TeraVM AI-RSG: Digital Twin & RIC Testing

• TeraVM AI-RSG is a digital twin platform that emulates large-scale Open RAN deployments with realistic RF, mobility, and traffic models.
• It supports closed-loop xApp testing by exposing E2-style metrics and REST API controls for dynamic cell on/off operations.
• The platform enables evaluation of hybrid heuristic-plus-ML strategies, achieving up to 13% energy savings with minimal throughput loss.

TeraVM AI RAN Scenario Generator (AI‑RSG) is an industrial RAN emulation and RIC-testing platform used as a digital twin of the RAN, namely a system-level simulator that reproduces behaviour of an O‑RAN deployment and exposes realistic KPMs and control hooks to external xApps. In published work it is used to develop and validate an energy-saving xApp in a realistic, large-scale emulated Open RAN scenario, with scalable emulation of many cells and UEs, 3GPP-based propagation and channel models, UE mobility and traffic profiles, cell RF and energy models, MAC scheduling and handover logic (Al-Tahmeesschi et al., 12 Sep 2025).

1. Definition and functional role

In the reported Open RAN use case, AI‑RSG serves two roles simultaneously. First, it is a digital twin of the RAN: a software replica that models topology, RF behaviour, protocol behaviour, traffic, mobility, and energy consumption. Second, it is a RIC tester: it exposes O‑RAN E2‑style metrics and control hooks and a REST API, allowing an external xApp to interact with the emulated network as if it were controlling real radio units (Al-Tahmeesschi et al., 12 Sep 2025).

The control pattern is explicit. An external application reads key performance measurements from the emulated RAN and issues control actions back to it, specifically cell on/off commands for MICRO radio units. In the published implementation, this interaction is carried out through HTTP calls to AI‑RSG, while the paper also notes that AI‑RSG can expose KPM reports and RC commands via E2 to external IPs. This is described as functionally equivalent to an xApp running on a near‑RT RIC and controlling real RUs via E2SM‑KPM, E2SM‑RC, and E2SM‑CCC, but all within a safe emulated environment (Al-Tahmeesschi et al., 12 Sep 2025).

This role fits a broader research direction in Open RAN experimentation that emphasizes controlled but realistic environments, automated data collection, and closed-loop xApp testing. OpenRAN Gym, for example, is presented as an open experimental toolbox that provides end-to-end design, data collection, and testing workflows for intelligent control in next generation Open RAN systems (Bonati et al., 2022). The AI‑RSG use case belongs to the same experimental lineage, but with explicit emphasis on digital-twin-driven energy management.

2. Digital-twin architecture and interface model

The digital twin instantiated in AI‑RSG models the main state components required for closed-loop control. These include cells and radio units, UE mobility and traffic, radio/channel conditions, and KPM streams derived from the emulated network state. The xApp described in the literature uses a REST API module, a simulation automation module, a KPM collection module, a logging module, and an action generation module to interact with AI‑RSG (Al-Tahmeesschi et al., 12 Sep 2025).

At the interface level, three API classes are central. Simulation control starts and stops AI‑RSG scenarios through HTTP POST and DELETE. Telemetry access queries KPMs through InfluxDB endpoints over HTTP GET. Actuation sends HTTP POST commands carrying a cell identifier and an on/off action. The same study notes that AI‑RSG can also expose live KPM export and cell control via E2‑style messages, which makes the platform suitable for RIC/xApp testing in addition to digital-twin simulation (Al-Tahmeesschi et al., 12 Sep 2025).

The principal telemetry streams used by the xApp are cell reports, UE serving-cell reports, and UE neighbour-cell reports. Cell reports include PRB utilisation, number of connected UEs, and RF state. UE serving-cell reports include serving cell ID, throughput, and SINR/RSRP. UE neighbour-cell reports include neighbour cell IDs and RSRP. AI‑RSG also reports aggregate outcomes such as total network power consumption, total and average UE throughput, number of UEs in throughput outage, number of heavily loaded MACRO and MICRO cells, number of idle MICRO cells, and number of inactive UEs (Al-Tahmeesschi et al., 12 Sep 2025).

These observables are sufficient for implementing a closed-loop control surface in which the external controller perceives network load, radio quality, and energy state, decides on an action, and immediately tests its consequences inside the twin. This suggests why AI‑RSG is described not only as a simulator but as an experimentation substrate for AI‑assisted RAN control.

3. Scenario and workload modelling

The energy-management study built on AI‑RSG uses a large-scale urban deployment over a $1.2 \times 0.6$ km area, with 10 MACRO cells, 41 MICRO cells, and 246 UEs, over a 2‑hour emulation window (Al-Tahmeesschi et al., 12 Sep 2025).

Parameter	MACRO	MICRO
Number of cells	10	41
Centre frequency	3900 MHz	4050 MHz
Channel model	UMa	UMi
Bandwidth	100 MHz	100 MHz
RF output power	45 dBm	32 dBm
Antenna height	20 m	10 m
Max. power consumption	379 W	172 W
Sleep state power consumption	N/A	8 W
Cell shutdown delay	N/A	10 s
Power reduction rate (shutdown)	N/A	3 dB/s

MACRO cells are always on and provide coverage. MICRO cells provide capacity and are switchable. This asymmetry is central to the platform’s use as an energy-efficiency digital twin: the controller is not asked to redesign topology, but to manage a dense underlay of small cells while preserving service continuity (Al-Tahmeesschi et al., 12 Sep 2025).

UE modelling is heterogeneous. The scenario includes 64 pedestrian UEs at 2 m/s and 20 Mbps target throughput, 50 indoor static UEs at 50 Mbps, 75 fast-car UEs at 15 m/s and 30 Mbps, and 57 slow-car UEs at 10 m/s and 23 Mbps. Average call duration is 30 s for all classes, while mean inter-call time varies by class. Indoor UEs are modeled with realistic heights in the 20–50 m range. These mobility and traffic classes are combined with 3GPP UMa and UMi path-loss models as in TS 38.901 (Al-Tahmeesschi et al., 12 Sep 2025).

The scenario therefore couples three forms of variability: radio heterogeneity, mobility heterogeneity, and traffic heterogeneity. Because AI‑RSG also includes cell RF and energy models, the same scenario can be observed simultaneously through a radio-performance lens and an energy-consumption lens.

4. Closed-loop xApp integration and decision logic

The xApp integrated with AI‑RSG is a hybrid ES‑xApp for MICRO RU sleep-mode management. Its architecture comprises a REST API module, simulation automation, KPM collection, logging, and action generation. The action generation module combines heuristic rules with unsupervised machine learning, specifically K‑Means clustering (Al-Tahmeesschi et al., 12 Sep 2025).

The unsupervised component is used for switch-on decisions when congestion appears. K‑Means is applied to spatial positions of active UEs in overloaded cells and nearby sleeping MICRO cells, minimizing within-cluster variance:

$$\underset{\{\mathcal{C}_k\}_{k=1}^K}{\min} \sum_{k=1}^{K} \sum_{\mathbf{x}_i \in \mathcal{C}_k} \| \mathbf{x}_i - \boldsymbol{\mu}_k \|^2.$$

Within each cluster, the controller computes a weighted distance

$$D_k = \sum_{i=1}^{N_k} w_i \cdot \| \mathbf{x}_i - \mathbf{c}_k \|,$$

where $w_i$ is UE throughput demand, and activates the sleeping cell with

$k^* = \arg\min_k D_k.$

The effect is to select the sleeping MICRO cell that is spatially closest to, and most beneficial for, high-demand UEs (Al-Tahmeesschi et al., 12 Sep 2025).

The heuristic component is used primarily for switch-off decisions. One rule switches off any MICRO cell with zero connected UEs, subject to a protection timer for recently activated cells. A second rule switches off lightly loaded MICRO cells when downlink PRB utilisation is below a threshold $\rho$, set to 50%, and every attached UE has a neighbour cell whose downlink PRB utilisation is also below $\rho$ and whose RSRP exceeds $R_{\min}=-110\ \text{dBm}$. This makes the sleep decision contingent on both load feasibility and handover feasibility (Al-Tahmeesschi et al., 12 Sep 2025).

From an encyclopedia perspective, the notable point is not merely the use of K‑Means, but the interaction pattern between a digital twin and a live controller. AI‑RSG is queried periodically for cell reports, UE serving-cell reports, and UE neighbour-cell reports; the xApp computes a sleep/wake decision; AI‑RSG applies the decision with modeled shutdown delay and RF power ramp-down; the next telemetry round reflects the updated state. This is a complete closed loop.

5. Experimental evaluation in large-scale Open RAN energy management

The published evaluation compares three strategies over 2‑hour AI‑RSG emulations: an All ON baseline, a Heuristic ES‑xApp, and the Proposed (hybrid) ES‑xApp (Al-Tahmeesschi et al., 12 Sep 2025).

Method	Power (kW)	DL Throughput (Gbps)
All ON	4.87	2.47
Heuristic	4.53	2.39
Proposed (hybrid)	4.32	2.46

Relative to the All ON baseline, the heuristic approach achieves a 6.98% power reduction with a 3.32% decrease in downlink throughput. The proposed hybrid xApp achieves a 13.27% power reduction with only a 0.4% throughput reduction. The paper summarizes this as approximately 13% energy savings without compromising user QoS (Al-Tahmeesschi et al., 12 Sep 2025).

The time-series analysis is equally important. The hybrid xApp’s power curve remains consistently below the baseline, while its downlink throughput curve closely tracks the baseline. Reported dips in throughput coincide with periods of reduced traffic demand, which is precisely when the controller is expected to switch off more MICRO cells. The result is therefore not only a static efficiency improvement but a temporally adaptive one (Al-Tahmeesschi et al., 12 Sep 2025).

In the underlying optimization problem, the objective is to maximize the number of MICRO RUs switched off while guaranteeing that every UE remains associated to exactly one RU, receives at least $R_{\min}$, and that no RU exceeds $C_{\max}$ connected UEs. The paper states that this mixed-integer problem is NP-hard, which motivates the low-complexity hybrid heuristic-plus-ML solution (Al-Tahmeesschi et al., 12 Sep 2025).

6. Broader AI‑RAN context, limitations, and extensions

The AI‑RSG use case sits within a broader shift from static RAN experimentation toward AI-native orchestration and digital-twin validation. OpenRAN Gym shows how xApps can be developed and tested on a softwarized RAN with automated data collection, while demonstrations on 7 base stations and 42 users establish that closed-loop xApp environments can be scenario-independent at the control-plane level (Bonati et al., 2022, Bonati et al., 2022). Other practical architectures combine AI-assisted planning, reinforcement-learning-based RAN optimization, real-time telemetry analytics, and digital-twin-based validation, again highlighting the role of the digital twin as a safety layer before live deployment (Saxena et al., 2 Feb 2026).

More recent AI‑and‑RAN orchestration work extends the same principle from radio control to shared compute control. CAORA, for example, couples a monitoring xApp, Y1 interface exposure of radio analytics, and a Soft Actor-Critic orchestrator that dynamically allocates Multi‑Instance GPUs between latency-sensitive RAN functions and AI workloads (Shah et al., 10 Mar 2025, Shah et al., 12 Jul 2025). Agentic AI‑native RAN proposals push this further by adding semantic intent abstraction, LLM-driven coordination, Digital Twin agents, and Configuration, Deployment and Monitoring agents at the SMO layer (Aroua et al., 20 Jun 2026). This suggests that AI‑RSG can be understood as a concrete digital-twin substrate inside a wider AI‑RAN research program: not limited to RU sleep-mode testing, but potentially usable for broader orchestration experiments involving conflict resolution, multi-objective optimization, and joint AI‑and‑RAN workload control.

At the same time, the currently documented AI‑RSG evaluation is deliberately bounded. It uses one urban scenario, UMa and UMi channel models, simple call arrival and duration statistics, and a control surface restricted to cell on/off actions. The evaluation is purely inside the digital twin; real hardware overheads and near‑RT deployment constraints are not measured in that study. The paper explicitly points to extensions such as testing more diverse AI‑RSG scenarios, adding additional control actions such as multi-level sleep states or transmit power adaptation, comparing with reinforcement-learning-based xApps, and integrating the AI‑RSG workflow with an actual near‑RT RIC (Al-Tahmeesschi et al., 12 Sep 2025).

Within those bounds, TeraVM AI‑RSG already occupies a distinct place in the literature: a digital twin that simultaneously supports large-scale O‑RAN emulation, KPI-grounded closed-loop xApp development, and direct measurement of the energy–performance trade-off under realistic radio, mobility, and traffic conditions.