Manufacturing has undergone a significant transformation from manual operations to automated, data-driven systems. Traditional manufacturing relies heavily on static models, rule-based control, and reactive maintenance, which limits adaptability and efficiency. The development of AI-enabled digital twin systems in manufacturing follows a logical progression from rigorous design verification and validation models, which include both the integration and co-evolution of products and processes (Tolio et al., 2010). These models are further extended to consider the product lifecycle (Maropoulos, 2010). Then, progressing to advanced sensing is necessary for effective process monitoring with capabilities to enable root cause analysis of quality defects (Ding et al., 2003). This ultimately leads to AI-driven control and autonomy of manufacturing processes (Ding et al., 2019; Gao et al., 2024). Early work on systematic validation frameworks established the conceptual foundation for virtual models of processes and products (Maropoulos, 2010; Tolio et al., 2010), while research in sensing led to (i) development of novel sensors for in-line/in-process monitoring of manufacturing systems (Shi, 2006); (ii) determining placement of in-line monitoring station(s) within a multi-stage manufacturing systems which is then synergistically integrated with measurement coverage analysis for more effective process monitoring, for e.g., reducing mean-time-to-detection by advancing process monitoring from off-line inspection toward in-line and in-process monitoring (Modoni et al., 2022) and system diagnosability (Ding et al., 2003); and (iii) increase monitoring data: variety (for e.g., ranging from categorical to spatio-temporal data) and volume (from small samples to 100% inspection rate) which are collected in-process with sufficient fidelity (i.e., within the required accuracy, repeatability and reproducibility) and veracity (i.e., accurate reflections of the process “defect condition” patterns). These advances provided the real-time input needed for both high fidelity and veracity analysis of the manufacturing process (Babu et al., 2019; Wang et al., 2021).

Subsequent methodologies for root cause analysis (RCA) and process variation modeling have provided a robust foundation for diagnosing product quality issues and quantifying sources of process variability, and as such evolved into critical components of intelligent manufacturing systems (Ceglarek et al., 2004; e Oliveira et al., 2023). RCA methodologies facilitate the systematic identification of latent defect drivers, while variation modeling quantifies stochastic process behavior and parameter sensitivities (Ceglarek et al., 1994; Franciosa et al., 2020). These analytical frameworks, when synergistically integrated with deep learning architectures and statistical reasoning techniques for predictive quality estimation (Sinha et al., 2020) and closed-loop process control, enable the development of high-fidelity digital twins (Soori et al., 2023). Simultaneously, the emergence of smart manufacturing theme supported by “Factory of the Future” and then Industry 4.0/5.0 and the “twin transition” (i.e., integration of digital and green transformations) initiatives (European Commission, 2020; Thelen et al., 2023) has further reinforce the need for real-time process monitoring, predictive analytics, and intelligent decision-making, i.e., enablers that currently drive the development of Digital Twin (DT) Technology (Wang et al., 2021; Villalonga et al., 2021).

Digital twin (DT) systems have emerged as a groundbreaking innovative framework in manufacturing (Leng et al., 2021; Soori et al., 2023), offering a virtual replica of physical processes and systems to enable real-time process monitoring (Franciosa et al., 2020; Ceglarek et al., 2015), optimization (Thelen et al., 2023), process control (Stavropoulos et al., 2021; Stavropoulos and Sabatakakis, 2024), and decision-making (Bocklisch Franziska, 2023). By integrating advanced sensor networks (Javaid et al., 2021), robust data management (Schadt et al., 2010), and artificial intelligence (AI)-driven analytics (Ertel, 2024), digital twins empower manufacturers to simulate, predict, and proactively resolve potential issues. While significant progress has been made in enabling in-line and in-process monitoring, comparatively fewer studies have addressed control-oriented digital twins, and even fewer have systematically examined the robustness of their underlying models when faced with uncertainties such as sensor noise, parameter variability, and environmental disturbances (Stavropoulos et al., 2021). Addressing these challenges remains critical for ensuring reliable and resilient digital twin deployment in industrial environments.

AI-driven approaches are expected to revolutionize digital twin technology by significantly expanding current reliance on traditional simulations by enhancing real-time decision-making, predictive modeling, and process optimization. Unlike traditional simulation-based methods (Mourtzis et al., 2015), which often rely on predefined scenarios and assumptions that limit responsiveness due to time-consuming calculations and dynamic environments (Tao et al., 2024), AI has the potential to learn from vast amounts of historical information, physical simulation and real-time monitoring data to generate accurate predictions, adapt dynamically to changing conditions, and provide context-based insights and ultimately better informed decisions (Nti et al., 2022). AI-driven defect detection (Ren et al., 2022), predictive maintenance (Achouch et al., 2022), and quality assurance (Sinha et al., 2020) can leverage various sources of data, information, and knowledge without relying only on a single resource-intensive simulation, offering manufacturers a more scalable and efficient approach.

In the context of current manufacturing trends (i.e., Industry 4.0 and twin transition) where manufacturing ecosystems are increasingly characterized by cyber-physical integration, interoperability, and data-driven decision-making (Ceglarek, 2014; Herrera Vidal, 2021), the limitations of traditional simulation-based digital twins in addressing scalability, responsiveness, and contextual awareness have become evident. To address these challenges, we propose an AI-enabled Digital Twin (AI-DT) framework that operationalizes key enablers of Industry 4.0/twin transition through the integration of advanced AI modalities: Generative AI for synthetic data augmentation and design-space exploration; Predictive AI for real-time forecasting and anomaly detection; Explainable AI for model interpretability and traceability; Context-Aware AI for environment-sensitive adaptability; and Autonomous Agentic AI for decentralized, self-governing control and decision-making. This composite AI-DT architecture enhances the semantic interoperability, cognitive autonomy, and adaptive intelligence of digital twins, thereby supporting resilient, self-optimizing manufacturing in line with Industry 4.0 and “twin transition” objectives.

The framework provided in Figure 1 outlines a comprehensive and modular architecture for the AI-enabled Digital Twin (AI-DT) system. It integrates physical and virtual systems, ensuring seamless interaction between real-world processes and their digital counterparts. The framework is designed to leverage cutting-edge technologies like sensors, advanced data management, AI and interactive user interfaces.

1. Sensor data acquisition layer (physical system, SDAL) forms the foundation of a digital twin system by accurately mirroring the physical system through real-time acquisition of diverse types of data from various sensors. This layer plays a critical role in monitoring, analyzing and simulating real-world processes by providing high-fidelity, precise, and contextually relevant inputs. The SDAL primarily focuses on data collection and transmission, serving as the entry point for real-world data into the virtual ecosystem. While the sensor hardware (e.g., photodiodes, cameras, and 3D scanners, et al.) is typically pre-selected and fixed based on system requirements, the emphasis lies on how data is efficiently captured and transmitted to meet system demands. This ensures the quality and timeliness of the data flow that underpins the accuracy and effectiveness of the Digital Twin.

2. Data management layer (bridging physical and virtual systems, DAML) acts as the bridge between the physical and virtual domains, responsible for processing, organizing, and managing the data collected by the sensors. Its main objective is to handle the vast amounts of data generated by the SDAL and make it usable for the Multi-Model AI Layer and the User Interface Layer.

3. Multi-model AI layer (virtual system, MMAL) is the core intelligence engine of the AI-DT system. It leverages multiple AI techniques to process different kinds of data, generate insights, predict outcomes, and provide actionable recommendations. This layer operates on data provided by the DAML, supports real-time and proactive decision-making in manufacturing.

Generative AI (GAI) helps simulate scenarios, generate synthetic data, create virtual representations of real-world systems, and is often the starting point in creating a Digital Twin system, it lays the foundation for building the virtual model of the physical system (Gozalo-Brizuela and Garrido-Merchan, 2023; Bandi et al., 2023; Feuerriegel et al., 2024). The main tasks include: (1) Synthetic Data Generation (Lu et al., 2023): it generates realistic training datasets to augment defect detection models, addressing challenges like data scarcity or class imbalances. (2) 3D Model Reconstruction (Liu et al., 2024): utilizes video or sensor data to generate high-resolution 3D models of manufacturing components for inspection and 3D environment building. GAI can effectively simulate manufacturing scenarios to test and validate predictive models without disrupting real-world operations.

Predictive AI (PAI) focuses on forecasting future system states, such as defects, based on historical and real-time sensor data. Its primary role is to analyse and infer probable outcomes, supporting proactive decision-making (e.g., predicting porosity or spatter formation based on welding video). The key tasks include (1) Defect Classification (Dai et al., 2024): categorizes defects into specific types (e.g., spatter, porosity, cracks) based on labeled data and predefined classes. This involves analyzing sensor data and assigning defect categories to improve the efficiency of quality inspection. (2) Defect Detection (Ren et al., 2022): identifies the presence and location of defects in manufacturing components using sensor data, video streams, or 3D models. It employs computer vision and sensor fusion techniques to locate defect areas in real-time. (3) Defect Prediction (Liu et al., 2022): predicts the occurrence and type of defects based on process parameters, sensor data, and historical defect records.

Beyond defect classification, detection, and prediction, PAI methods are also widely applied to non-intrusive or indirect feature measurement. For example, in welding, PAI can infer weld quality metrics without destructive testing, while in additive manufacturing, it can estimate over-deposition or surface geometry deviations from sensor data (Stavropoulos and Sabatakakis, 2024; Bourlesas et al., 2024). This extended role demonstrates the flexibility of PAI in providing quality assurance through indirect, real-time measurements that avoid part removal or offline inspection.

Explainable AI (EAI) is critical for interpreting predictions and decisions made by the system, fostering trust and transparency, it provides insights into the “why” and “how” behind system behaviors, which is important for audits, compliance, and user trust (Hrnjica and Softic, 2020; Chen, 2023; Ahmed et al., 2022). The key tasks include (1) Model Explainability (Dai et al., 2024; Li et al., 2018): it could provide clear explanations of how AI models arrive at their classifications or predictions. (2) Causal Analysis (e Oliveira et al., 2023): Identifies the underlying factors influencing AI decisions by establishing cause-and-effect relationships between input variables and predictions. This helps in understanding why specific defects occur, how process variations impact outcomes, and what corrective actions can mitigate identified issues. (3) Trust Building (Choung et al., 2023): It increases trust in AI systems by offering insights into decision-making processes, especially in high-stakes scenarios.

Context-aware AI (CAI) ensures adaptability and robustness in real-time decision-making, making predictions more relevant and actionable by incorporating environmental and operational variations (Hribernik et al., 2021). The key tasks include: (1) Dynamic Decision-Making (Villalonga et al., 2021): adjusts recommendations based on current process parameters (e.g., speed, material type, temperature) and evolving conditions. (2) Context Integration (Reiman et al., 2021): incorporates external factors such as environmental conditions, operator inputs, or production goals to tailor AI actions. (3) Real-Time Feedback (Pantelidakis et al., 2022): provides adaptive feedback loops for optimizing processes as they occur.

Agentic AI (AAI) represents an autonomous decision-making agent that can act based on insights from modules such as GAI, PAI, EAI, and CAI, shown in Table 1. AAI is designed to close the loop by executing decisions. It empowers the system to act without human intervention, enabling full autonomy in complex environments (Acharya et al., 2025). The key tasks include: (1) Autonomous Operations (Leng et al., 2023): executes routine tasks like defect classification, report generation, or process adjustments without human input. (2) Decision Support (Psarommatis and Kiritsis, 2022): provides ranked recommendations or options for operators to choose from, balancing automation with human oversight. (3) Learning and Adaptation (Li C. et al., 2022): continuously learns from feedback and evolving scenarios to improve its decision-making capabilities. For example, an Agentic AI module in a welding application can operate as a self-correcting welding robot that autonomously adjusts process parameters, such as laser power, welding speed, or focal position, in response to defect signals. Unlike traditional automation systems that follow predefined rules, this self-correcting behavior allows the robot to learn from sensor feedback and adapt dynamically to process variations, ensuring consistent weld quality without human intervention.

4. USer interface layer (virtual system, USIL) forming the virtual interface of the AI-DT System, acts as the primary point of interaction between operators and the decisions (Banaeian Far and Imani Rad, 2022). This layer is designed to ensure seamless communication, real-time feedback, and actionable insights for effective decision-making and process optimization. In addition to its role as the primary interaction layer, the USIL enables bi-directional engagement with both MMAL. For example, operators can initiate tasks such as 3D reconstruction and synthetic data augmentation through the USIL, directly visualizing outputs generated by GAI. Likewise, such as defect classification, detection outputs, and predictive trends (e.g., spatter evolution) from PAI's results, are presented via real-time dashboards, supporting informed decision-making. Operator feedback provided through the USIL can be fed back into GAI and PAI, like by refining simulation parameters, adjusting predictive thresholds, or prioritizing specific defect categories. This continuous feedback loop ensures that AI outputs are not only accessible and interpretable but can also be iteratively improved and adapted within production workflows.

In this paper, we use welding as a case study, present a comprehensive framework for designing and implementing a DT system, our approach integrates data collection, advanced AI to develop and apply for industrial processes. The primary objective of this virtual system is to enable operators to carry out their tasks within an augmented or virtual reality environment. Furthermore, evaluate whether welding products meet the required tolerances based on specified process parameters. This framework not only bridges the gap between theory and practical application but also demonstrates the potential of DT systems to transform manufacturing processes by improving operational efficiency and product quality.

Figure 2a serves as the home page of the AI-DT System for welding, offering an overview of the welding system (Figure 2b), sensors, datasets, defects and AI analysis. Building on this foundational overview, the detailed technical framework (Figure 3a) provides a structured methodology for the design and implementation of the AI-DT system. This framework leverages AI-driven methodologies to enhance modeling, monitoring, and decision-making capabilities, ensuring a seamless interaction between physical and virtual systems. (1) 3D Generative Modeling Environment starts with the creation of a 3D virtual model that accurately replicates the welding environment. Using Generative AI, the model is constructed from multi-view videos, integrating spatial, geometric, and contextual data for an immersive digital representation. (2) AI-Driven Modeling and Learning is predicted using Predictive AI while incorporating physics-informed constraints to ensure realistic process simulations. AI learns from historical welding data, simulation outputs, and real-world observations, enabling precise quality assessment and defect forecasting. (3) Optimization and Explainable Refinement undergoes iterative root cause analysis and performance optimization through Explainable AI. This step ensures that AI decisions, such as defect predictions and recommended process adjustments, are transparent and interpretable, fostering trust in automated decision-making while improving accuracy over time. (4) Real-Time Monitoring and Adaptive Interaction remains dynamic through real-time monitoring, capturing live data from sensors, video feeds, and contextual information. Context-Aware AI adapts the digital twin's behavior in response to changes in environmental conditions, process variations, or external influences, ensuring continuous feedback and actionable insights. (5) Digital Twin System Construction and Validation integrates Agentic AI for autonomous decision-making, allowing it to proactively adjust parameters, detect defects, and implement corrective actions without direct human intervention. This ensures that the digital twin can self-optimize and maintain peak efficiency in dynamic operational environments.

This paper presents a modular and scalable AI-enabled Digital Twin (AI-DT) framework that integrates Generative AI (GAI) and Predictive AI (PAI) to enhance real-time monitoring and defect detection in manufacturing. Although demonstrated within the context of welding, which includes sensor data acquisition, multi-model AI layers, and user interface components. The proposed framework can be adapted to other manufacturing scenarios, such as machining, assembly, and additive manufacturing, by reconfiguring sensor inputs, retraining the AI modules with domain-specific data, and updating process parameters to reflect task-specific operational conditions.

As manufacturing environments are inherently noisy and unpredictable, several limitations remain. Future developments will focus on enhancing the robustness and adaptability of our AI models, incorporate metamodeling approaches and automation workflow (Stavropoulos et al., 2023). Training on diverse datasets that capture noise, occlusion, and variability will strengthen model resilience, while metamodels and workflow databases can accelerate Digital Twin development by enabling modular reuse, systematic workflow design, and alignment with business and policy constraints. Furthermore, we plan to integrate Explainable AI (XAI) for actionable root cause analysis, deploy Context-Aware AI to support real-time parameter optimization, and incorporate Agentic AI to facilitate autonomous decision-making and adaptive control. Collectively, these enhancements will advance the scalability, reliability, and practical applicability of the proposed framework across a wide range of industrial scenarios.