Site-Specific Digital Twin

• Site-specific digital twins are virtual models anchored to individual sites, integrating unique geometry, assets, and local decision logic.
• They combine high-resolution data acquisition with domain-specific simulation and real-time telemetry to accurately mirror operational conditions.
• These twins enable optimized control and decision-making across sectors such as wireless networking, facility management, and industrial process monitoring.

Searching arXiv for the primary and closely related papers on site-specific digital twins to ground the article in current literature. A site-specific digital twin (DT) is a digital twin whose geometry, assets, materials, sensing infrastructure, operating context, and decision logic are instantiated for a particular physical site rather than abstracted into a generic model. In the wireless literature, it is described as a “synchronized, site-specific, cross-layer virtual representation” of a deployment (Mefgouda et al., 10 Jun 2026); in mmWave beamforming, as a “virtual replica of a particular deployment” used to generate synthetic channels (Luo et al., 1 Dec 2025); in geotechnical construction, as a probabilistic belief state updated from site investigation and monitoring data (Cotoarbă et al., 2024). This notion is closely aligned with spatial digital twins that include “precise location and dimensional attributes” (Ali et al., 2023), and with cross-domain views in which the twinning object may be a unit, a system, or a system of systems tied to a concrete location and operating environment (Xiong et al., 2024).

1. Conceptual definition and boundaries

A site-specific DT is not merely a high-resolution 3D model. The recurring pattern across the literature is a virtual representation that is anchored to a particular site and kept relevant by data, models, and operational processes. In 6G heterogeneous wireless networks, the conceptual shift is stated explicitly: network selection moves “from” a one-shot operation on an instantaneous decision matrix “to” a decision process over an evolving DT state $\mathcal{T}_k$ (Mefgouda et al., 10 Jun 2026). In industrial process testing, the DT is tied to the induction heating line at Bharat Forge Kilsta AB, including its five furnace zones, coil layout, pyrometer locations, conveyor speeds, and operating modes (Ma et al., 2023). In campus facility management, the DT is built from the as-is condition of the Price Gilbert Building rather than from an archetypal building template (Siv, 13 Dec 2025).

Site-specificity therefore has at least three persistent dimensions. First, it is geometric: buildings, streets, tunnels, rooms, structural members, antenna placements, and trajectories are modeled as they exist at the site. Second, it is operational: local process modes, service intents, maintenance rules, or control policies are embedded in the DT. Third, it is evidential: the twin is updated, validated, or calibrated using measurements, documents, or observations from that same site. A plausible implication is that “site-specific” refers less to visual realism than to the coupling between local structure, local data, and local decision-making.

The literature also shows that site-specific DTs occur at multiple scales. They may represent a single asset or room, a building or production line, an aerial corridor, or a system of systems such as a city district or infrastructure corridor (Xiong et al., 2024). What remains invariant is that the DT is defined relative to an identifiable physical locus and its constraints.

2. State representation and constituent structure

The internal representation of a site-specific DT is domain-dependent, but the surveyed works share a layered state conception. In the 6G wireless formulation, the DT state at epoch $k$ is

$$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$

where $\mathcal{E}$ is the site-specific 3D environment and materials, $\mathcal{R}$ is the RAT set and configuration, $\mathbf{p}_k$ is the UE position, $\mathcal{A}(k)$ is the candidate RAT set, $\mathbf{H}_k$ is the channel state, $\mathbf{Q}_k$ is the packet-level QoS state, $\mathbf{U}(k)$ and $k$0 encode user preferences and service profile, $k$1 is the LLM-generated requirement profile, and $k$2 is DT memory (Mefgouda et al., 10 Jun 2026). This is a particularly explicit example of a site-specific DT as an evolving state space rather than a static simulator.

Wireless channel twins express the same principle in a reduced form. In site-specific mmWave MIMO, real channels are written as $k$3, while the twin generates

$k$4

where $k$5 is the 3D electromagnetic replica of the deployment and $k$6 preserves hardware characteristics (Luo et al., 1 Dec 2025). The twin is site-specific because $k$7 encodes a particular urban canyon, base-station placement, service area, and array orientation rather than a stochastic channel family.

Probabilistic site-specific DTs generalize this state concept to uncertainty. In geotechnical construction, the digital state is defined as

$k$8

that is, a belief distribution over the physical state conditioned on all observed data and past actions (Cotoarbă et al., 2024). This replaces single-point parameterization with a posterior over soil parameters, consolidation state, and predicted response.

Freshness can itself become part of state quality. In UAV-assisted industrial IoT, the “Age of Digital Twin” is introduced to capture DT staleness. For physical entity $k$9, monitored by a group of IoT devices $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$0, the average AoDT is

$$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$1

with $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$2 and $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$3 the upload delay (Khalaf et al., 22 Apr 2025). Here the site-specific DT is not only spatially grounded but also temporally constrained.

3. Acquisition, synchronization, and fidelity

Site-specific DT construction is typically a fusion pipeline that begins with local geometry and ends with operationally aligned digital artifacts. In wireless aerial corridors, a 3D model of the Howard University campus is built from OpenStreetMap in Blender, annotated with ITU materials, combined with base-station locations from OpenCelliD, and passed to NVIDIA Sionna RT to produce site-specific channel realizations and angle information (Tarafder et al., 3 Feb 2026). In the 5G aerial-corridor variant, the resulting channel twin is assembled as $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$4, with $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$5 the complex gain from base station $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$6, antenna $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$7, to UAV $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$8 (Tarafder et al., 6 Jul 2025). In building facility management, the pipeline is terrestrial laser scanning $$\mathcal{T}_{k} = \left( \mathcal{E}, \mathcal{R}, \mathbf{p}_{k}, \mathcal{A}(k), \mathbf{H}_{k}, \mathbf{Q}_{k}, \mathbf{U}(k), \mathbf{S}(k), \mathbf{q}(k), \mathcal{M}_{k} \right),$$9 FARO Scene and ReCap processing $\mathcal{E}$0 Revit BIM enrichment $\mathcal{E}$1 WireTwin deployment, producing a DT with 509 equipment items and OmniClass-coded spaces and systems (Siv, 13 Dec 2025).

Operational synchronization is treated explicitly in industrial DT testing. For the forging line DT, asynchronous telemetry from OPC-UA tags is assembled into “snapshots” containing power, temperatures, positions, speed, and timestamp; consecutive snapshots are then used to initialize the DT, run it forward, and compare predicted and observed states (Ma et al., 2023). This yields a practical online testing loop for fidelity monitoring rather than a one-time validation exercise.

Fidelity is not a single quantity. In site-specific beam codebook learning, three fidelity factors are separated: 3D geometry fidelity, electromagnetic material fidelity, and ray-tracing fidelity. The main finding is that ray-tracing accuracy is the dominant factor, because limiting reflections to first order drastically reduces NLOS coverage and degrades the learned codebook, whereas geometry and material simplifications cause relatively minor loss (Luo et al., 1 Dec 2025). In channel-precoding twins for a Ballston cellular site, fidelity is improved by a two-stage calibration: first tune ray count and material configuration in the ray tracer, then learn an RT-to-real-CIR mapping with a modified U-Net. The reported NMSE changes from $\mathcal{E}$2 dB for RT-only CIR to $\mathcal{E}$3 dB after AI refinement (Haider et al., 27 Jan 2025).

A compact way to summarize the acquisition logic is to view a site-specific DT as the alignment of four artifacts: local geometry, local semantics, local measurements, and local computation. The exact mix varies by domain, but the requirement that these artifacts refer to the same site is what differentiates site-specific DTs from generic digital models.

4. Decision-making and control over site-specific twins

Once built, site-specific DTs act as decision substrates. In application-aware network selection, the DT does not merely provide context; it is the state over which two network-selection branches operate: MADM–LLM–NS and LLM–NS. The framework introduces history-aware adaptive normalization and DT-memory-driven retrieval-augmented in-context learning to stabilize rankings under candidate-set evolution, and reports reduced rank reversal, fewer unnecessary handovers, and improved service-aware QoS satisfaction relative to representative MADM baselines (Mefgouda et al., 10 Jun 2026).

In mmWave MIMO, the site-specific DT is used as an offline surrogate environment for beam codebook learning. Synthetic channels from the DT feed clustering and DDPG-based beam design, and the learned codebooks outperform a DFT codebook in the high-fidelity target scenario. The same work also shows that learning separate LOS and NLOS codebooks yields clear gains for LOS users and smaller positive gains for NLOS users in the studied layout (Luo et al., 1 Dec 2025).

In next-generation aerial corridors, the DT becomes a training environment for DRL. A channel twin built on Sionna RT and 3GPP antenna models is used to generate channel datasets, while Stage 1 precomputes beam directions and Stage 2 trains a Multi-Head PPO policy for BS association and beam selection. The reported gains are a 44%–121% improvement over DQN and a 249%–807% gain over heuristic schemes in a dense UAV scenario, with inference latency around $\mathcal{E}$4–$\mathcal{E}$5 ms for $\mathcal{E}$6–$\mathcal{E}$7 (Tarafder et al., 3 Feb 2026).

A closely related aerial-corridor formulation uses a two-stage CT-enabled optimization in which dual annealing first selects scan angles maximizing array gain and a Hungarian assignment then solves the BS–beam association problem. In the 100 m scenario, the average throughput per UAV is $\mathcal{E}$8 Mbps for the HF-CT method, compared with $\mathcal{E}$9 Mbps for LF-CT and $\mathcal{R}$0 Mbps for the model-based CT baseline (Tarafder et al., 6 Jul 2025).

Site-specific DTs also support maintenance and health monitoring. In resource-constrained systems, an on-board lightweight DT ranks windows by uncertainty and anomaly score, stores only the top $\mathcal{R}$1 windows permitted by memory, and relies on off-board fleet-aware anomaly detection and $\mathcal{R}$2-SP fine-tuning to update the asset-specific model without learning faults as normal behavior (Montana et al., 2023). This suggests a recurrent pattern: the site-specific twin is not only a mirror of the local asset, but also a gatekeeper for which site data should be preserved, transferred, and used for model evolution.

5. Representative domain instantiations

The variety of site-specific DTs is best seen through concrete instantiations rather than generic definitions.

Domain	Site anchor	Primary DT role
Heterogeneous wireless networking	Abu Dhabi urban region with 3D buildings, streets, RAT layout, and $\mathcal{R}$3 UE positions (Mefgouda et al., 10 Jun 2026)	Evolving DT state for LLM-grounded RAT selection
mmWave MIMO	Times Square-like urban canyon with BS at 20 m height and site-specific UE grid (Luo et al., 1 Dec 2025)	Synthetic channel generation and LOS/NLOS codebook learning
Aerial corridor communications	Howard University campus with real BS locations and circular UAV paths (Tarafder et al., 6 Jul 2025)	Channel-twin CSI for beam and association optimization
UAV-assisted industrial IoT	$\mathcal{R}$4 industrial field with grouped IoT devices and stationary UAVs (Khalaf et al., 22 Apr 2025)	UAV placement and resource allocation under AoDT constraints
Industrial process assurance	Bharat Forge Kilsta AB induction heating line (Ma et al., 2023)	Snapshot-based online testing of DT fidelity
Geotechnical construction	Highway 73 embankment near Stockholm (Cotoarbă et al., 2024)	Bayesian updating of settlement and surcharge decisions
Smart campus FM	Price Gilbert Building at Georgia Tech with 509 equipment items (Siv, 13 Dec 2025)	BIM-, document-, and maintenance-centered facility DT
Site-specific channel precoding	Ballston, Arlington, with real BS and drive-test CIR measurements (Haider et al., 27 Jan 2025)	RT- and AI-calibrated CSI inference for precoding
Site-specific hybrid precoding	Downtown Boston urban macro scenario (Luo et al., 2024)	DT-generated synthetic channels for learned compressive sensing and hybrid precoding

Across these instantiations, the same structural logic reappears. The site is first captured or reconstructed; then domain models are bound to that reconstruction; then site data, documents, or telemetry are linked to the modeled entities; finally, the integrated DT supports decisions such as scheduling, selection, prediction, testing, maintenance, or control. This suggests a spectrum ranging from “channel twins” and “process twins” to broader spatial digital twins, but all remain site-specific because their validity depends on local geometry, local state, and local constraints.

6. Limitations, transferability, and research directions

A common misconception is that greater geometric detail alone guarantees a better site-specific DT. The evidence is more qualified. In mmWave codebook learning, the most critical fidelity factor is ray-tracing accuracy rather than geometry or material detail alone (Luo et al., 1 Dec 2025). In wireless channel inference, additional AI refinement is required after ray tracing to bring synthetic CIR close to measured CIR (Haider et al., 27 Jan 2025). In geotechnical PDTs, uncertainty propagation and Bayesian updating are central precisely because sparse site data do not fully determine the subsurface state (Cotoarbă et al., 2024).

Another misconception is that site-specific DTs transfer easily across sites. Several works say the opposite. In the 6G wireless DT, a different site requires rebuilding $\mathcal{R}$5, rerunning Sionna RT and ns-3, and rebuilding DT memory (Mefgouda et al., 10 Jun 2026). In mmWave beam learning, moving to a different environment requires rebuilding the DT and retraining the codebook (Luo et al., 1 Dec 2025). In aerial corridors, the policy is specific to the Howard University corridor, and cross-site transfer is left as future work (Tarafder et al., 3 Feb 2026).

Scalability remains a major obstacle. Sionna RT, ns-3, and high-fidelity ray tracing are computationally expensive in multi-UE or multi-site settings (Mefgouda et al., 10 Jun 2026). LiDAR acquisition, 3D reconstruction, and manual BIM enrichment are labor-intensive for building-scale DT deployment (Siv, 13 Dec 2025). HP2C-DT addresses part of this challenge by making HPC an active element of the computing continuum; in the reported power-grid experiments, edge-side aggregation reduces communication bandwidth by roughly an order of magnitude, dynamic offloading improves response times by up to $\mathcal{R}$6, and compute-intensive workflows maintain near-ideal strong scaling across a practical resource range (Iraola et al., 12 Jun 2025). A plausible implication is that future site-specific DTs will rely increasingly on explicit edge–cloud–HPC partitioning rather than on a single execution tier.

Safety, calibration, and observability also remain open issues. In online forging DT testing, snapshot assumptions can be violated by power changes between snapshots, and some internal states are not directly observable (Ma et al., 2023). In resource-constrained health-monitoring DTs, update logic must avoid learning anomalies as normal behavior (Montana et al., 2023). In campus FM, simulated IoT data validate workflow feasibility but do not yet validate real-time building behavior (Siv, 13 Dec 2025). These limitations do not negate the site-specific DT concept; they clarify that the twin’s authority depends on synchronization discipline, calibrated models, and the quality of the local evidence pipeline.

Taken together, the literature portrays the site-specific DT as a technically heterogeneous but conceptually coherent construct: a DT whose usefulness arises from being about this deployment, this corridor, this building, this process, or this embankment. Its defining feature is not simply fidelity in isolation, but fidelity organized around a specific site and made actionable through local data, local models, and local decisions.