshold. For example, for a threshold of 0.15, the score remains below the limit for most of the run, which indicates that DT is well aligned with the RW. 4) OpenTwin-Assisted Energy Saving xApp: This experiment demonstrates how the energy-saving xApp employs our OpenTwin to manage the network in a way that reduces energy consumption. We measure the energy consumption when the energy-saving xApp is set to both active and inactive modes to clearly illustrate the amount of energy that can be
10
3.51%
4
5.84%
5.83%
5.81%
5.81%
5.10%
2.67%
80
3.88%
5
MAPE [%]
100
4.64%
6
RW DT
3
40
0.52%
1
1.01%
2
60
(a) DL PRB usage versus time for UE3.
e l3 ser id vin .bu gs ffe inr rsi tb. ze. tot qo nb s.u rdl eid ini tb. tia tot l . 1 nb 6q rdl am ini tia l.6 4 qa tb. m tot nb rdl qo ini sfl tia ow rru l .pd .pr cp bu pd sed uv dl olu me dl_ filt er dlp rbu sag e
nb
rdl .
1.u
ue tb.
tot
brd
tot
aln
drb
800
err
600
tb.
400
Time (s)
nb
200
tot
0
rdl i
nit ia
l.q
ps
20
l.1 .
k.u e
id
id
0
tb.
ue_3_dlprbusage
7
KPI
(b) Mean absolute percentage error versus KPM.
Fig. 5: OpenTwin adaptability under network drift.
(a) Deviation score S over iterations.
(b) Energy consumption versus KPM.
Fig. 6: Total energy consumption in DT and real-world under xApp active and inactive.
saved. The dashed lines represent the scenario in which the energy-saving xApp is inactive. We observe that the energy consumption is high in this case. When the energy-saving xApp is activated, this control xApp immediately observes the network within the virtual space and interacts with the OpenTwin via the A1 interface to orchestrate the virtualized network. A logic-based configuration approach is applied in this experiment, in which the xApp periodically detects cells without active UEs and puts them into sleep mode, a state in which power consumption is reduced. The xApp then sends a control message to the real network via the E2 interface to implement the necessary changes according to the network reconfiguration in the OpenTwin. Fig. 6 (b) demonstrates that the xApp significantly reduces energy consumption. Notably, the energy consumption of the OpenTwin and the real-world network closely overlap, demonstrating that the energy savings in the real implementation are consistent with the performance achieved within the OpenTwin. This result sufficiently highlights the value of our OpenTwin in providing a virtual space that allows xApps to control and manage the network without any disruption to real operations.
VI. C ONCLUSION In this paper, we develop OpenTwin, a simulator-assisted DT framework that allows for closed-loop configuration inference, KPM calibration, and downstream network control within an O-RAN–compliant architecture. Through an XGBoost-based configuration generation and a time-aware RLS tuner for continuous KPM generation calibration, our OpenTwin emulates the target O-RAN environment and adapts to dynamic changes in mobility, load, and radio conditions. An additional deviation-aware, lightweight supervisor microservice ensures fidelity assessment and automatic regeneration upon detection of significant network drift. Through extensive experiments, we demonstrate that OpenTwin reproduces KPM streams, including serving-cell SINR and PDCP volume over varying load conditions, along with accurate predictions of the configuration parameters and KPM evolution over time. We also illustrate the capability of OpenTwin to host downstream control applications, such as an energy-saving xApp, to validate control policies safely and repeatably without a physical deployment. This ensures our approach is a pragmatic and extensible foundation for self-calibrating DTs in O-RAN systems. Because of its modular architecture, high fidelity
11
under dynamic conditions, and compatibility with existing O-RAN interfaces, it can be a useful platform for research on ML-assisted RAN control, xApp/rApp development, and further work toward integrating live telemetry and partial realworld synchronization. ACKNOWLEDGMENT This work is supported in part by the NSF #2120411 and #2443035, CNS #2440756, and CNS #2312138. R EFERENCES [1] N. D. Tripathi and V. K. Shah, Fundamentals of O-RAN. John Wiley & Sons, 2025. [2] A. S. Abdalla, P. S. Upadhyaya, V. K. Shah, and V. Marojevic, “Toward next generation open radio access networks: What O-RAN can and cannot do!” IEEE Netw., vol. 36, no. 6, pp. 206–213, 2022. [3] J. F. Santos, A. Huff, D. Campos, K. V. Cardoso, C. B. Both, and L. A. DaSilva, “Managing O-RAN networks: xApp development from zero to hero,” IEEE Commun. Surveys Tuts., 2025. [4] Research nester global market size, forecast, and trend highlights over 2024 - 2036. Accessed:2026-03-25. [Online]. Available: https://www.researchnester.com/reports/open-radio-accessnetwork-market/2781 [5] B. Agarwal, R. Irmer, D. Lister, and G.-M. Muntean, “Open RAN for 6G networks: Architecture, use cases and open issues,” IEEE Commun. Surveys Tuts., 2025. [6] G. Reus-Muns, P. S. Upadhyaya, U. Demir, N. Stephenson, N. Soltani, V. K. Shah, and K. R. Chowdhury, “SenseORAN: O-RAN-based radar detection in the CBRS band,” IEEE J. Sel. Areas Commun., vol. 42, no. 2, pp. 326–338, 2023. [7] O-RAN Alliance, “O-RAN massive MIMO use cases technical report 1.0,” O-RAN Alliance, Tech. Rep., 2022, release v01.00. [Online]. Available: https://specifications.o-ran.org/specifications [8] P. Almasan, M. Ferriol-Galmés, J. Paillisse, J. Suárez-Varela, D. Perino, D. López, A. A. P. Perales, P. Harvey, L. Ciavaglia, L. Wong et al., “Network digital twin: Context, enabling technologies, and opportunities,” IEEE Commun. Mag., vol. 60, no. 11, pp. 22–27, 2022. [9] Z. Zhang, Z. Peng, H. Yu, M. Chen, and Y. Liu, “Digital network twins for next-generation wireless: Creation, optimization, and challenges,” IEEE network, 2025. [10] ——, “Digital network twins for next-generation wireless: Creation, optimization, and challenges,” arXiv preprint arXiv:2410.18002, 2024. [11] T. Li, Q. Long, H. Chai, S. Zhang, F. Jiang, H. Liu, W. Huang, D. Jin, and Y. Li, “Generative AI empowered network digital twins: Architecture, technologies, and applications,” ACM Comput. Surveys, vol. 57, no. 6, pp. 1–43, 2025. [12] S. M. Raza, R. Minerva, N. Crespi, M. Alvi, M. Herath, and H. Dutta, “A comprehensive survey of network digital twin architecture, capabilities, challenges, and requirements for edge–cloud continuum,” Comput. Commun., vol. 236, p. 108144, 2025. [13] Z. Zhang, M. Fang, D. Chen, X. Yang, and Y. Liu, “Synergizing AI and digital twins for next-generation network optimization, forecasting, and security,” IEEE Wireless Commun., vol. 32, no. 3, pp. 98–105, 2025. [14] ITU-T-3090, “Digital twin network – requirements and architecture,” International Telecommunication Union, Recommendation Y.3090, 2 2022, available: https://www.itu.int/rec/T-REC-Y.3090-202202-I/en. [15] ITU-R WP 5D, “IMT towards 2030 and beyond (IMT-2030),” https://www.itu.int/en/ITU-R/study-groups/rsg5/rwp5d/imt-2030/pages/ default.aspx, 2023, accessed: 2025-11-30. [16] J. Yang, M. Andersin, T. Balercia, and C. Chong, “Developing nextgeneration wireless networks with NVIDIA aerial omniverse digital twin,” https://developer.nvidia.com/blog/developing-next-gen-wirelessnetworks-with-nvidia-aerial-omniverse-digital-twin/, 2024, accessed: 2025-11-30. [17] M. C. Kirana, P. Keyela, F. Rostamian, D. H. Tashman, and S. Cherkaoui, “ML-enabled open RAN: A comprehensive survey of architectures, challenges, and opportunities,” IEEE Commun. Surveys Tut., vol. 28, pp. 4948–4979, 2026. [18] M. Polese, L. Bonati, S. D’Oro, P. Johari, D. Villa, S. Velumani, R. Gangula, M. Tsampazi, C. Paul Robinson, G. Gemmi, A. Lacava, S. Maxenti, H. Cheng, and T. Melodia, “Colosseum: The open RAN digital twin,” IEEE Open J. Commun. Soc., vol. 5, pp. 5452–5466, 2024.
[19] M. H. Rahman, M. S. Hossen, N. H. Stephenson, V. K. Shah, and A. Da Silva, “SAJD: Self-adaptive jamming attack detection in AI/ML integrated 5G O-RAN networks,” in Proc. IEEE Military Commun. Conf. (MILCOM), 2025, pp. 1130–1135. [20] M. H. Rahman, M. S. Hossen, N. H. Stephenson, V. K. Shah, and A. D. Silva, “A demonstration of self-adaptive jamming attack detection in AI/ML integrated O-RAN,” in Proc. IEEE Military Commun. Conf. (MILCOM), 2025, pp. 901–902. [21] D. Lin, S. Bhargav, A. Chiejina, M. I. Ibrahem, and V. K. Shah, “ZT-RIC: A zero trust RIC framework for ensuring data privacy and confidentiality in open RAN,” in Proc. IEEE Consum. Commun. Netw. Conf. (CCNC), 2025, pp. 1–7. [22] EURECOM, “RIC-TaaP: RIC testing as a platform,” Accessed: 202603-25. [Online]. Available: https://github.com/Orange-OpenSource/nsO-RAN-flexric [23] M. Polese, L. Bonati, S. D’oro, S. Basagni, and T. Melodia, “Understanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,” IEEE Commun. Surveys Tuts., vol. 25, no. 2, pp. 1376–1411, 2023. [24] K. Alam, M. A. Habibi, M. Tammen, D. Krummacker, W. Saad, M. Di Renzo, T. Melodia, X. Costa-Pérez, M. Debbah, A. Dutta et al., “A comprehensive tutorial and survey of o-ran: Exploring slicing-aware architecture, deployment options, use cases, and challenges,” IEEE Communications Surveys & Tutorials, 2025. [25] J. L. Herrera, S. Montebugnoli, D. Scotece, L. Foschini, and P. Bellavista, “A tutorial on O-RAN deployment solutions for 5G: From simulation to emulated and real testbeds,” IEEE Commun. Surveys Tut., 2025. [26] A. Lacava, M. Polese, R. Sivaraj, R. Soundrarajan, B. S. Bhati, T. Singh, T. Zugno, F. Cuomo, and T. Melodia, “Programmable and customized intelligence for traffic steering in 5G networks using open RAN architectures,” IEEE Trans. Mobile Comput., vol. 23, no. 4, pp. 2882–2897, 2023. [27] F. Kavehmadavani, V.-D. Nguyen, T. X. Vu, and S. Chatzinotas, “Intelligent traffic steering in beyond 5G open RAN based on LSTM traffic prediction,” IEEE Trans. Wireless Commun., vol. 22, no. 11, pp. 7727– 7742, 2023. [28] L. Bonati, M. Polese, S. D’Oro, S. Basagni, and T. Melodia, “NeutRAN: An open RAN neutral host architecture for zero-touch RAN and spectrum sharing,” IEEE Trans. Mobile Comput., vol. 23, no. 5, pp. 5786–5798, 2023. [29] A. Scalingi, S. D’Oro, F. Restuccia, T. Melodia, and D. Giustiniano, “Det-RAN: Data-driven cross-layer real-time attack detection in 5G open RANs,” in Proc. IEEE INFOCOM Conf., 2024, pp. 41–50. [30] A. Wadud, F. Golpayegani, and N. Afraz, “QACM: QoS-aware xApp conflict mitigation in open RAN,” IEEE Trans. Green Commun. Netw., vol. 8, no. 3, pp. 978–993, 2024. [31] A. Abouaomar, A. Taik, A. Filali, and S. Cherkaoui, “Federated deep reinforcement learning for open RAN slicing in 6G networks,” IEEE Commun. Mag., vol. 61, no. 2, pp. 126–132, 2022. [32] G. F. Riley and T. R. Henderson, “The ns-3 network simulator,” in Model. Tools Netw. Simul. Springer, 2010, pp. 15–34. [33] L. Bonati, M. Polese, S. D’Oro, S. Basagni, and T. Melodia, “OpenRAN Gym: An open toolbox for data collection and experimentation with AI in O-RAN,” in Proc. IEEE Wireless Commun. Netw. Conf., 2022, pp. 518–523. [34] A. Lacava, M. Bordin, M. Polese, R. Sivaraj, T. Zugno, F. Cuomo, and T. Melodia, “ns-O-RAN: Simulating O-RAN 5G systems in ns-3,” in Proc. Workshop ns-3, ser. WNS3 ’23. NY, USA: Association for Computing Machinery, 2023, p. 35–44. [35] W. Garey, T. Ropitault, R. Rouil, E. Black, and W. Gao, “O-RAN with machine learning in ns-3,” in Proc. Workshop ns-3. NY, USA: Association for Computing Machinery, 2023, p. 60–68.