Causal Digital Twin (CDT) Overview

• Causal Digital Twin (CDT) is a digital replica that integrates explicit causal models (e.g., SCM) to answer interventional and counterfactual queries.
• CDT frameworks leverage structured causal graphs, counterfactual reasoning, and simulation engines to perform root-cause analyses and stress tests in telecom and industrial systems.
• Key challenges include ensuring causal fidelity, computational scalability, and robust validation while emerging methods aim to integrate multimodal data and prescriptive analytics.

Searching arXiv for recent and foundational papers on "Causal Digital Twin" and related formulations. Causal Digital Twin (CDT) denotes a digital twin whose internal model is explicitly causal rather than purely correlational. Across the literature, a CDT couples a virtual replica of a physical or cyber-physical process with a causal representation—typically a Structural Causal Model (SCM), a causal graph, or a potential-outcomes simulator—so that the twin can answer not only observational questions such as “what is happening?” but also interventional and counterfactual questions such as “what if we intervene?” and “what would have happened otherwise?”. In telecom, CDT appears as a digital twin explicitly driven by causal knowledge and used to perform interventions and counterfactual stress tests; in industrial control systems it is defined as a twin whose core is an SCM and do-calculus; and in formal counterfactual work it is a stochastic mapping that generates simulated potential outcomes subject to fidelity and validation assumptions (Sriram et al., 9 Jan 2026, Homaei et al., 16 Sep 2025, Laudy, 1 Apr 2026).

1. Conceptual scope and distinction from conventional digital twins

A conventional digital twin is typically a high-fidelity simulation model that mirrors the state and dynamics of a physical system and is used for monitoring, planning, or reactive troubleshooting. In that setting, the twin mainly answers questions such as “If current conditions continue, what will happen?” or “What happened?”, often using physics-based or data-driven models that do not encode causal structure explicitly. A CDT adds an explicit causal representation of system structure, systematic support for interventions, counterfactual scenario generation, and, in several formulations, closed-loop interaction with generative models or decision modules (Sriram et al., 9 Jan 2026, Homaei et al., 16 Sep 2025).

This distinction also appears in digital-twin taxonomies developed for IoT. One line of work separates Display Digital Twin, Forward (Simulation) Digital Twin, and Causal Digital Twin. In that taxonomy, display twins emphasize visibility and monitoring; forward twins emphasize simulation, often for design; and CDT “captures the dynamics of interconnections among assets in terms of cause and effect relationships,” targeting operations, maintenance, prescriptive analytics, and counterfactual experimentation (Madhavan, 2021). Related work on IoT further characterizes CDT as a graph-based causal representation built from multichannel sensor time series, in which the twin is not merely a simulator but a causal map of the system that supports root-cause identification, energy-coupling analysis, and “what-if” experiments (Madhavan, 2022).

A broader representational perspective argues that digital twins should be understood as outcome-driven, dynamic digital replicas that explicitly model task-specific entities and interactions. In that view, “digital twin representations” encode geometry, physical constraints, semantic relations, and domain rules rather than relying on tokenized correlations alone. That paper does not use the term CDT explicitly, but it states that its arguments directly support such a notion by aligning digital twin representations with causal structures, interventions, and counterfactuals (Shen et al., 1 May 2025).

2. Formal models and mathematical formulations

One canonical CDT formulation is the SCM. In the industrial control systems setting, the system is represented as a causal graph

$\mathcal{G} = (\mathcal{V}, \mathcal{E}),$

with observed variables $\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$ and directed edges $\mathcal{E}$ denoting direct causal relationships. Each variable obeys a structural equation

$V_i := f_i(\mathrm{PA}_i, U_i),$

and the full SCM is

$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$

The corresponding causal digital twin function is explicitly defined as

$$\mathcal{T}_{\text{causal}} : (\mathcal{G}, \{f_i\}_{i=1}^n, \text{do}(X = x)) \rightarrow P(Y \mid \text{do}(X = x)),$$

which makes the twin a mechanism for interventional prediction rather than observational prediction alone (Homaei et al., 16 Sep 2025).

In telecom, the causal model is operationalized through a Knowledge Graph $G = (V, E)$ and a Digital Twin Execution Engine. The Knowledge Graph encodes routers, links, services, traffic flows, resource constraints, and their dependencies. The generated scenario is a sequence of interventions

$$\text{Scenario} = (e_1, e_2, \dots, e_T), \quad e_t \sim \text{LLM}(\text{KG context}, e_{<t}),$$

with each event constrained to be a structurally valid $do(\cdot)$ on a node in the graph. The paper explicitly interprets the resulting system as an operational SCM: the KG encodes parent relations $\mathrm{pa}(v)$, and the simulation implements the structural equations $\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$0. The twin then evaluates

$\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$1

where $\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$2 is a multi-step intervention sequence (Sriram et al., 9 Jan 2026).

A distinct formalization embeds CDT directly in the potential-outcomes framework. The Digital Twin Counterfactual Framework defines the twin simulator as a stochastic mapping

$\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$3

with simulated potential outcomes

$\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$4

Using the same latent noise $\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$5 across treatment and control induces a joint distribution over $\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$6, and hence a simulated individual treatment effect

$\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$7

This formalization makes the twin explicitly causal in the sense that it generates simulated potential outcomes, counterfactuals, and treatment-effect distributions (Laudy, 1 Apr 2026).

A third major formulation models CDT through multichannel time series. In the IoT literature, a Structural Vector Autoregressive (SVAR) process is used: $\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$8 or equivalently

$\mathcal{V} = \{V_1, V_2, \ldots, V_n\}$9

where $\mathcal{E}$0 or $\mathcal{E}$1 captures contemporaneous structural causality, and the lag matrices capture generalized Granger causality. The resulting causal factors are visualized as a Fence graph or ladder graph, which is identified directly with the CDT (Madhavan, 2022, Madhavan, 2021). In the stochastic formulation, the causal factors are recast as latent states in a state-space model,

$\mathcal{E}$2

and estimated using Kalman Filter or Kalman Smoother (Madhavan, 2021).

3. Architectural patterns and operational workflows

In telecom, CDT is instantiated as a closed-loop framework called Adversarial Network Imagination. Its three core components are a Causal LLM, a Knowledge Graph, and a Digital Twin Execution Engine. The workflow proceeds as follows: the KG provides a dependency subgraph and context; the Causal LLM generates a multi-step failure scenario that must respect those dependencies; the Adversarial Scenario Generator converts the description into an executable event sequence; the Digital Twin simulates the scenario and records degradation metrics; the Mitigation Engine applies recovery actions such as rerouting and resource reallocation; and feedback on propagation depth, impact, and mitigation effectiveness is used to refine future scenario generation in a closed loop. The “causal” property of the LLM is not autoregressive causality but generation constrained by a dependency graph and by causal consistency conditions such as valid interventions, temporal precedence, and dependency-respecting propagation (Sriram et al., 9 Jan 2026).

In industrial control systems, the CDT framework is organized into five major components: Data Acquisition and Temporal Augmentation, Causal Graph Discovery, Structural Causal Model Estimation, Interventional Digital Twin, and Causal Anomaly Detection, Root Cause, and Counterfactual Reasoning. Causal discovery uses a time-series PC algorithm with lagged conditional-independence tests, while orientation rules enforce temporal precedence, physical constraints, and control logic. SCM estimation uses linear-plus-quadratic additive noise models for continuous variables and logistic regression for binary actuators. Online operation computes SCM predictions, causal scores, the Mechanism-based Causal Anomaly Index (MCAI), and then performs root-cause analysis and counterfactual defense analysis when anomalies are flagged (Homaei et al., 16 Sep 2025).

In multichannel IoT, CDT architecture is centered on the collective treatment of sensor channels as a vector time series rather than as separate univariate streams. The generalized SVAR state-space model treats the causal factors themselves as time-varying latent states, often with a value-and-slope state design, and a time-varying Kalman filter is used for estimation. This representation distinguishes Self-Node Lagged (SNL), Inter-Node Lagged (INL), and Inter-Node Structural (INS) causal factors. The ladder graph then displays lagged edges from $\mathcal{E}$3 to $\mathcal{E}$4 and structural edges within $\mathcal{E}$5, with edge thickness and line style encoding magnitude and sign. In that formulation, the ladder graph is not merely a visualization layer; it is the CDT (Madhavan, 2021).

A complementary architectural direction proposes that foundation models should operate over digital twin representations rather than raw tokens. The proposed pattern has a DT construction layer that maps raw multimodal data into structured representations $\mathcal{E}$6 and an FM reasoning layer that operates over $\mathcal{E}$7: $\mathcal{E}$8 That perspective is not itself a complete CDT architecture, but it supplies a representation-theoretic substrate for CDT systems that need explicit entities, interactions, geometry, and domain rules (Shen et al., 1 May 2025).

4. Intervention semantics, counterfactuals, and reasoning tasks

CDT systems are designed to support the three levels of Pearl’s causal hierarchy. In the ICS setting, these are stated explicitly as Association,

$\mathcal{E}$9

Intervention,

$V_i := f_i(\mathrm{PA}_i, U_i),$0

and Counterfactuals,

$V_i := f_i(\mathrm{PA}_i, U_i),$1

The interventional twin modifies the SCM by cutting incoming edges to the intervened variable and propagating effects through structural equations and adjustment formulas. Counterfactual reasoning follows Pearl’s three-step procedure—abduction, action, prediction—and is used to evaluate defense strategies under hypothetical interventions (Homaei et al., 16 Sep 2025).

In telecom, scenario generation is explicitly framed as interventions on links, routers, traffic, and services. The LLM first produces human-readable failure descriptions; these are then encoded into machine-readable events such as "fiber_link_failure", "router_overload", "traffic_spike", or "multi_node_outage". The Digital Twin evaluates end-to-end latency, packet loss, reroute convergence time, congestion, and impacted nodes or services; the Mitigation Engine then tests actions such as traffic rerouting, resource reallocation, rollback, and failover. The process turns the twin from a passive mirror into a mechanism for counterfactual stress testing and proactive mitigation analysis (Sriram et al., 9 Jan 2026).

In ICS security, anomaly detection is defined in terms of violations of causal mechanisms rather than raw statistical deviations. For each variable $V_i := f_i(\mathrm{PA}_i, U_i),$2 at time $V_i := f_i(\mathrm{PA}_i, U_i),$3, CDT computes a causal anomaly score comparing the measured value to the SCM-based expected value under $V_i := f_i(\mathrm{PA}_i, U_i),$4, normalized by the expected variability under normal causal mechanisms, and aggregates these scores into the Mechanism-based Causal Anomaly Index (MCAI). Root-cause analysis then quantifies $V_i := f_i(\mathrm{PA}_i, U_i),$5 and uses Shapley causal values to distribute anomaly impact among interacting causes, producing top-1 and top-$V_i := f_i(\mathrm{PA}_i, U_i),$6 rankings for operators (Homaei et al., 16 Sep 2025).

The IoT CDT literature emphasizes the same intervention semantics in a time-series setting. A Fence graph or ladder graph can be used for counterfactual experiments by modifying or removing causal paths in the estimated SVAR structure. In the bearing case studies, the CDT is used for “what-if” and counterfactual experiments, with the stated goal of moving from prediction toward prescriptive analytics: not only “what will happen?” but also “what should we change to improve outcomes?” (Madhavan, 2022, Madhavan, 2021).

5. Validation, falsification, and empirical performance

A central issue in CDT is not only how to construct counterfactuals, but how to justify them. The Digital Twin Counterfactual Framework introduces a fidelity hierarchy that separates marginal fidelity, joint twin fidelity, and structural fidelity. Marginal fidelity concerns the correctness of arm-specific conditional outcome distributions and supports quantities such as ATE, ATT, ATU, CATE, and QTE. Joint fidelity additionally requires the correct copula between potential outcomes and is therefore necessary for the ITE distribution, probability of benefit or harm, and variance of treatment effects. Structural fidelity is required for mediation and dynamic-regime effects. The same work proposes a five-level validation architecture: Level 0 – Marginal calibration, Level 1 – Conditional calibration, Level 2 – Individual-level calibration, Level 3 – Treatment effect calibration, and Level 4 – Distributional stress testing. The explicit point is that marginal causal claims can become increasingly testable, whereas copula-dependent claims remain assumption-indexed (Laudy, 1 Apr 2026).

A complementary perspective is causal falsification. In that framework, a digital twin is interventionally correct if, for all relevant action sequences,

$V_i := f_i(\mathrm{PA}_i, U_i),$7

The paper proves that observational data cannot, in general, certify such interventional correctness without strong assumptions such as no unmeasured confounding. Instead, it proposes a longitudinal Manski-bound procedure that tests falsifiable implications of correctness. Rejecting those hypotheses yields reliable evidence that the twin is not correct in the relevant causal sense, even when the data are confounded (Cornish et al., 2023).

Empirical CDT results are most developed in the ICS setting, where both structural validation and operational performance are reported. The following figures summarize the reported quantitative results.

Aspect	Reported result	Dataset / context
Physical Constraint Compliance	$V_i := f_i(\mathrm{PA}_i, U_i),$8	SWaT
Average normalized SHD	$V_i := f_i(\mathrm{PA}_i, U_i),$9	Synthetic ground truth
F1-score	$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$0	SWaT
F1-score	$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$1	WADI
F1-score	$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$2	HAI
False-positive reduction	$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$3	Relative to correlation-based methods
Root-cause Top-1 accuracy	$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$4 vs $$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$5	CDT vs best baseline
Counterfactual defense effect	attack success reduced by $$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$6	Enhanced sensor monitoring
Online anomaly-detection latency	$$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$7	Real-time deployment

These results are accompanied by $$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$8 after Bonferroni correction over 21 comparisons, real-time feasibility up to approximately 150 variables, and intervention-prediction accuracy of 91–96% for observed interventions (Homaei et al., 16 Sep 2025). In telecom, evaluation is qualitative rather than fully numerical, but the reported findings are that the causal framework generates more multi-component and cascading scenarios than rule-based and historical-replay baselines, that causal conditioning leads to deeper cascades in the twin, and that removing the KG or causal conditioning reduces mitigation effectiveness (Sriram et al., 9 Jan 2026).

6. Application domains, limitations, and emerging directions

CDT has already been instantiated in multiple domains. Telecom CDT uses causal knowledge graphs, LLM-constrained scenario generation, and digital twin execution for adversarial failure imagination and proactive resilience analysis (Sriram et al., 9 Jan 2026). ICS CDT uses SCM-based anomaly detection, root-cause analysis, and counterfactual defense evaluation (Homaei et al., 16 Sep 2025). IoT CDT uses SVAR, Kalman estimation, and Fence or ladder graphs to analyze connected assets in industrial, manufacturing, automotive, consumer, building, and smart-city settings (Madhavan, 2022, Madhavan, 2021). The same design pattern is explicitly proposed for power grids, manufacturing, and transportation, where a KG or causal graph encodes dependencies, a twin operationalizes those dependencies by simulation, and interventions are applied in simulation rather than in the physical system (Sriram et al., 9 Jan 2026). Related discussions extend the CDT perspective to smart grids, autonomous vehicles, healthcare cyber-physical systems, and manufacturing sabotage scenarios (Homaei et al., 16 Sep 2025).

The main technical limitations recur across formulations. Digital twin fidelity depends on accurate topology, resource, or process data; incomplete or outdated information reduces causal fidelity (Sriram et al., 9 Jan 2026). Scalability remains a central issue: ICS causal discovery scales as $$\mathcal{M} = \langle \mathcal{V}, \mathcal{U}, \mathcal{F}, P(\mathcal{U}) \rangle.$$9 and is reported as practically feasible for medium-scale systems of up to about 150 variables, while large cascading-failure simulations in telecom are computationally costly (Homaei et al., 16 Sep 2025, Sriram et al., 9 Jan 2026). Stable causal structure assumptions may fail in settings with frequent equipment changes or control-logic modifications, and stealthy attacks remain harder to detect than single-point attacks, with reported detection rates of approximately $$\mathcal{T}_{\text{causal}} : (\mathcal{G}, \{f_i\}_{i=1}^n, \text{do}(X = x)) \rightarrow P(Y \mid \text{do}(X = x)),$$0 and $$\mathcal{T}_{\text{causal}} : (\mathcal{G}, \{f_i\}_{i=1}^n, \text{do}(X = x)) \rightarrow P(Y \mid \text{do}(X = x)),$$1, respectively (Homaei et al., 16 Sep 2025). Data access, scenario quality, and causal identifiability are persistent constraints, and formal counterfactual work stresses that copula-dependent individual-level claims remain fundamentally untestable from standard observational data alone (Laudy, 1 Apr 2026).

Several emerging directions broaden the CDT agenda. The representation-oriented argument that foundation models should operate on digital twin representations rather than raw tokens implies a route toward causally grounded multimodal reasoning, because the representations explicitly encode entities, interactions, constraints, and state variables rather than only statistical co-occurrence (Shen et al., 1 May 2025). In disease modeling, Digital Twin Generators are presented as conditional generative models of longitudinal clinical trajectories; the same work states that these models are not explicitly causal, but it also shows that adding explicit interventions, treatment variables, and causal structure would be the natural path toward a CDT for medicine (Alam et al., 2024). In cardiac electrophysiology, a patient-specific whole-heart and torso model is used to study how interventions on anatomy, conductivities, and electrode placement affect the ECG; the study concludes that diagnostically relevant ECG features and overall morphology appear relatively robust against the investigated uncertainties, which suggests a mechanistic route to causal digital twins in biophysical settings (Zappon et al., 2024).

Taken together, these strands converge on a common understanding. A CDT is a digital twin in which the internal state, dynamics, and intervention semantics are organized around cause-and-effect mechanisms rather than around prediction alone. Its practical value lies in converting digital twinning from reactive mirroring into intervention-aware analysis, counterfactual simulation, and prescriptive decision support. Its methodological difficulty lies in simultaneously achieving causal fidelity, computational tractability, and defensible validation.