A unified unsupervised deep learning and stride threat modeling framework for multi-domain cyber-physical security in nuclear digital twins

1 Introduction

1.1 The nuclear digital twin paradigm and background

The transition toward decentralized and data-centric energy systems has led to the integration of Industry 4.0 principles into the nuclear industry. One key technology in this modernization effort is the DT of a plant’s equipment and processes. A DT is a dynamic, high-fidelity representation of a physical object that is updated in real time based on sensor input. It mirrors the physical object in terms of both state and behavior, allows prediction of future states under various operating conditions, and enables optimization of the operation of the physical object (Tao et al., 2019). For example, if there is a micro-vibration increase in the frequency of a coolant pump in a nuclear reactor, the DT will update immediately, calculating the accumulated fatigue and predicting the optimal maintenance schedule for the pump (Fuller et al., 2020).

In current-generation nuclear plants, these cyber-physical systems exist in a three-layered architecture (Humayed et al., 2017):

• Layer one represents the heavy machinery used for generating electricity through the primary process (i.e. steam turbine operation, containment vessel pressurization and boiler dynamics).
• Layer two represents the continuous fluid dynamics responsible for the critical secondary cooling loops and water treatment systems required to regulate temperature.
• Layer three consists of the distributed edge-computing networks, which rely upon Industrial Internet-of-Things (IIoT) sensors to collect data on environmental variables and send that data to other parts of the facility.

Figure 1 illustrates the baseline operational architecture of the multi-domain nuclear digital twin, demonstrating how telemetry flows from individual physical equipment at the facility (e.g., steam turbines, motorized cooling valves, and edge sensors) up through the edge computing layer to a centralized virtual control layer. Building upon this structural foundation.

Figure 2 maps the specific attack vectors across the architecture, highlighting the critical IT/OT convergence points and the localized areas where cyber-physical threats may occur.

As outlined in Table 1, given the complexity of a nuclear facility, security solutions must be able to protect multiple, highly specialized operational layers at once.

TABLE 1. Functional domains and hardware mapping within a nuclear cyber-physical ecosystem.

1.2 The IT/OT vulnerability gap

Until recently, the majority of nuclear infrastructure was built around the principle of the “air gap,” where OT-the hardware and software (such as Programmable Logic Controllers (PLC)) used to monitor and control physical components were completely isolated from IT networks connected to the internet (Tuptuk and Hailes, 2018). However, because creating DTs and implementing remote monitoring systems require merging previously separate IT and OT systems, they also eliminate the isolation provided by the air gap, thus exposing historical, legacy industrial communication protocols to new forms of cyber threats (Liang et al., 2017).

Because current enterprise security methods depend almost exclusively on signature-based intrusion detection systems (IDS), which operate much like digital fingerprint scanners, comparing incoming network traffic to a database of already identified malicious code (Langner, 2011), they cannot detect previously unseen cyber-physical attacks. Signature-independent attacks refer to those that have never been seen before.

One common type of cyber-physical attack is a false-data injection attack (FDIA). An FDIA occurs when an attacker gains unauthorized access to an IT system and modifies digital telemetry going to a PLC without sending any recognizable malware. In addition to modifying digital telemetry, attackers can create false representations of real-world events. For example, an attacker could modify digital pressure readings indicating low-pressure boiler levels. The IDS recognizes that nothing malicious occurred and does not raise any alerts. However, in response to the artificially lowered pressure readings, the PLC increases the boiler pressure (Yu et al., 2025). Because this type of attack presents itself as a series of logically plausible sensor readings related to a physical event (in this case, lowering boiler pressure), most existing security controls are unable to recognize it.

1.3 Research questions

The inability of current systems to understand complex relationships between multi-domain physical phenomena creates a need for behavioral anomaly detection. Therefore, this research aims to address potential issues created by merging nuclear facility IT and OT systems by asking several fundamental questions:

• RQ1: How can unsupervised deep learning architectures effectively capture the normal spatiotemporal dynamics of diverse nuclear subsystems to detect previously unseen anomalies without relying on prior attack signatures?
• RQ2: To what extent does domain-specific data scaling—specifically accommodating both bounded physical sensor data and extreme network traffic variances—improve generalized detection accuracy across distinct cyber-physical environments?
• RQ3: Can high-dimensional deep learning feature attribution be algorithmically mapped to established threat modeling frameworks to provide operators with actionable, human-readable security intelligence?

1.4 Contributions of the research

Current literature frequently treats Industrial Control Systems (ICS) security in silos, developing anomaly detectors that excel in specific physical domains but fail to generalize to network-edge environments (Goh et al., 2017). While others criticize deep-learning approaches for being “black-box” (i.e. providing simple alerts–e.g. “anomalies detected”) without explaining either the mechanical or networking context of the issue. Four main contributions result from this research:

• Unified Cyber-Physical Security Framework: We present an end-to-end framework that integrates domain-aware preprocessing, unsupervised temporal anomaly detection, adaptive thresholding, and STRIDE-based threat attribution into a single workflow designed for heterogeneous nuclear DT environments.
• Domain-Aware Processing of Heterogeneous IT/OT Telemetry: Rather than treating all IT and OT telemetry equally through uniform preprocessing methods, this method utilizes domain-aware normalizing techniques to maintain statistical characteristics associated with bounded physical sensor measurements and highly variable edge-network traffic.
• Integrated Detection and Threat Attribution: This method combines unsupervised anomaly detection with deterministic STRIDE-based feature attribution, allowing for the conversion of unsupervised anomaly scores into interpretable cyber-physical threat categories useful for informing operational decisions.
• Comprehensive Multi-Domain Validation: The proposed solution has demonstrated consistency across three publicly available nuclear-relevant proxy datasets that simulate primary power production processes, secondary cooling processes and edge-IIoT monitoring environments.

2 Background and threat landscape

2.1 Digital twins in nuclear engineering

DT is a developing Cyber Physical System technology that uses a dynamic virtual representation of a physical asset to monitor the physical asset during its operation. The DT creates a continuous synchronization of real-time data collected from the physical asset and a virtual environment. Unlike conventional simulation models, the DT facilitates two-way data exchange between the physical infrastructure and the virtual environment, thus allowing for continuous monitoring, predictive analytics, and optimal operations of the entire system’s lifecycle (Tao et al., 2019; Fuller et al., 2020).

Recent advances in DT technology have increased the use of DTs in the nuclear industry for the purposes of reactor monitoring, predictive maintenance, fault detection, and operational support. Recent studies have shown that incorporating Artificial Intelligence (AI) into DTs improves modeling of complex reactor behaviors, detection of abnormal operating states, and improvement of data-driven maintenance practices while improving operational risk (Wang et al., 2025; Zideh et al., 2024; Arhouni et al., 2026). All of these applications will continue to grow in importance as nuclear facilities begin to utilize highly interconnected Cyber Physical Environments utilizing industrial control systems, Edge Computing, and Industrial Internet of Things (IIoT) technologies (Humayed et al., 2017; Arhouni et al., 2026).

Despite all of the recent advancements in DT technology, Cyber Security continues to be one of the least studied aspects of DT Research. While most of the previous studies focused on either system performance, equipment reliability, or predictive maintenance, only a few have addressed the issue of continuous cyber threat monitoring within DT Architectures. Since OT and IT continue to converge, cyber-attacks against industrial communication networks could cause a direct impact on the integrity of physical processes. Therefore, it becomes clear that cyber-physical security becomes a required element versus an optional extension (Humayed et al., 2017; Zideh et al., 2024; Kheddar et al., 2023).

The previously mentioned limitations create a need for integrated AI-driven security frameworks that are able to concurrently monitor heterogeneous physical process data and industrial network traffic. In addition to detecting anomalies in both physical process data and industrial network traffic, the developed frameworks would also need to provide interpretable threat information so that operators could quickly make decisions regarding their operational status (Wang et al., 2025; Kheddar et al., 2023). These factors form the basis for the unified anomaly detection and STRIDE-based Threat Attribution Framework proposed in this study.

2.2 AI-based anomaly detection in industrial control systems

With the increase in digitalization of ICS, there has been a significant expansion of the cyber-attack surface of critical infrastructures. Intelligent intrusion detection is therefore an essential component of modern cyber-physical security. Unlike traditional IT environments, ICS networks operate under severe real-time constraints. For example, cyber-attacks against ICS could result in damage to equipment, disrupt operational processes, or even pose safety hazards. It is for these reasons that anomaly detection has gained widespread acceptance as a viable method for detecting previously unknown or “signature-independent” attacks that cannot be identified by traditional signature-based IDS (Humayed et al., 2017; Kheddar et al., 2023).

Previous Machine Learning (ML) methods used many different types of supervised classification techniques like Support Vector Machines (SVM), Random Forests (RF), and decision trees. Supervised classification methods did an excellent job with identifying known attacks and classifying them into either benign or malicious. However, all of those previous ML techniques required a significant amount of pre-labeled datasets to be successful. The use of pre-labeled datasets limits the ability to classify new attacks that may occur after initial deployment of the system (Jakotiya et al., 2023; Mahadik et al., 2024). As a result, there is a growing need to explore unsupervised and semi-supervised learning models. Those types of models do not require extensive amounts of labeled data to create a baseline representation of what “normal” looks like.

In addition, Deep Learning models including auto-encoders (AE) and long short-term memory auto encoders (LSTM-AE) have gained popularity for anomaly detection in industrial systems. One reason for this increased interest is due to the fact that both AE’s and LSTM-AEs can generate compacted forms of normal operational behavior using sequential data collected from sensors. Furthermore, both AEs and LSTM-AEs can identify anomalies based on high levels of reconstruction errors that are produced during anomaly detection without needing a large number of examples for each type of attack. One benefit of utilizing LSTM-AE architecture is its ability to learn long-term dependencies within multivariate continuous time-series data found in many industrial processes. (Malhotra et al., 2016; Audibert et al., 2020; Zideh et al., 2024).

More recently, advanced architectures including Convolutional Neural Networks (CNN), Transformer-based models, Graph Neural Networks (GNN), and hybrid Deep Learning frameworks have been presented to further improve anomaly detection performance in increasingly complex industrial environments. These architectures greatly improve feature extraction and temporal modeling; however, they typically require considerable computational power, extensive hyperparameter tuning, or large volumes of training data, limiting their feasibility for implementation in resource-constrained industrial systems (Yang et al., 2025; Shi et al., 2026) (L(y)u et al., 2023).

Despite the many improvements made in anomaly detection methodologies in recent years, some issues remain unaddressed. Many of the previous studies evaluated their methodology on only a single dataset, limiting assessments of architectural reuse across multiple domain applicability. Additionally, although many anomaly detection methodologies reported accurate anomaly detection rates, few provided explanations for the detected anomalies; therefore, reducing their practical utility to industrial operators. Finally, many studies processed homogeneous IT and OT data uniformly regardless of their fundamental differences in statistical properties. Processing heterogeneous data sets uniformly reduces anomaly detection accuracy across hybrid cyber-physical environments (Kheddar et al., 2023; Kim et al., 2022; Wang et al., 2025). These shortcomings demonstrate the necessity for frameworks that integrate domain-aware processing, robust anomaly detection, and explainable threat attribution into a singular architecture.

2.3 AI applications in nuclear systems

Nuclear engineering utilizes AI in various ways, such as reactor monitoring and fault diagnostics, as well as predictive maintenance and safety assessments. This is due to improvements in ML and access to large amounts of time-series data, which allow for the use of data-driven methodologies to complement existing physics-based methodologies used to improve the reliability and operationally efficient performance of nuclear plants (Xue et al., 2025; IAEA, 2022).

Data-driven methodologies utilizing ML have been extensively researched and applied for the purposes of equipment health monitoring and anomaly detection in nuclear facilities. ML methodologies such as CNN’s, recurrent neural networks (RNN’s), and LSTM networks have shown capabilities to learn complex temporal relationships contained in multi-variate reactor data, thus enabling the ability to detect abnormal operating conditions and degradation of equipment at an earlier point than traditional methodologies (Wang et al., 2025). Recently, there has been significant interest in using physics-informed ML (PIML), which incorporates both physical knowledge and data-driven models. PIML can be beneficial when developing models that will operate under safe but possibly changing operating conditions (Zideh et al., 2024).

DT-enabled nuclear systems are another area being researched. DTs utilize AI methodologies to continuously monitor real-time sensor data collected from a facility, providing predictive maintenance recommendations, location information regarding anomalies, and assistance with making operational decisions. A spatio-temporal learning framework was introduced by Wang et al., to enable the localization of faults in DT-based nuclear power systems. They found that incorporating temporal dependencies into device-to-device relationships enhances fault localization accuracy (Wang et al., 2025). A recent review article indicated that utilizing AI with DTs enables significant benefits in terms of improved operational resiliency, reduced maintenance costs, and increased safety for all stages of a nuclear plant’s life cycle (Xue et al., 2025; Arhouni et al., 2025).

While there are many advantages associated with developing DT-enabled systems utilizing AI methodologies, much of the current research has focused on three primary areas: i.e. equipment condition monitoring, fault diagnostics, and operational optimization. In contrast, relatively little attention has been given to cyber-physical security concerns related to nuclear DT systems. Anomaly detection and threat attribution are often treated as separate problems. Therefore, they do not provide sufficient support for operational decision-making activities occurring during security incident response. These issues create a need for unifying AI-driven frameworks that can concurrently identify anomalous activity, and provide interpretable threat attribution across diverse nuclear cyber-physical systems.

Table 2 presents a comparative summary of recent articles about DTs, AI-based anomaly detection, and nuclear AI applications. The information contained within this table forms the foundation for identifying the previously mentioned critical research gaps.

TABLE 2. Comparative summary of recent studies on DTs, AI, and cyber-physical security.

2.4 Critical research gap

The three aforementioned research areas, namely DTs, AI-based anomaly detection, and monitoring of nuclear systems, have developed separately. Although DT research has emphasized the importance of operational optimizations and predictive maintenance, ICS security research has concentrated on anomaly detection through the utilization of ML and deep learning technologies. Similarly, recent applications of AI in nuclear engineering have focused on reactor monitoring, fault diagnostics, and equipment health analysis; however, with significantly less focus placed on integrated cyber-physical security.

Therefore, several open questions persist. For example, few studies consider heterogeneous IT and OT data together since each type of data has distinctly different statistical properties. Secondly, anomaly detection methods are generally assessed separately from actual threat interpretation, resulting in a lack of utility in operational decision-making. Lastly, very few studies explore ICS security within DT-enabled nuclear environments using a unified methodology.

These findings illustrate the necessity for an integrated cyber-physical security methodology that integrates domain-aware data preprocessing with robust anomaly detection and structured threat interpretation for heterogeneous nuclear DT systems. Thus, addressing these challenges motivates the methodological contributions described in detail in the subsequent sections.

3 Proposed unified methodology

The primary challenge in securing a multi-domain nuclear DT is the extreme structural variance of the data. Thermodynamic sensor arrays output bounded, continuous telemetry, whereas edge IoT network monitors output bursty, discrete packet data characterized by massive outliers. To address this, we propose a domain-aware, unsupervised Deep Learning pipeline that ingests heterogeneous raw data, identifies latent physical anomalies, and translates those mathematical anomalies into actionable cybersecurity threat intelligence.

Unlike studies that investigate individual anomaly detection algorithms or standalone threat modeling techniques, the proposed methodology focuses on integrating multiple complementary components into a unified cyber-physical monitoring pipeline. This systems-oriented design enables heterogeneous operational data to be processed, analyzed, and translated into actionable security intelligence within a single framework.

3.1 Framework architecture overview

The proposed framework operates as a continuous, streaming data pipeline designed for edge deployment within a nuclear facility. The architecture is executed sequentially across five primary stages:

• Raw Data Ingestion and Segmentation: Telemetry is gathered from diverse facility subsystems and chunked into contiguous segments to preserve temporal dependencies across file boundaries.
• Domain-Specific Scaling: A dynamic scaling algorithm applies specific mathematical normalizations based on the physical origin of the data (physical sensing vs. network monitoring).
• Sliding Window Sequencing: The scaled continuous data is transformed into overlapping multivariate time-series matrices suitable for recurrent neural network ingestion.
• LSTM-Autoencoder Reconstruction: An unsupervised deep learning model compresses the data into a latent representation and attempts to reconstruct it. High reconstruction error indicates an anomaly.
• STRIDE Feature Mapping: The specific sensor driving the anomaly is isolated, and its functional keywords are evaluated to categorize the anomaly into the STRIDE threat matrix.

3.2 Segment-aware preprocessing and dual-scaling

ICS security typically uses Min-Max scaling techniques for data preprocessing. However, this method of using a single scale across multiple domains is problematic because it distorts the feature space (Jakotiya et al., 2023). For example, a sudden spike in network activity (i.e. Denial-of-Service attacks) would contain large extremes of packet size, therefore squashing the variance of a low amplitude sensor (e.g., a radiation monitor).

Our framework addresses these problems by introducing a domain-aware Dual Scaling preprocessing mechanism. Our framework determines the scaling mechanisms to be used for each domain:

3.2.1 Bounded physical telemetry (HAI, SWaT)

Bounded physical telemetry can include, for instance, sensors measuring temperature, pressure, and rotational speed in primary power generation and secondary cooling loops. Sensor values are limited to bounded thermodynamic parameters. Therefore, we utilize Min MaxScaler (Equation 1) to normalize operational data into the range [0,1].

x s c a l e d = x − x min x max − x min (1)

3.2.2 Edge network telemetry (WUSTL-IIoT)

Standard scaling methods do not provide effective results for network telemetry monitors that have large, unbounded packet floods and bursty transmission patterns. We use a RobustScaler (Equation 2) to process the large outlier-driven bursts. RobustScaler uses the median (Q₂) and inter-quartile range (IQR = Q₃ - Q₁) to make the normalization robust against extreme attack-driven outliers.

x r o b u s t = x − Q 2 Q 3 − Q 1 (2)

In addition to maintaining the temporal continuity needed for recurrent networks, our pipeline also includes Segment Aware Sequencing. No windows are allowed to span files that represent disjoint historical logs or independent test periods. Therefore, the model is precluded from being able to learn artificial “gaps” in time.

3.3 LSTM-autoencoder design

We utilize an autoencoder design consisting of an LSTM layer to capture the complex relationships between the time series data in order to detect attacks without needing labels.

Let X represent the continuous multivariate time-series stream. The data is partitioned into overlapping windows X w ∈ R W × F, where W is the number of time steps, and F is the number of sensor features.

The Encoder utilizes an LSTM architecture to compress the high-dimensional window into a compact, latent representation vector z. The final hidden state of the LSTM layers captures the essential thermodynamic or network baseline of that specific time window (Equation 3):

h t , c t = L S T M e n c x t , h t − 1 , c t − 1

z = R e L U W e n c · h W + b e n c (3)

The Decoder attempts to reconstruct the original input sequence from this compressed latent vector z. The latent vector is repeated across all time steps W and passed through a corresponding LSTM layer, culminating in a linear output layer to generate the reconstructed window X ^ w(Equation 4):

x ^ t = W d e c · L S T M d e c z , h t − 1 ′ , c t − 1 ′ + b d e c (4)

Because the LSTM-AE is trained exclusively on normal operational data, it learns the fundamental physics and baseline network behavior of the DT. When presented with a previously unseen cyber-physical attack, the unseen anomalous pattern forces the encoder into an unfamiliar latent space, causing the decoder to fail. We quantify this failure using the Mean Absolute Error (MAE) for the entire window(Equation 5):

L M A E = 1 W × F ∑ t = 1 W ∑ j = 1 F x ^ t , j − x t , j (5)

If L M A E exceeds a dynamically calculated threshold τ (e.g., the 99.5^th percentile of the training MAE, the window is flagged as anomalous.

3.4 Memory-safe heuristic STRIDE-Oriented Labeling

A fundamental limitation of existing unsupervised deep learning detectors is their lack of contextual output; providing facility operators with a scalar reconstruction error offers no actionable intelligence during a critical infrastructure incident. In order to bridge the gap between quantitative measurement of anomalous behavior and qualitative cyber-physical defense strategies, this research presents a deterministic STRIDE mapping layer.

Once a given sliding window is determined to be anomalous L M A E > τ, the framework replaces traditional binary alerts with attribute-based alerts at the feature level. In complex cyber-physical attacks, multiple features frequently spike simultaneously (e.g., a massive influx of network packets occurs exactly as a physical valve is forced open). To resolve categorization conflicts and prevent ambiguous, overlapping alarms, the framework employs a deterministic maximum-deviation rule. Localized MAE values are computed for every individual feature j in the window W. The algorithm then ranks these features by their absolute reconstruction error and isolates the single feature, f_max, that exhibits the greatest mathematical deviation from its learned baseline. The alert is exclusively mapped based on this primary point of failure (Equation 6).

f max = argmax 1 W ∑ t = 1 W x ^ t , j − x t , j (6)

Through the identification of f_max, the framework identifies which specific actuators, pumps, or network protocols have deviated significantly from their learned temporal baselines. The framework then cross-references nomenclature associated with f_max against a predetermined, domain-specific STRIDE keyword dictionary. This localized mapping performs sequentially for each window, thereby eliminating the need for storage of full feature-error matrices for the entire dataset.

To promote transparency and reproducibility explicit keyword mappings are detailed for primary generation, secondary cooling and edge networks. Table 3 details this dictionary which was constructed by mapping naming conventions native to testbeds HAI, SWaT and WUSTL IIoT to corresponding theoretical attack vectors. This mechanism is deterministic and heuristic. To avoid systematic classification bias in control rooms any feature that does not match a defined keyword strictly is defaulted to ‘Unknown/Unresolved’ and requires manual investigation by operators. Furthermore, the ‘Elevation of Privilege’ category has been deliberately excluded from this dictionary; verifying privilege escalation typically requires OS-level kernel logs and active directory audits, which are not present in standard ICS thermodynamic telemetry streams.

TABLE 3. Domain-Specific STRIDE keyword dictionary used for Heuristic STRIDE-oriented labeling.

3.5 Formal algorithmic implementation

The synthesis of the dual-scale preprocessing phase, the LSTM-AE reconstruction component, the dynamic thresholding component, and the STRIDE attribution component is presented in Algorithm 1.

Algorithm 1 provides an implementation that allows for continual execution of streaming data on edge devices. Through isolation of processing to a single sliding window X_w and through dynamic routing of data through the correct domain scaler, the framework provides timely classification of threats compatible with real-time DT synchronization for nuclear plants. As described in Algorithm 1, by limiting processing only to feature-level attribution whenever L_MAE exceeds the 99.5^th percentile threshold, the framework is guaranteed to be memory-safe.

Algorithm 1 Unsupervised Multi-Domain Anomaly Detection and STRIDE Categorization.

4 Experimental setup and multi-domain datasets

The empirical validation of the proposed framework requires data sets that reflect the extreme heterogeneities present in today’s nuclear cyber-physical systems. Since obtaining destructive cyber-attack data from a live nuclear reactor would be both technically impossible and ethically wrong, the data used in this study was derived from three well-established and widely recognized hardware-in-the-loop (HIL) industrial testbeds. Although these are not native to active nuclear systems, HAI, SWAT, and WUSTL-IIoT were selected to serve as nuclear-relevant industrial proxy datasets. They effectively emulate the operational characteristics, thermodynamic boundaries, and network variability associated with a DT representation of a nuclear ecosystem. It is important to note that this evaluation represents an offline conceptual emulation of a DT monitoring pipeline; it functions as a foundational cyber-physical anomaly detector rather than a live, closed-loop system with bidirectional synchronization. As a whole, these proxy datasets provide a multi-domain approximation of the tripartite nature of a nuclear facility (primary power generation, secondary fluid cooling, and edge-network telemetry).

In order to avoid temporal data leakage as well as to eliminate forward-looking bias, in addition to being split by domain; all datasets (training, validation, testing) were also split based on time. Therefore, randomization of the training/test sets was prohibited to ensure that each testbed’s thermodynamic inertia remained unbroken and to allow for an unsupervised baseline to be completely unaware of future anomalies. The pre-processed dataset information can be found in Table 4. The summary provides dimensionality information for each dataset, Window size for the time series segmentation phase before deep learning ingestion, and imbalance ratios for the classes contained within each dataset.

TABLE 4. Summary of the Multi-domain ICS datasets utilized for framework validation.

Window sizes represent empirical values derived from the time series segmentation phase before deep learning ingestion based upon an operationally defined sequence length of ten-time steps.

4.1 Domain 1: primary power generation (HAI dataset)

Primary steam production is the process in which thermal energy produced via nuclear fission is converted into high-pressure steam to generate electricity through massive turbine generators. A proxy for this domain is provided by the HIL-based Augmented ICS (HAI) dataset (Shin et al., 2020).

HAI is comprised of a complex assembly of physical components consisting of boilers, turbine simulators, and water treatment equipment, along with commercially available Programmable Logic Controllers (PLCs) such as GE, Emerson, and Siemens. Telemetry consisted of 79 continuous physical features representing boiler pressure-related parameters, steam flow rate, and turbine speed (RPM). The data from the original 79 physical input channels is expanded by a factor of four during the automated pre-processing and alignment phase when it is converted into 335 inputs in preparation for ingestion as sequences. This increase is strictly due to the conversion of all discrete categories (i.e., discrete PLC states, binary on/off valve states, and discrete setpoints) to be mathematically compatible with a feed forward neural network through the use of one-hot-encoding. This process aligns the structure across the multiple different files that comprise the entire HAI dataset while at no point creating artificial lags or target leaks. For purposes of this study, HAI served as a proxy for the reactor’s primary steam-generating cycle. An attacker would have to conduct False Data Injection (FDI) attacks at multiple points to create the type of thermodynamic stress described here. Thus, within HAI, this type of attack would be viewed as an attempt to hide increasing core temperatures or cause artificial mechanical failure due to improper function of turbine governor valves.

4.2 Domain 2: secondary cooling loops (SWaT dataset)

Safety in a PWR depends exclusively on its secondary and tertiary fluid dynamics systems. Millions of gallons of water are pumped through heat exchanger banks to cool the primary circuit and prevent thermal runaway. This domain is represented by the Secure Water Treatment (SWaT) dataset created by the iTrust Center for Research in Cyber Security (Goh et al., 2017).

SWaT is a six-stage industrial water treatment test bed with complex fluid dynamics systems utilizing motorized valves, reverse osmosis (RO) membranes, and chemical dosing pumps. Features captured include exact amounts of fluids being pumped through various parts of the process (i.e., exact amounts of fluid level), actual flow rates, and status of pumps operating (i.e. Open/closed/Running). In our DT architecture, SWaT represents the cooling loops necessary to maintain temperature stability throughout the plant. Attackers in this domain directly affect valve operations. For example, an attacker could force a primary coolant pump into “off” while simultaneously transmitting false sensor readings indicating that it remains operational. Such actions could result in localized boiling or steam explosions in a real-world reactor environment.

4.3 Domain 3: edge monitoring (WUSTL-IIoT dataset)

Many modern nuclear facilities are currently deploying IIoT sensors to collect data related to the local environments surrounding their physical assets (e.g., perimeter radiation levels, seismic vibrations, and ambient temperatures). These edge devices communicate with other devices across internal networks, creating additional layers of potential vulnerability to IT-layer attacks that exploit weaknesses in physical access controls. The WUSTL-IIoT dataset is an example of one of these types of attack surfaces (Zolanvari et al., 2021).

While HAI and SWaT produce bound physical outputs (e.g. 0 PSI to 100 PSI), WUSTL-IIoT produces raw IIoT edge network telemetry transformed into 41 discrete features (e.g., source bytes, destination packets, TCP sequence duration) whose properties are quite different from those produced by HAI and SWaT. Specifically, WUSTL-IIoT contains large numbers of outlier values during periods of intense network flooding. Finally, this dataset includes DoS attacks, network reconnaissance attacks, and spoofing attacks. Therefore, when modeling this domain in our DT framework, we model an adversary attempting to disable the control room by cutting off all communication between the field sensors and the central DT controller.

4.4 Implementation and reproducibility details

All experiments used PyTorch version 2.12.0 to ensure complete reproducibility. One LSTM layer with 32 hidden units was used to compress the multivariate time series windows into a lower-dimensional latent space of 16. During the compression step, a Rectified Linear Unit (ReLU) activation function was applied for the latent space mapping. No dropout was applied (dropout = 0.0). To ensure complete transparency of the methodology, the full configuration is detailed in Table 5. Architecture standardizes across all domains and scales automatically with input feature dimension (F). Sequence length was fixed at W = 10 timesteps; this duration was selected empirically because it captures immediate mechanical inertia (such as delay between digital command and valve actuation) effectively without inducing excessive buffering latency that would interfere with real inference at the edge. However, depending on the native telemetry frequency of each domain (Table 1), W = 10 represents substantially varying physical durations—from milliseconds in the primary power generation (HAI) to sub-seconds in fluid dynamics (SWaT), to variable, event-driven network flows in the edge monitoring domain (WUSTL-IIoT). The lack of domain-specific window-length sensitivity tuning is a recognized limitation of this baseline evaluation.

TABLE 5. Comprehensive LSTM-autoencoder architectural and hyperparameter Configuration.

The model’s parameters were optimized using the Adam optimizer. A constant learning rate of 0.001 was selected along with a batch size of 256. Training stopped after a maximum of 50 epochs. Instead of employing early stopping, the best weights for each of the models corresponding to the smallest mean squared error (MSE) during validation were saved through checkpointing (Epoch 18 for HAI, Epoch 42 for SWaT, and Epoch 50 for WUSTL-IIoT). Although MSE was used in training to encourage rapid gradient descent by heavily penalizing the size of individual error terms; The use of MAE (Median Absolute Error) when calculating Anomaly Scores during Inference was done intentionally as it provided an error distribution with no outliers which is mathematically better than MSE for creating Stable Percentile Based Anomaly Boundaries.

To strictly simulate the resource-constrained nature of edge-assisted deployments, the entire multi-domain training and inference pipeline was run on standard CPU hardware restricted to only two CPU core threads. Run times for the three domains were calculated under these constrained conditions.

4.5 Evaluation metrics and threshold sensitivity

Evaluating anomaly detection in safety-critical infrastructure requires specialized statistical scrutiny. Because cyber-attacks are exceedingly rare compared to standard operations, the datasets are severely imbalanced. Consequently, standard “Accuracy” is a misleading metric; a model could simply predict “Normal” 100% of the time and achieve 95% accuracy while completely failing to detect an attack.

Therefore, this research evaluates model performance using Precision, Recall, and the F1-Score, defined as follows:

P r e c i s i o n = T P T P + F P

R e c a l l = T P T P + F N

F 1 − S c o r e = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l

Where TP (True Positives) represents correctly identified attacks, FP (False Positives) represents normal states falsely flagged as attacks, and FN (False Negatives) represents undetected attacks.

We evaluate the ability to distinguish between normal and anomalous data sets by computing both the Receiver Operating Characteristics (ROC) Area Under the Curve (AUC) and the Precision-Recall Area Under the Curve (PR-AUC). In addition, because of the significant imbalance between positive and negative samples in many real-world applications, we emphasize PR-AUC over ROC-AUC (Davis and Goadrich, 2006).

4.5.1 The reactor scram imperative and percentile thresholding

In the Nuclear Power Industry, there is a major financial and operational cost associated with triggering a “reactor scram”. An error-induced false positive can cause an automated safety system to insert control rods into a reactor, causing an instantaneous stoppage of the nuclear fission reaction. After this action, the reactor may take up to 72 h to restart due to Xenon-135 poisoning in the core. This can result in millions of dollars being lost by Utility Operators (Qi et al., 2023). In order to prevent such costly errors, we have developed a mathematical method to establish a well-defined boundary between common fluctuations in process variables and windows exhibiting statistically significant anomalies that warrant immediate security review. This has been illustrated graphically in Figure 3.

To circumvent this, our methodology deliberately suppresses False Positives by implementing a highly restrictive, dynamic anomaly threshold (τ). This threshold is established a priori—completely blind to the test set. Following the unsupervised training phase, τ is calculated exclusively using the 99.5^th percentile of the MAE distribution generated by the normal training windows. By locking this operating point before inference, we ensure zero data leakage. By absorbing 99.5% of standard operational fluctuations and shifts due to benign sensor drift into the baseline, the method ensures that any signal crossing the threshold represents a statistically profound deviation from the established latent representation (Equation 7).

τ = P 99.5 L M A E t r a i n (7)

By absorbing 99.5% of all normal operational fluctuations and shifts due to benign sensor drift within the bounds of what would otherwise be classified as ‘normal’, the method assures that any alert crossing the threshold represents a statistically profound deviation from the established latent representation. Although this extremely restrictive and thus mathematically exacting threshold does produce a reduction in recall by permitting some weak signals to exist beneath its decision boundary, the suppression of false positives remains substantial and independent of the threshold. As such, despite potentially lower values for recall and F1-score at this conservatively operating point, the method can continue to provide very high values for metrics such as ROC-AUC.

5 Results

This section presents the empirical data generated during the validation of the multi-domain LSTM-Autoencoder framework. The evaluation is strictly structured to report the mathematical outcomes of the training phase, the statistical thresholding behavior, the detection efficacy across the withheld test sets, and the output of the automated threat categorization algorithm.

5.1 Unsupervised training convergence and computational efficiency

The framework was trained exclusively on normal operational telemetry to establish a thermodynamic and network baseline. Training was executed over 50 epochs for each of the three operational domains (HAI, SWaT, and WUSTL-IIoT). As depicted in Figure 4, the Mean Squared Error (MSE) loss successfully converged across all three domains without exhibiting severe overfitting.

The convergence trajectories indicate that the dual-scaler preprocessing successfully standardized the learning environment. Despite the extreme structural differences between physical sensor data and network packets, the validation loss stabilized across all domains. Notably, the WUSTL-IIoT training curve (Figure 3, right) exhibits an immediate plateau with a persistent scalar gap between the training and validation splits. This behavior is an expected artifact of modeling edge network telemetry; unlike the continuous Markovian gradients of thermodynamic sensors, discrete network traffic is highly bursty. The LSTM-AE rapidly converges to a central-tendency representation of the normal traffic baseline in the initial epochs. The scalar gap (10¹³) reflects inherent variance disparities between the training and validation splits, amplified by the Mean Squared Error calculation. However, because the framework relies on relative anomaly thresholding rather than perfect time-series forecasting, this rapid baseline convergence seems to be highly effective, ultimately yielding an ROC-AUC of 0.985 during the attack testing phase.

In terms of computational overhead, processing over two million combined sliding windows required highly manageable runtimes. Figure 5 details the computational footprint required to process the time-series sliding windows.

5.2 Threshold sensitivity analysis

To mathematically justify apriori selection of τ as 99.5^th percentile as deployed anomaly boundary, a post hoc sensitivity analysis was performed. The framework evaluated the MAE boundaries generated solely using normal training data over a range of percentiles (90th to 99.9th). The blindly calculated boundaries were then applied to test set in order to observe the mathematical trade-off between Predicted Attack Rate (False Positives) and resulting F1-score.

Figure 6 illustrates the extreme volatility of the F1-score when subjected to lower percentile thresholds. The empirical data demonstrate that conventional thresholds (e.g., the 90th or 95th percentiles) yielded an unacceptably high Predicted Attack Rate, mathematically crushing the Precision metric across all datasets. As the threshold tightened toward the 99.5th percentile, the framework absorbed the standard operational variance (such as minor sensor drift or routine network packet retransmissions). The F1-score achieved its optimal balance at this 99.5^th percentile boundary, successfully suppressing False Positives while isolating statistically profound deviations. This behaviour illustrates the inherent trade-off between maximizing overall ranking performance and selecting a conservative operating threshold suitable for deployment in safety-critical industrial systems.

5.3 Multi-domain anomaly detection performance

Utilizing the optimized 99.5^th percentile threshold, the LSTM-Autoencoder was evaluated against the withheld test datasets containing the actual cyber-physical attacks. Performance is quantified using the Receiver Operating Characteristic Area Under the Curve (ROC-AUC), Precision-Recall Area Under the Curve (PR-AUC), and the F1-score. The comprehensive detection metrics are summarized in Table 6 and Figure 7 quantify the framework’s efficacy across heterogeneous data structures.

TABLE 6. Framework validation performance metrics at the 99.5^th percentile threshold.

To evaluate the correlation dynamics and cross-metric trade-offs between precision and recall across individual domains, Figure 8 provides a normalized metrics heatmap.

5.4 Interpretation of threshold-dependent performance

Although the WUSTL-IIoT dataset achieved an excellent ROC-AUC of 0.9858, its Recall (0.0153) and F1-score (0.0297) appear substantially lower than those observed for the physical process datasets. This apparent discrepancy arises from the distinction between threshold-independent and threshold-dependent evaluation metrics. ROC-AUC evaluates the model’s ability to rank anomalous samples above normal samples across all possible decision thresholds, whereas Precision, Recall, and F1-score are computed using the operational threshold selected for deployment.

When evaluating these metrics, it is vital to contextualize the sharply reduced F1-score for WUSTL-IIoT (from 0.799 at a baseline 95th percentile down to 0.030 at the 99.5th percentile). We explicitly note that the 99.5th-percentile threshold is unsuitable for high-sensitivity network attack detection. While this conservative operating point is highly effective at preventing false-positive physical safety actuations (e.g., reactor scrams), it misses approximately 98.5% of positive network attack windows. Consequently, this specific operating point cannot be characterized as uniformly robust for edge environments. This severe operational limitation fundamentally necessitates the deployment of the multi-tiered alarm hierarchy (detailed in Section 5.8), which relies on a more sensitive 90th-percentile advisory tier to successfully capture network-layer intrusions.

5.5 Statistical robustness and initialization stability

To ensure the reported performance is resilient to the stochastic nature of deep learning weight initialization, the framework was systematically evaluated across five independent training runs using distinct random seeds. Table 7 reports the mean and standard deviation (μ ± σ) for the primary discriminative metrics across these runs.

TABLE 7. Statistical robustness across 5 independent initializations.

The exceptionally tight standard deviations (σ ≤ 0.007) across all three domains demonstrate the consistent convergence stability the preprocessing pipeline and the dynamic thresholding mechanism under random weight initialization. The framework consistently converges to a highly discriminative latent baseline regardless of initial state conditions. However, because overlapping time-series windows are not statistically independent, it is important to note that this seed-to-seed variability strictly measures training optimization stability rather than true operational reliability against temporal drift or novel attack variations.

5.6 Automated STRIDE categorization output

During the testing phase, the framework executed Algorithm 1 to perform feature-level attribution for every window flagged as anomalous (L M A E > τ). Table 8 provides a direct sample of the framework’s algorithmic output, demonstrating the translation from mathematical error to threat taxonomy.

TABLE 8. Top feature-level attributions and their automated mapping to Heuristic STRIDE-Oriented threat categories during the testing phase.

The algorithmic mapping successfully transformed high-dimensional reconstruction errors into specific threat categories. The distribution in Figure 9 confirms that the framework did not default to a single alert type; rather, it dynamically categorized the anomalies based on the specific nomenclature of the sensor or network protocol driving the mathematical deviation. Overall, the testing phase confirmed that the framework detects multi-domain anomalies while successfully providing interpretive, operator-assistance labels. We explicitly note that these outputs function as heuristic guides rather than empirically validated threat classifications, as quantitative attribution accuracy cannot be formally established without ground-truth or expert-annotated STRIDE labels.

5.7 Comparison with state-of-the-art methods

To contextualize the efficacy of the proposed framework, a quantitative baseline comparison was conducted against a representative set of established anomaly-detection methods. The datasets were evaluated against Principal Component Analysis (PCA), Isolation Forest (IF), and a conventional Dense (Feed-Forward) Autoencoder. All baselines utilized identical data splits, windowing logic, and evaluation metrics to ensure acontrolled comparison as detailed in Table 9.

TABLE 9. Qualitative baseline comparison of ROC-AUC performances across representative anomaly detection models.

Bold values denote the best-performing ROC-AUC metrics achieved by the proposed LSTM-AE framework compared to the baseline models across each dataset.

The quantitative results demonstrate that while conventional methods like IF and PCA provide baseline detection capabilities, they fail to capture the complex temporal dependencies of thermodynamic sensors, resulting in lower overall discriminative capacity. The proposed LSTM-AE framework provides superior performance relative to these simpler alternatives, confirming its efficacy before the subsequent application of the multi-tiered thresholding and threat attribution layers.

Furthermore, while advanced architectures like Transformer-AEs and Graph Neural Networks (GNNs) reported in recent literature achieve high detection rates in isolated physical domains, the primary distinguishing capability of the proposed LSTM-AE framework is its integration of deterministic feature mapping. This framework bridges the gap between raw mathematical performance and operational utility, providing the human-readable threat explainability that traditional “black-box” methods generally lack.

5.8 Ablation study of framework components

To isolate and quantify the specific performance contributions of the proposed architectural enhancements, an ablation study was conducted. The framework was systematically stripped of its core innovations—dual-scaling, dynamic thresholding, and heuristic STRIDE-Oriented mapping—to establish a standard LSTM-Autoencoder baseline.

The baseline configuration utilized a universal MinMaxScaler across all domains, a static 95th-percentile anomaly threshold, and yielded only binary alert outputs. Components were re-introduced sequentially. Because ROC-AUC is threshold-independent, both ROC-AUC and F1-Score are reported to demonstrate the isolated impacts on both overall discriminative capacity and the False Positive Rate.

The ablation results in Table 10 mathematically validate the necessity of the unified architecture. The introduction of the Dual-Scaler pipeline is directly responsible for a massive leap in the network-layer (WUSTL-IIoT) ROC-AUC (from 0.812 to 0.985), proving that applying standard thermodynamic scaling to unbounded packet floods destroys latent discriminative variance.

TABLE 10. Ablation study demonstrating the sequential impact of proposed architectural components on detection metrics and operational capability.

Bold values denote the optimal or best-performing metrics (ROC/F1 scores) achieved across the configurations for each respective dataset during the ablation study.

Furthermore, transitioning from a static 95th-percentile threshold to the Dynamic 99.5^th-Percentile Threshold drastically improved the F1-Scores for the physical domains (HAI and SWaT) by suppressing standard operational noise. However, this same transition severely degraded the F1-Score for the edge network domain (WUSTL-IIoT) from 0.799down to 0.030. Reporting only 99.5^th percentile as conservative safety choice is operationally inadequate; nuclear facilities cannot ignore attacks of lower intensity to reduce false alarms. To maximize Recall and Precision in deployment then this study shows that using a framework in a hierarchy of three tiers is essential.

• Tier 1 (90th Percentile) - Advisory Warning: Low level anomalies go unnoticed by operators and are logged silently. Recall is maximized capturing ∼86% of WUSTL IIoT attacks. Events are logged to a SIEM but do not trigger alerts.
• Tier 2 (95th Percentile) - Operator Review: Deviations that are moderate trigger alerts. STRIDE attribution engine displays taxonomy on control room screen and requires operators to verify telemetry before acting.
• Tier 3 (99.5th Percentile) - Emergency Escalation: Only very extreme deviations cross this threshold. Because false positives are mathematically eliminated at this extreme upper limit, automated safety responses are triggered immediately (like shutting down a compromised feedwater pump) without risking reactor scrams.

Integration of STRIDE mapping does not change underlying math probabilities but translates raw anomaly vectors into structured taxonomy needed for this hierarchical incident response. Finally, to explicitly answer RQ2 and demonstrate the necessity of the domain-aware dual-scaling strategy, a targeted scaling ablation was conducted. Table 11 compares the ROC-AUC performance of the LSTM-AE when subjected to no scaling, Standard scaling, Min-Max scaling, and Robust scaling across the three datasets. All comparative runs utilized identical seeds, data splits, and architectural parameters.

TABLE 11. Scaling Ablation demonstrating the impact of standardization techniques on ROC-AUC performance across domains.

The results mathematically justify our domain-specific assignment: Min-Max scaling is required to preserve the strict physical boundaries of the thermodynamic testbeds (HAI and SWaT). Conversely, applying standard physical scaling to unbounded network packet floods destroys the latent discriminative variance; Robust scaling was the only technique capable of standardizing the extreme outlier variance of the WUSTL-IIoT network traffic without crushing the anomaly signal, validating the dual-scaler pipeline.

6 Discussion

This research confirmed the central hypothesis of the study: for a Nuclear DT, which is made up of many domains, one single architecture has to exist to measure/understand physical thermodynamic behavior and network-wide telemetry. A significant part of the contribution of this research is the systematic integration of domain-aware pre-processing, adaptive anomaly detection, and STRIDE-based threat attribution into a single framework. This was done to fill in a large gap in practice often seen in current research studies, where anomaly detection and threat interpretation are developed separately. In this section, we will interpret the findings from Section 5 to answer the basic research questions that guided this research.

6.1 Addressing RQ1: unsupervised deep learning for previously unseen threat detection

Traditional IDS fails in nuclear environments due to catastrophic cyber-physical attacks like turbine shearing or forced thermal runaway, lacking historical signatures. Results show that the stacked LSTM-Autoencoder succeeded in bypassing this limitation with normal, steady-state thermodynamics mapping of the plant by using the high-dimensional physical baseline it created by training exclusively on normal operational data. When subject to simulated previously unseen events (i.e. False Data Injection Attacks masquerading as boiler pressure in the HAI dataset), the adversarial inputs caused the encoder to enter an unknown latent space, leading to extreme reconstruction failure. The model’s profound discriminative power is evidenced by its robust ROC-AUC scores across the primary power generation (0.944), secondary cooling (0.948), and edge monitoring (0.985) domains. Directly comparing these findings with the literature (Table 8), the proposed framework exceeds standard unsupervised baselines like the Variational Autoencoder (VAE) by up to 14.1% in the edge network domain. Furthermore, it outperforms recent Transformer-AE architectures, which demonstrate high efficacy in isolated physical systems but fail to generalize to broader DT environments. This confirms that the proposed architecture provides a mathematically rigorous defense against previously unseen attacks that exceeds current state-of-the-art capabilities.

6.2 Addressing RQ2: the impact of domain-specific dual-scaling

The empirical results demonstrate the dual-scaler architecture. A persistent limitation in existing ICS security literature is the application of homogeneous preprocessing techniques (e.g., universal Min-Max normalization) across heterogeneous IT/OT datasets, which mathematically suppresses the statistical characteristics of physical sensors when combined with highly variable network traffic. The proposed domain-aware strategy addresses this by preserving operational characteristics, allowing the framework to detect subtle physical deviations while remaining robust to extreme network variations (e.g., total_packet_length, 835.5118 MAE).

Compared with existing literature that treats ICS domains in silos, our methodology uniquely generalizes across three distinct domains. State-of-the-art models (i.e., GNNs and Transformers) achieve highly robust detection (>0.91 ROC-AUC) when detecting anomalies within a limited range of physically bounded data; however, their ability to detect anomalies degrades to ∼0.85 when attempting to detect anomalies within raw network variance. The dual-scalar architecture prevented all degradation completely and maintained >0.94 ROC-AUC throughout the entire cyber-physical ecosystem.

More importantly, the preprocessing allowed the successful utilization of the 99.5^th percentile adaptive anomaly threshold. While the WUSTL-IIoT dataset exhibited an impressive ROC-AUC of 0.9858 (strongly indicating threshold-independent discrimination capability), its lower Recall and F1 score reflected its conservatively selected deployment threshold. The chosen operating point purposefully traded off minimizing false positive alarm occurrences vs. maximizing attack recall in order to avoid costly and unnecessary reactor scrams that may occur in a safety-critical environment.

6.3 Addressing RQ3: actionable intelligence through automated STRIDE mapping

One of the largest barriers to implementing deep learning in a nuclear control room is the “black box” output problem; operators cannot execute emergency cooling procedures based upon a scalar reconstruction error. The algorithmic implementation described in this paper closed this gap between mathematical deviation and operational security by instead providing a specific physical target (the fmax) as opposed to merely outputting a binary alert. Rather than providing a list of contributing anomalies, as shown in Table 8, the framework provided immediate STRIDE category assignments for each anomalous physical valve/pump or structure (e.g. Tampering, DoS, etc.) by automatically cross-referencing them against a domain-specific STRIDE keyword matrix.

Unlike other deep learning methods (e.g. Standard Autoencoders, Isolation Forests, Multi-Model Ensembles) that produce only opaque alerts, our method produces deterministic STRIDE attribution. The results demonstrate that unsupervised deep learning models do not have to be black boxes; they can be algorithmically forced to create the structured, readable threat intelligence required for rapid cyber-physical incident response, closing an essential operational gap previously ignored in previous studies.

6.4 Limitations and deployment considerations

While the SWaT data set has a good representation of the complexity of industrial fluid dynamic systems it is missing key variables related to radioactive decay heat and primary coolants’ radioactivity which are present in an operating pressurized water reactor (PWR) system. Likewise, while the HAI testbed has high fidelity of boiler and turbine steam telemetry, actual nuclear SCADA systems have commercially unavailable, radiation hardened hardware whose different levels of latency and noise cannot be represented by available HIL test beds. Therefore, although this validation of the methodological framework was successful, translating this architecture into an operational Nuclear Digital Twin will require calibration to each location utilizing its own native facility telemetry in order to account for such domain dependent physical property representations.

In addition, to clearly define the limits of what this evaluation represents, Table 12 shows how the industrial proxy data sets used in this study compare to their equivalent nuclear data sets. This table illustrates both the similarities of the cyber physical aspects of the two types of systems and the differences in those aspects that represent variables in nuclear power plants that can’t be replicated in commercial test beds.

TABLE 12. Comparison of industrial proxy datasets against active nuclear subsystems.

Finally, several statistical and architectural limitations must be acknowledged. First, the evaluation treats highly overlapping sliding windows as independent samples; future evaluations should employ event-aware block bootstrap procedures and report event-level detection rates to provide a more rigorous measure of operational reliability. Second, the framework utilizes a fixed sequence length (W = 10) across all datasets without conducting exhaustive sensitivity analyses to optimize window lengths for each domain’s unique physical sampling frequency. Lastly, the Heuristic STRIDE-Oriented Labeling module is designed strictly as an interpretive operator-assistance tool; because the datasets lack ground-truth threat taxonomies, the empirical correctness of the assigned STRIDE classifications remains unvalidated.

7 Conclusion and future work

7.1 Summary of contributions and findings

Cyber-physical security frameworks for complex nuclear DTs need to monitor physical processes, ICS, and distributed edge-network infrastructure simultaneously. This study presents a unified deep learning unsupervised cyber-physical security framework for domain-aware preprocessing, temporal anomaly detection, adaptive thresholding, and STRIDE-based threat attribution to heterogeneous IT/OT environments. Unlike proposing a new anomaly detection algorithm, this study’s principal contribution is the systematic integration of complementary techniques to provide an interpretable and practically deployable cyber-physical solution for nuclear DT systems.

The proposed framework successfully integrates normalization across the three representative operational domains using domain-specific normalization as well as a stacked LSTM autoencoder architecture, which accommodates the statistical characteristics of each domain. The dual preprocessing strategy preserves the bounded dynamics of variables in a physical process as well as robustness against highly variable distributions encountered in telemetry data from an industrial network environment. Additionally, the adaptive 99.5^th percentile threshold effectively suppresses false positive alarms while retaining the capability of detecting statistically significant deviations from normal operating conditions. Beyond anomaly detection, the proposed framework addresses one of the primary limitations of many unsupervised deep learning approaches by incorporating deterministic feature-level STRIDE-based threat attribution. As opposed to reporting only reconstruction error values, the proposed framework identifies which operational features contribute most strongly to anomalies and translates the detected events into cyber-physical threat categories that are more easily understandable, thereby supporting faster diagnostic decision-making and incident response. Experimental validation utilizing the HAI, SWaT, and WUSTL-IIOT industrial proxy datasets demonstrated that the unified framework consistently achieved high performance levels across representative primary power generation, secondary cooling, and edge-network monitoring environments with ROC AUC scores of 0.944, 0.948, and 0.985, respectively. These results indicate that the integrated framework provides a practical means for enhancing cyber-physical security within heterogeneous DT environments representative of modern nuclear infrastructure.

7.2 Operational impact and deployment readiness

Alongside detection performance, the design of this framework has been driven by strict operational deployment constraints. To mathematically validate its readiness for edge-assisted monitoring, the complete training and inference pipelines were executed on a resource-constrained, Linux-based environment utilizing PyTorch 2.12.0 allocated to merely two standard CPU core threads. Even under these edge-proxy limitations, the total unsupervised training times were highly efficient: 807.6 min for the 3.4-million-window HAI dataset (converging optimally at epoch 18), 179.9 min for the SWaT dataset, and 39.4 min for the WUSTL-IIoT dataset.

Most critically for real-time deployment, the lightweight, single-layer LSTM architecture requires a minimal memory footprint and guarantees deterministic inference speeds. During testing, the framework achieved an inference throughput of approximately 4,000 sliding windows per second on the 2-core CPU configuration. This yields an isolated forward-pass inference latency of roughly 0.25 milliseconds per window. Because the primary power generation telemetry (e.g., steam flow rates and turbine speeds) operates at high millisecond-level frequencies, this sub-millisecond inference latency definitively proves that the framework can synchronize with a live Nuclear Digital Twin without inducing buffering delays, queue overflows, or violating the strict timing boundaries required for automated incident response.

7.3 Future directions

While the proposed framework provides effective anomaly detection across multiple domains along with interpretable threats, there are several avenues of investigation for future work. A potentially interesting area for exploration would be to integrate PINNs into the framework, allowing physical constraints derived from reactor thermal/hydrodynamic behavior to constrain the models during the optimization process. The incorporation of such additional physics could enhance the physical validity of anomaly detection and decrease false alarms due to normal operating conditions.

Future research may also explore other explainable AI methods, such as attention-based explanation and SHAP-based attribute explanations, to provide complementary insights to the existing deterministic STRIDE mapping method. Additionally, utilizing localized LLMs could allow for automated creation of contextualized incident response recommendations for operators based on identified threat types, enhancing their ability to respond effectively in next-generation nuclear DT environments.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Fei Xu, The University of Texas at El Paso, United States

Reviewed by: Wulfran Fendzi Mbasso, Applied Science Private University, Jordan

Rayhanul Islam Sony, Trine University, United States