Digital Twins in 3D Printing Processes Using Artificial Intelligence

Abstract

Digital twins (DTs) provide accurate, data-driven, real-time modeling to create a digital representation of the physical world. The integration of new technologies, such as virtual/mixed reality, artificial intelligence, and DTs, enables modeling and research into ways to achieve better sustainability, greater efficiency, and improved safety in Industry 4.0/5.0 technologies. This paper discusses concepts, limitations, future trends, and potential research directions to provide the infrastructure and underlying intelligence for large-scale semi-automated DT building environments. Grouping these technologies along these lines allows for a better consideration of their individual risk factors and use of available data, resulting in an approach to generate holistic virtual representations (DTs) to facilitate predictive analyses in industrial practices. Artificial intelligence-based DTs are becoming a new tool for monitoring, simulating, and optimizing systems, and the widespread implementation and mastery of this technology will lead to significant improvements in performance, reliability, and profitability. Despite advances, the aforementioned technology still requires research, improvement, and investment. This article’s contribution is a concept that, if adopted instead of the traditional approach, can become standard practice rather than an advanced operation and can accelerate this development.

1. Introduction to Digital Twins and Support of DT by AI/ML

Properly implemented data management, analysis, and visualization, especially AI-based classification and prediction, helps all types of organizations and people achieve greater success in the implementation of most tasks. This is due to the fact that, although processes are becoming more and more complicated, we are able to measure them more effectively and precisely and describe them as objectively as possible. Of course, large amounts of data to be processed (including in real time) and the optimization of inferences and predictions from data, so that their use is tailored to the needs of users, have become a barrier. There are two solutions here: artificial intelligence (AI), including data-based approaches (machine learning—ML), and digital twins (DTs)—models reflecting real objects accurately enough to predict their correct and incorrect operation. DTs provide accurate, data-driven modeling in real time to create a digital representation of the physical world. They are part of an ecosystem of integrated new technologies, such as virtual reality (VR), mixed reality (AR), and AI/ML, which enable modeling and research into ways to achieve better sustainability, greater efficiency, and improved safety in Industry 4.0 and Industry 5.0 technologies. The aforementioned group of technologies allows for a better consideration of individual risk factors and the use of available data towards putting humans and the environment at the center of events and efforts. This results in an approach to generating holistic virtual representations such as DTs to facilitate predictive analyses in industrial practice [1,2,3].

The evolution of digital twins from basic computer simulations to sophisticated, real-time virtual representations powered by AI has been driven by advances in computing, data analysis, and connectivity. Although the term “digital twin” has been popularized relatively recently, its basic concepts have a history spanning several decades.

• Early foundations (1960s and 1980s) encompassing the concept of using computers to simulate physical systems, along with the development of CAD (computer-aided design) systems to create digital models of physical objects and the creation of physical and digital replicas of spacecraft on Earth for the Apollo missions, which allowed for monitoring and troubleshooting problems during spaceflight;
• The emergence of the digital twin concept (1990s and 2000s) for product lifecycle management (PLM), which required the integration of data and processes for design, manufacturing, and maintenance (in 2002, the term “digital twin” was first used by Dr. Michael Grieves at the University of Michigan to denote a virtual representation that serves as a real-time digital equivalent of a physical object or process);
• Technological growth and advancement (2010s) due to the proliferation of the Internet of Things (IoT), embedding sensors in physical objects, real-time data collection and communication with their digital counterparts, and the integration of artificial intelligence (AI) with DTs, which has enabled the practical implementation of digital twins in many areas, from manufacturing to healthcare and smart cities;
• Modern DT applications (since the 2020s) including complex dynamic DTs based on AI, edge computing, and augmented reality (AR), developed under the auspices of organizations such as the Digital Twin Consortium.

AI-based digital twin architectures include a virtual model that mirrors a physical entity, enabling real-time monitoring, analysis, and optimization. This architecture typically includes several layers, including data acquisition, processing, and AI-based analysis, that work together to create a dynamic representation of the physical counterpart. Data from sensors on the physical object are continuously streamed to the digital twin, where they are processed and analyzed using machine learning algorithms. These algorithms predict future states, identify potential problems, and suggest optimizations based on the twin’s understanding of the physical system. The digital twin is semantically linked to the physical counterpart via a common data ontology that provides a consistent interpretation of the data and processes. This semantic layer ensures that changes to the physical entity are accurately reflected in the digital twin and that the twin’s insights can be efficiently applied back to the physical system. The architecture supports bidirectional communication, allowing the digital twin to send feedback or control signals to the physical entity for autonomous or semi-autonomous adjustments. The AI component continuously refines the twin model by learning from new data, thus increasing its accuracy and predictive capabilities over time. Functionally, the digital twin can be used for simulation, diagnostics, and predictive maintenance, providing a comprehensive understanding of the behavior of the physical system under different conditions. The relationship between the digital and physical counterparts is symbiotic; the digital twin improves the operation and performance of the physical entity, while the physical entity provides live data that improve the twin model. This architecture is essential in environments where real-time decision making and system optimization are essential, such as in manufacturing, smart cities, and healthcare systems.

State-of-the-art DT component solutions include advanced sensor networks for precise data acquisition, high-performance real-time data processing, and AI-based analytics for predictive insights. Sensor technologies such as IoT devices have improved data accuracy and granularity, but they still face challenges in harsh environments and in ensuring consistent data quality. Edge computing and cloud infrastructure enable real-time processing, but data latency and synchronization remain an issue, especially in distributed systems. AI models, especially deep learning and reinforcement learning, have revolutionized predictive maintenance and anomaly detection, outperforming traditional differential equations by handling complex, nonlinear relationships without explicit modeling. AI greatly enhances the predictive capabilities of digital twins by learning from large, diverse datasets, enabling the twin to predict failures or optimize performance under different conditions. However, AI models require significant amounts of data to train and can suffer from interpretability issues, making it difficult to understand the reasoning behind their predictions. Technologies such as federated learning are helping to address data privacy and security concerns by allowing AI models to be trained on distributed datasets without centralizing sensitive information. The application of digital twins to complex systems, such as smart cities or advanced manufacturing, benefits greatly from AI, where traditional differential equations would be impractical due to the immense complexity and variability of the systems. Despite these advances, challenges remain in creating fully autonomous digital twins, as integrating AI models with physical systems still requires human oversight to manage edge cases and unforeseen scenarios. AI-powered digital twins are gradually moving from reactive to proactive and even prescriptive systems, offering recommendations based on future predictions that traditional methods would struggle to provide due to their limited scope in handling dynamic, real-world conditions.

The applications of DTs are wide-ranging and continue to cover new areas, from investigating the deformation of objects such as bridges [4] to emulating cyber attacks [5] and emulating museum objects [6]. DTs are also often designed to monitor production processes or their components [7,8] and the lifecycle of products/services [9]. For the time being, the unattainable dream is to generate a DT of a patient or at least their organs (physiological—healthy and pathological—with dysfunction) [10]. The aforementioned solutions, based on their specific requirements and data collection technologies, are often prototypical and require dedicated solutions, but nevertheless, it is already possible to identify key mechanisms and developments that can be clothed in a technological basis/framework for the creation and use of DTs. These processes are being reinforced by the use of AI, including ML (a data-driven approach in which the links between input and output data are extracted automatically in the learning process so that there is no need to know the rules of operation). AI/ML facilitates data collection and selection and allows for better preparation of data for computational analyses and a higher level of accuracy of inference, classification, and prediction, which is crucial for the development of DTs [11,12,13].

This paper aims to discuss the concepts, limitations, challenges, future trends, and potential research directions to provide infrastructure and underlying intelligence for large-scale semi-autonomous built environments such as DTs.

Introducing our AI into DTs can increase production quality. This applies to both production planning and configuration, as well as quality control and traceability. As process simulation is quite well known and understood as a technology, it has already become a key step in the preparation, design, and monitoring of production processes. The introduction of AI/ML into the simulation of the above-mentioned processes will reduce their costs and will become less time-consuming, including the possibility of carrying out these processes within the production cycle and becoming useful in controlling processes in real time [14]. As a result, AI/ML-based DTs have the potential to transform existing simulation processes into real-time, closed-loop production systems. The costs of such a transformation are acceptable enough that they can be realistically implemented in industrial conditions.

Desheng et al. [8] showed, using DT technology, that the properties and rolling stability in hot rolling are strongly dependent on the final rolling temperature, where diagnosing anomalies in industrial conditions has, so far, been difficult. The proposed method includes a number of elements subjected to computational processing, listed as follows:

• The elimination of samples with large deviations using the Hausdorff distance algorithm;
• Assessment of the tape head setting value based on a random forest model;
• Assessment of other rolling parameters based on the isolation forest algorithm;
• Curve risk assessment performed using the LCSS algorithm;
• A method for identifying the causes of anomalies, combining the knowledge graph and the Bayesian network (the occurrence of nodes in the Bayesian network determines the cause of the anomaly);
• Model verification “on site”.

The above combination of mechanistic modeling and AI techniques allows for fast, automatic, and accurate detection and analysis of final rolling temperature anomalies, with an accuracy of 92% [9].

Abio et al. [15] investigated the use of a substitute model in their analysis of hardening in the stamping process of a 22MnB5 steel sheet. The key stage of heat treatment is directly related to the final quality of products—among others, in the automotive industry. Finite element simulations were proposed for unsteady heat transfer analysis using ABAQUS software Version 6.9, which generates training data for an ML-based DT, predicting key process outcomes for the production of entire batches of products (Figure 1). The DT predicted changes in the most important process temperature variables in many scenarios, with an average error of about 3 °C and a reaction time that was four orders of magnitude shorter compared to traditional simulations. The described technology can be expanded to a larger DT for autonomous process control [14]. The data were prepared by ABACUS for the DT, which included ML; this concerned the production of knowledge, not the data themselves.

Understanding the future evolution of engineering (including its leading areas) is closely linked to the integration and development of AI, particularly in the development of DTs. As systems become increasingly complex and interconnected, AI will play a key role in creating accurate, real-time digital representations of these systems, enabling engineers to simulate, predict, and optimize performance. The evolution of engineering is likely to focus on areas such as smart manufacturing, autonomous systems, and sustainable design, all of which can benefit from AI-powered digital twins. These digital twins will enable engineers to remotely monitor systems, predict failures before they occur, and optimize operations for efficiency and sustainability. Additionally, AI can facilitate the analysis of the vast amounts of data generated by these digital twins, uncovering insights that may not be immediately apparent through traditional engineering methods. As engineering moves toward more predictive and adaptive designs, AI will play a key role in developing and refining digital twins that can evolve alongside their physical counterparts. This symbiotic relationship will also foster innovation in new materials, energy systems, and robotics, where AI-based simulations can test and validate concepts before physical prototypes are built. Additionally, collaboration between AI and engineering via digital twins will increase the ability to create more personalized and adaptive systems that are tailored to specific needs and conditions. Ultimately, integrating AI with digital twin technology will redefine the boundaries of engineering, leading to more efficient, sustainable, and intelligent systems.

The general concept of using AI/ML in digital twins in engineering involves creating a virtual replica of a physical system or component that can simulate its behavior under different conditions. AI and machine learning algorithms analyze real-time data from the physical system, learning from its performance to predict future behavior, identify potential failures, and optimize operations. This approach allows engineers to conduct virtual testing, reducing the need for physical prototypes and accelerating the design process. By continuously updating the digital twin with new data, the model becomes more accurate over time, improving its predictive capabilities. Ultimately, this integration of AI/ML improves decision making, reduces maintenance costs, and increases the reliability of systems. In practice, this means that AI/ML in DTs in engineering is used to predict and prevent equipment failures by analyzing sensor data from machines, enabling predictive maintenance. In automotive engineering, DTs simulate the behavior of engines or transmissions under different driving conditions, optimizing performance and fuel economy. AI-powered DTs in aerospace engineering help design more efficient aircraft components by simulating aerodynamic properties and stress factors under different conditions. In manufacturing, DTs use AI/ML to optimize production lines by predicting bottlenecks and suggesting improvements in real time. Finally, AI-enhanced digital twins are used in robotics, where they simulate robot movements and interactions, increasing precision and reducing the risk of operational errors [28].

2. Our Concept of DTs and the Results of Our Studies

AI and ML in DTs have become popular in recent years, with DTs integrating an AI component at the interface of these two fields [14].

Our concept of an AI-based DT is based on the creation of a highly detailed and dynamic virtual model of a physical system that is constantly updated with data from a physical counterpart in real time. The DT uses AI/ML technologies to simulate, predict, and optimize the performance of a system throughout its lifecycle. This gives it an advantage over traditional solutions that are still used [15,16]. The key components of our DT include the following:

• The physical entity, i.e., the actual reflected system or component (machine, engine, or production line) [17];
• A digital twin model, i.e., a precise digital (virtual) representation of a physical entity, including all its physical properties and behaviors. This model is built using computer-aided design (CAD) data, finite element analysis (FEA) models, and other engineering simulations [17];
• A data acquisition process is carried out in the following way: IIoT sensors and devices collect real-time data from a physical unit (temperature, pressure, vibration, speed, and other relevant parameters) [18];
• Data integration and processing: real-time data are sent to the digital twin either in the form of full data or (in selected cases) in the form of vectors or matrices of features extracted within the EC.AI/ML algorithms process these data to dynamically update the digital model, ensuring an accurate reflection of the current state of the physical entity, usually in real time [18];
• AI algorithms analyze data to detect patterns, predict future behavior, and identify potential problems. ML models can also be trained partially on historical data to improve prediction accuracy and optimize performance [19];
• Simulation and analysis: the DT can run simulations to test various scenarios and predict results without affecting the physical system. This helps one understand how changes in one part of the system can affect the entire or practice scenarios that are difficult to implement in a real system (damage, cyber attacks, etc.) [17,18,19].

Based on the literature review and our own research, we can conclude that the priority for applications of AI-based DTs, relatively fast in development and industrial application, is given to the following:

• Predictive maintenance: AI-based DTs can predict when a particular component is likely to fail, allowing for timely (preemptive) maintenance before an actual failure occurs [20].
• Performance optimization: through continuous data analysis, digital twins can identify inefficiencies and suggest optimizations, such as recommending adjustments for operating conditions to improve performance or extend component life; the final decision here belongs to a person (e.g., a technologist or head of maintenance) [21].
• Design and testing: DTs can be used to test new designs in a virtual environment before their physical implementation; this allows for rapid prototyping and the optimization of designs with reduced costs and time [22].
• Fault diagnosis and troubleshooting: DTs can help diagnose the problem by comparing real-time data with expected behavior. This helps to quickly identify and eliminate faults [23].
• Lifecycle management: digital twins can monitor and manage the entire lifecycle of a system, from design and production to operation and decommissioning, ensuring optimal performance and resource utilization [24].

This approach to DTs allows for the following direct benefits to be achieved:

• Enhanced decision making: provides actionable insights and data-driven decision support [25];
• Cost savings: reduces maintenance costs, prevents downtime, and minimizes the need to create physical prototypes [26];
• Improved reliability and performance: provides continuous monitoring and optimization increase in system reliability and operational efficiency [27];
• Greater flexibility: the use of AI/ML enables testing and implementation of changes in a virtual environment, often in real time, offering greater flexibility in operations and design adjustments (e.g., planning several variants in switching the production line to new products) [28].

According to our concept, creating an AI-based DT for any problem involves several key steps, listed as follows:

• We start by collecting comprehensive data from the physical system, including sensor data, operational history, and environmental conditions.
• We then preprocess and clean these data to ensure accuracy, removing noise and incomplete or outlier values that could skew the AI model.
• We select the appropriate AI model (traditional artificial neural network, deep learning algorithm, or a hybrid AI combination, e.g., with fuzzy logic or fractal analysis) depending on the complexity of the problem and the nature of the data.
• We train the AI model using the preprocessed data, allowing it to learn the behavior and patterns of the real system over time.
• We validate the model’s performance by comparing its predictions with real-world data, adjusting parameters, and retraining as needed to improve accuracy.
• We integrate the trained AI model with the simulation environment to create a DT, ensuring that it can replicate the behavior of the system under different conditions with the accuracy required for our application.
• We continuously update the DT with real-time data from the physical system, enabling it to adapt to changes and improve its predictive capabilities (also as the behavior of the real system changes, e.g., as it wears out).
• We use the DT to simulate and predict outcomes for different scenarios, providing insight into potential issues and optimizing system performance.
• We implement a feedback loop where the performance of the physical system is monitored and compared to the DT, improving the model over time to increase its accuracy.
• We implement the DT of the digital twin in a user-friendly interface, enabling engineers and decision makers to interact with it, analyze results, and make informed decisions about maintenance, design, and operational improvements.

AI uses the employee’s experience, intuition, and invention. It generates knowledge from data. The lack of clear relationships and rules of thumb prevent/hinder the use of traditional methods.

DT improvement is an ongoing process that never ends as long as the real system is in operation. The operation of a DT can be affected by modifications to both the real system and the DT (including basing the DT on a new simulation engine, environment, or interface). Synchronization and optimization constitute a continuous process (Figure 2).

The architecture of our artificial intelligence-based DT system using 3D printing processes as an example is described in Figure 3. It provides a dynamic virtual representation of a physical object/process at selected stages of its lifecycle. Our DT system covers these selected physical processes: 3D model creation, 3D model preparation for printing, product printing, and product quality assessment. In the databases, for product printing, we have descriptions of these physical processes (e.g., production type, material type, layer height, coating thickness, bottom thickness, top thickness, fill density, print speed, bed temperature, print temperature, second nozzle temperature, build orientation, number of contours, and tensile strength). The systems can both provide observations on the current state of the system and answer “what if” questions. Multiple neural networks have been defined for each decision problem at the product printing stage and their specific tasks. In the knowledge base, we have descriptions in the form of NN models. When an order comes in from a customer, the control module distributes the tasks to the different departments of the company. The order is processed. New physical processes are the basis for expanding the data and knowledge in the system. In this way, the system learns continuously, and knowledge is accumulated as the system runs [34].

Experimentation and testing have been introduced into the DT as a new entity. In the new entity, experiments and tests concern the creation and testing of new physical processes and the creation, testing, and validation of new neural network models for new physical processes. In addition, neural network models for existing physical processes are being improved.

In DT research, the selection of appropriate ML methods and the tuning of their hyperparameters is crucial for accurate and efficient modeling. The choice of the method and settings depends on the specific application, the characteristics of the dataset, and, indirectly, the available computational resources. For this purpose, researchers use a combination of theoretical knowledge, empirical testing, and automated optimization techniques to achieve optimal results.

3. The Roles of Our DT

A DT is a virtual representation parameterized based on real process data in order to model, simulate, monitor, analyze, and optimize the physical systems they represent. DT theory involves creating virtual models of physical objects, systems, or processes to simulate, analyze, and optimize their real-world counterparts [29,30,31]. These digital replicas enable real-time monitoring, control, simulation, and predictive analytics, providing insights into performance, efficiency, potential problems, and optimization opportunities [29,30,31,32]. By integrating data from sensors and IoT devices, DTs can reflect real-time changes and predict future behaviors using advanced analytics and AI/ML. This improves decision making by providing a risk-free environment for scenario testing and process improvement. Originating from space technology, DTs have expanded into various industries, including manufacturing, healthcare, and urban planning. They facilitate preventive maintenance, reducing downtime and costs by identifying problems before they occur. In smart cities, digital twins help optimize asset management and infrastructure planning. They also play a key role in personalized healthcare by modeling individual patient conditions to adjust treatment plans. As digital twins evolve, they are expected to become more autonomous, with self-optimization capabilities that will further enhance operational efficiency and innovation [33,34]. Insight into how DTs achieve their benefits is provided by examples of their use in households.

• Home energy management: DTs of home energy systems can monitor and optimize energy consumption by simulating different scenarios and suggesting the most efficient solutions, such as switching appliances on and off at the right times to reduce costs and energy consumption [35].
• Smart refrigerators: DTs of refrigerators can analyze their interior, monitor the freshness of products, and suggest purchases and menus based on available ingredients; they can also simulate a refrigerator’s energy consumption and suggest optimal temperature settings [35].
• Automation of home appliance maintenance: DTs of household appliances such as washing machines, dishwashers, or air conditioners can monitor their health and predict failures, warning household members of the need for service before a major failure occurs [35].
• Heating and cooling system management: DTs of the HVAC (heating, ventilation, and air conditioning) system in the home can simulate different weather conditions and adjust settings in real time to provide the required comfort to household members while saving energy [35].
• Smart garden management: a garden DT (together with devices for watering, mowing, etc.) can monitor soil conditions, sunlight, and the water needs of plants, optimizing watering and fertilization, leading to healthier plants and water savings [34,35].

DTs can solve problems by enabling real-time monitoring and analysis of machines and systems, allowing for the early detection of anomalies and the prevention of unexpected failures. They provide a virtual testing ground for new designs and modifications, reducing the need for physical prototypes and speeding up the process of technology and equipment development. By simulating various operating conditions and risk scenarios, DTs help optimize performance, increase efficiency, and extend the life of components. They facilitate predictive maintenance by analyzing data trends to predict when maintenance will be required, thereby minimizing downtime and associated costs. In addition, DTs enhance the ability to customize and tune machines to specific requirements, ensuring optimal functionality. With continuous feedback loops and data integration, digital twins support the continuous improvement and innovation of solutions [36,37,38].

It seems that the application of DTs in the field of mechanical engineering should be widespread, which we will explore in the review section later in this article.

3.1. Rules of DT Construction according to Our Concept

Considered in the DT, the subsequent stages of our DT preparation include the following:

• Sensors, measurements, and data collection;
• Data cleaning and preparation for use in the model;
• Process modeling;
• Component performance and the overall DT;
• Evaluation, adjustment, and development;
• Monitoring and control;
• Use in practical applications and conclusions/improvements from them [6].

Sensorization provides opportunities to measure and describe the characteristics of the tested system, process monitoring, and control. This allows us to isolate important parameters, connections between them, and mechanisms implemented in the process, which are often simultaneous and not stated directly. Sometimes this allows us to describe the process more precisely in order to improve its final quality, expressed not only in greater accuracy but also in shorter duration, lower energy consumption, and the amount of waste. Acquiring and analyzing signals is useful for understanding the correct behavior of a given system and detecting situations of incorrect operation, or even predicting them, which we will describe in more detail in the section devoted to the use of AI to create DTs. It should be noted that research on sensors and measurement equipment is closely related to the monitoring and development of monitoring strategies described below [9].

Moreover, the quantity and quality of data are crucial. Their quantity, diversity, certainty, completeness, lack of bias, and balance in classes ensure the correct data input to the model. For this reason, data preparation is as critical to the model as the modeling process itself and the selection of its parameters/algorithms.

Modeling is one of the main objectives, as models provide the possibility to understand, monitor, and control production processes in the laboratory and in industrial applications. Such modeling is based on both the results of empirical experiments and theoretical motels, with many hypothetical mechanisms. Reproducible experiments and models reflecting them are useful as being oriented more towards real-world applications than discussion. They also allow for consideration of the main controls and their algorithms. This part of DT development often builds on simplified solutions (e.g., an Ishikawa diagram—Figure 4) and develops them progressively to more complex, multivariate, and plausible models. In the following section, we will discuss how AI/ML streamlines this stage, allowing for movement straight to more advanced models [9].

A significant part of the research when creating a DT is devoted to the modeling and characterization of individual key components (system components). For this purpose, the main characterization studies are described and commented on—both theoretical (e.g., sets of relevant equations) and experimental (descriptions of the experimental equipment and experimental procedures used to characterize the performance of the components in terms of efficiency and performance and experimentally obtaining the relevant coefficients). This applies to both proper operation and cases of natural wear and damage (e.g., cracks, operation in improper conditions, etc.). This gives not only better coefficient values and a reference to optimal operating conditions but also a classification of the operating conditions of a given element as correct or defective (e.g., leading to premature wear or damage) [9].

Advanced use cases are a good way to better understand the process itself. Practical applications of DTs not only allow for detailing the process itself and taking into account their goodness of fit for application but also for isolating of their main features desired by users and increasing the role of DTs in solving potential problems and ensuring user satisfaction with DTs (Figure 5) [9].

3.2. Rules of Our DT Operation

The most important goal for research (both in the area and in interdisciplinary research using AI and engineering solutions) dealing with DTs and considering the industrial improvement of the processes described by DTs as an important goal is to focus efforts on DTs as a complete model of a machine and process capable of exchanging data from a real twin in order to accomplish the following:

• Improve the ability to design new operating cycles (including their optimization) [9];
• Improve the final quality of the workpiece/service in an economical and sustainable way (energy-saving, with minimal waste, and, in some cases, saving water or coolant) [9];
• Anticipate and prevent unexpected phenomena, damage, and downtime [9];
• Ensure predictive maintenance while maintaining production smoothness [9].

We would like to point out that the above-mentioned principles apply to both traditional processes, such as milling and turning, and much more demanding ones, such as waterjet cutting, laser cutting, or additive manufacturing (3D printing). The description methods are universal, but, for example, when cutting with a water jet, there are some delays that should be taken into account [6]. In any such case, the entire DT development process must be followed and reorganized to define the original DT concept for a given application.

We would like to point out that, depending on the application, there is already CAD/CAM/simulation software on the market enabling the simulation of production processes. It is worth using them, even to identify shortcomings. In addition, to create appropriate communication between the elements of the digital data flow in the process, the industrial Internet of Things (IIoT) should be used, and process and product/service data should be protected against unauthorized access, leakage, or cyber attacks. This requires not only investment and an open approach from machine designers to innovative process monitoring but also the preparation of engineering staff. When developing DT, CAM systems can not only pay attention to traditional parameters (tool path, process parameters such as type and dimensions of the material or cutting depth, and tool movements) but can also constitute an advisory module in selecting the best tool configuration depending on the application (cutting trial, target material, high precision, etc.). CAM can also take into account and compensate for simultaneity, convergence, and delay, and, in some global applications, time zone differences. Future solutions may include programs for multi-tool machines (including combinations of machining and 3D printing), for the production of products from various materials, for feeders (abrasive tools) depending on the position on the tool path, and for receivers with a disinfection and packaging robot. It is also possible to introduce new functions for closing the control circuit on the same machine. The DT simulation must be performed offline before the actual machining operation. This will create nominal patterns of (virtual) signals that can be compared to real ones (both during simulation and during operations). However, this increasingly requires the division of roles into edge computing (EC) as close as possible to the production line and cloud computing (CC) at the level of the control of the entire line or smart factory. It is in the company’s interest that all signals from the machine are received and processed—this will not only allow for the development of a more reliable DT, but in the event of errors, it will allow them to be found faster [9].

The principles of our DT operation ensure its effective functioning and delivery of expected benefits. These principles include data integration, model accuracy, real-time processing, and security measures. This can be presented in detail as follows:

• Accurate representation: initial model fidelity, i.e., the DT must start with an accurate and comprehensive model of the physical entity, including all relevant physical properties and behaviors, and then be subject to continuous calibration; i.e., the model should be regularly updated based on new data (but also staff observations) to maintain accuracy over time [22];
• Real-time data integration: continuous data flow from sensors and IIoT devices installed in the physical system to capture real-time operational data and reliable data-processing capabilities to efficiently handle large volumes of real-time data [22];
• Data management and storage: the scalability of data storage as the amount of data from a physical unit increases and the integrity, accuracy, and completeness of collected data [22];
• The adoption of industry standards for data formats and communication protocols to facilitate seamless data integration and exchange [22];
• The integration of the DT system with other digital tools and systems (ERP, PLM, SCADA, and others) to provide a comprehensive view of the situation and the selection of appropriate AI/ML algorithm(s) based on specific needs and characteristics of the system [22];
• Regularly training AI/ML models, using both historical and real-time data, to improve their predictive capabilities and accuracy [22];
• Testing various scenarios without affecting the physical system as predictive analysis to forecast potential problems and proactively optimize system performance [22];
• Implementing feedback loops where the digital twin can inform and adapt the performance of the physical system based on analysis and simulations [22];
• Incorporating user feedback into the operation of the digital twin to continually improve its performance and usefulness [22];
• Implementing proven cybersecurity measures to protect data and models from unauthorized access and cyber threats [22];
• Intuitive and user-friendly interfaces for interaction with the DT, granting authorized users easy access to information (on a need-to-know basis) and control over the system [22];
• Providing users with training and a user support system (Helpdesk) to maximize the usability of the DT [22];
• Use of the DT to monitor and manage the entire lifecycle of a system, from design and production to operation and decommissioning [22];
• Relying on key performance indicators (KPIs) and metrics to monitor the performance of both the digital twin and the physical system [22];
• Developing DT adaptations to changes in the physical system, such as updates or modifications, without significant rework [22].

We will discuss the challenges in the Discussion section, but we would like to highlight them here. The main challenges for AI-based DTs include the following:

• Data management: handling and processing large amounts of data in real time can be complex and resource-intensive [22].
• Integration: seamlessly integrating DTs into existing systems and workflows can be a challenge [22].
• Accuracy: ensuring that the DT remains an accurate representation of the physical system requires sophisticated models and constant updates [22].
• Security: protecting sensitive data and maintaining the cybersecurity of a DT system is critical [22].

3.3. Interoperability of Our AI-Based DTs in an Innovative Industry Ecosystem

Industry 4.0 and Industry 5.0 have transformed the manufacturing industry towards full connection between the elements of the production line, with more effective use of data from production processes and the placing of humans and their environment at the center of industry attention [39]. The integration of sensors, effectors, communication technologies, computing platforms, AI/ML, control systems, and predictive tools is used to effectively manage processes in real time [40]. New technologies, such as DTs, can further intensify the above-mentioned processes towards the development of smart factories [41]. The combination of real (in the production line) and digital (in one or more DTs) versions of the system allows for real-time data exchange for the purposes of the control, monitoring, prediction, and classification of events [42,43,44,45,46,47,48,49,50,51,52]. Thus, classic production control turns into a forward-looking process, reacting in advance, making optimal use of all opportunities that will appear, and reacting, in advance, to the needs of repairs and adjustments, anomalies, and production disturbances [42,43,44,45,46,47,48,49,50,51,52].

In the current comprehensive approach, it has been shown that AI/ML in DTs is an effective tool for handling production data in the field of second opinion and quality control systems [45,46,47,48,49], including process quality monitoring [53,54]. It is also possible to identify service patterns for large amounts of unlabeled data generated in the production process, e.g., when detecting and classifying anomalies, predicting service, faults, and quality [46,47,48,49,50]. Generative adversarial networks (GANs) or a hybrid combination of different ML techniques work well in anomaly detection [51] but also in oversampling too-small experimental datasets to feed ML algorithms for forecasting [52].

A separate problem is predicting the life of production lines, products, and services. For practical reasons, a hybrid approach is most often used—a combination of decision trees and artificial neural networks (ANNs) [53]. Reinforcement learning can ensure that BTs are goal-oriented and useful for decision making, such as planning, schedule optimization, maintaining production continuity, and quality control [54,55].

The availability of process data is crucial for the proper operation of AI/ML methods and techniques for DTs. Unfortunately, conducting all experiments (e.g., regarding damage) for the purpose of obtaining data from a real production line is often impossible, because it is expensive and time-consuming, involves stopping and reprogramming the production line, and wastes raw materials. Therefore, production simulations are mainly a source of data and knowledge about the uninterrupted operation of the real production system [56,57]. Only DTs can enable research into many scenarios of the anomalous operation of the production line and its configurations, as well as experimental tests in the scope of modifying the line’s operating conditions for new experimental conditions, products, services, and, e.g., susceptibility to cyber attacks.

4. Materials and Methods

4.1. Dataset

When developing bibliometric analysis methods, we focused on understanding the research picture in the area of AI-based DTs in engineering and the possible progress of knowledge and practice in this area of research. We formulated a research question (RQ) that will help us discover the key aspects of the research area regarding finding the concept of creating an AI-based DT in engineering. This is consistent with both the title and the Introduction section and the objectives of this paper. The RQ formulated in this way is significant and can be investigated based on bibliometric data (literature review). The application of AI in engineering is closely related to the generation of DTs, as AI enhances the simulation, analysis, and optimization of physical systems, making DTs more accurate and effective. As industries recognize the value of DTs in improving design, testing, and maintenance, the emphasis on AI applications in this context is likely to increase, reflecting a significant trend in both academia and in industries. Overall, the intersection of AI and DT in engineering represents a key research area with great potential for future development and innovation. Such analysis and interpretation of bibliometric data can make a key contribution to the ongoing discussion and build a more solid foundation for further analyses and research (Figure 6).

4.2. Methods

In this review, we used tools built into the Web of Science (WoS) database.

5. Results of the Review

5.1. General Results

Despite expectations, the number of publications in the WoS database with the keywords (in Topic) “digital twin”, “mechanical engineering”, and “artificial intelligence” was zero, and, with the keywords “digital twin”, “mechanical engineering”, “machine learning”, there were three [42,43,44]. There are many more publications in areas other than mechanical engineering, as Figure 7, Figure 8, Figure 9 and Figure 10 show. This indicates the need to intensify research on AI-based DTs used in mechanical engineering and publish them in the best journals indexed in the WoS database. This may also be due to the limited amount of interdisciplinary research.

On the basis of such results, it can be concluded that the concept of mechanical engineering is a very general one, and perhaps that is why the literature cannot be found. Hence, we checked the following:

• “digital twin”, “3D printing”, and “artificial intelligence”: 16 (WoS, 2020–2024);
• “digital twin”, “3D printing”, and “machine learning”: 21 (WoS, 2019–2024).
• This also indicates a research gap regarding our example in Figure 2 and Figure 3.

Virtual inspections and assessments of infrastructure assets (e.g., bridges) are capable of creating an accurate digital representation of existing assets (DTs). They are an alternative to in-person and site-based assessments, while being cheaper, more reliable, and requiring fewer distributive bridge inspections. For bridge monitoring, the most popular technologies include unmanned aerial vehicle (UAV) photogrammetry and terrestrial laser scanning (TLS). These are characterized by their potential to provide qualitative digital models based on the assessment of generated point clouds in terms of quality and geometric accuracy for an asset of this size, such as a bridge. The assessment addresses point distribution, outlier noise level, data completeness, surface deviation, and geometric accuracy [45]. Thus, Bridge Information Modelling (BrIM), being a specific form of Building Information Modelling (BIM), provides a digital link between bridge assets related to geometric and non-geometric control data. A comprehensive asset management system has been developed that uses BrIM data to enhance and facilitate the BMS. It has shown potential in replacing traditional paper documentation and managing bridges efficiently and effectively, in addition to being less prone to subjectivity in assessment [58].

5.2. Examples of Tasks Performed Better than Previous Methods

Despite the concept of AI-based DTs being at the beginning of their development, several studies have already been published showing their advantages over traditional solutions. In order to solve the problem of noise and similar gray values between the foreground and background of pellet images, an AI-based DT based on superpixel features was developed. The binary SVM classification model was used to transform the image segmentation problem into a foreground and background classification problem, and the four neighbor search algorithm was proposed to reduce the mis-segmentation rate of edge superpixels. The high accuracy of the proposed method was achieved, at 95.87%, which enabled visual analysis and decision making [59]. An autonomous mobile robot platform capable of simulation-based navigation using NVIDIA’s Isaac Simulation 4.1.0 software was developed. AI was used here to analyze environmental data and simulation inputs, enabling the robot to dynamically change course or perform specific tasks. This resulted in an increase in the robot’s adaptability and performance in different scenarios [60]. AI-based fault detection and remote robot control are being explored in the AI-enabled Cyber–Physical In-Orbit Factory project, which shows how AI is being used to make manufacturing more robust, fault-tolerant, and autonomous, including in a robotic Assembly, Integration, and Testing (AIT) system where a small satellite could be assembled by a robotic manipulator from modular subsystems [61].

The aforementioned research has shown that artificial intelligence (AI)-based digital twins outperform traditional digital twins due to their ability to perform more advanced data analysis and prediction. AI allows for dynamic learning from large datasets in real time, enabling more accurate modeling and prediction of the behavior of complex systems. In traditional digital twins, simulations are typically based on fixed models, which can be less flexible and slower to respond to change. With AI, digital twins can better identify patterns and anomalies and optimize processes, leading to higher operational efficiency rates. This research also highlights AI’s ability to adapt and make autonomous decisions, which increases the value of digital twins in the context of managing and optimizing industrial systems.

6. Discussion

So far, we have presented our concept and its applicability to already formulated and partly investigated problems. However, this does not exhaust the possibilities, as solutions such as AI-based DTs are only at the threshold of their development, requiring further research to address the many challenges they face. The complexity of the issues discussed requires the introduction of a coherent development strategy and the spreading of projects over time. The existing approaches described in the literature were not as comprehensive as we would like; hence, the limitations and challenges were much greater. Here, we would like to narrow them down to the key ones in an interdisciplinary approach.

As an example, we consider the difference between an AI-based DT for autonomous vehicles and AI-based driving software. Although both are AI-based, they have different implications for the user and use different data sources, so the ethical and bias implications of the two are different. The DT for autonomous vehicles is a virtual replica of a physical vehicle that simulates its behavior, interactions with its environment, and reactions to various road conditions. It allows for the testing and optimization of systems in a controlled environment, which minimizes the risk of errors during real-world use. By contrast, autonomous driving software relies on AI algorithms that continuously process data from vehicle sensors, making real-time decisions on the road. The DT uses a wide range of data, including environmental simulations and historical vehicle data, while autonomous driving software relies primarily on live data. This difference has ethical implications, as the DT can help identify and correct bias in algorithms in a safe environment before being utilized on the road. Autonomous driving software must not only be precise but also transparent in its decision making to minimize the risk of poor decisions. The DT can be used to analyze long-term environmental impacts, something that autonomous driving software cannot do in real time. Both solutions, while based on AI, serve different purposes; one focuses on improvement and testing, while the other focuses on safe driving. AI bias in the DT can lead to erroneous conclusions during testing, while bias in autonomous driving software can directly threaten road safety. Therefore, ethical considerations related to the use of these technologies must take into account the differences in their application and impact in the real world [62,63,64].

6.1. Main Limitations and Challenges

Highlighting the main limitations and challenges related to the use of AI to generate digital twins should be prioritized, as they must be solved, opening up opportunities for further development of the discussed group of technologies. Chief among these is the collection of high-quality, comprehensive datasets, which are required for accurate AI models, but obtaining such data can be challenging due to sensor limitations and data inconsistency. Moreover, creating and operating DTs requires expensive computational resources, which may constitute a barrier to their widespread use. Moreover, integrating AI into existing engineering systems and software can be complex and time-consuming, requiring specialized knowledge. Achieving real-time data processing and simulation in DTs is difficult due to the previously mentioned high computational requirements. Also, ensuring that AI models accurately reflect the physical behavior of complex systems is difficult, especially in dynamic environments. Scaling DT solutions to cover large and complex systems without sacrificing performance or accuracy is a significant challenge. Ensuring compatibility and seamless communication between the various tools and software platforms used in engineering can be problematic. There is a lack of specialists with all the skills necessary to develop and implement artificial intelligence-based DTs. In this case, it is easier to rely on interdisciplinary teams, but assembling them is also not easy, and training specialists will take at least several years. The development, deployment, and maintenance of AI-based DTs can be prohibitively expensive, limiting their availability to smaller organizations. A system of subsidies or implementation grants may help here. Protecting sensitive data related to the creation and operation of DTs from cyber threats is crucial but challenging. AI applications in DTs must meet industry standards, and regulatory requirements can be complex (e.g., the EU AI Act). Effectively managing uncertainty in both the data and the models themselves is an ongoing challenge. Continuously monitoring and updating AI models to reflect changes in the physical system or new data is necessary but resource-intensive. From the above reasons, it is necessary to rigorously validate and verify that DTs accurately represent physical systems, which requires extensive testing and expertise. Ensuring that the use of AI in DTs complies with ethical standards, including privacy and bias mitigation, adds an additional layer of complexity [65,66,67].

In the case of virtualized systems, AI-assisted DTs may be limited by the quality and accuracy of the input data, which affects the precision of simulations and predictions. AI models may not fully account for all the complex physical interactions, leading to simplifications or representation errors. The process of training AI models requires a large amount of data, which can be time-consuming and expensive and not always available for all systems. DTs may also not be able to handle unpredictable events or failures that were not accounted for in models trained on historical data. Lastly, computational limitations can affect the ability to perform complex simulations in real time, especially for highly complex systems. The current trends in digital twin development is shown in

Table 1. Current trends in digital twin development.

Trend	Description
Manufacturing Efficiency	In manufacturing, DTs are used to optimize production processes, predict maintenance needs, and reduce downtime by accurately modeling machines and the flow of work, intermediates, and materials.
Healthcare	DTs are developed for healthcare, where they simulate patient-tailored models to improve diagnosis, personalize treatment plans, and assess patient outcomes.
Smart cities and territories	Cities are deploying DTs to model urban environments, improving planning, infrastructure management (including transport and energy), and emergency response, providing a detailed, real-time virtual representation of urban operations.
Aerospace	DTs are increasingly being used for predictive maintenance, increasing aircraft performance and optimizing fleet management through data analysis and real-time simulation.
Sustainability	There is an increasing emphasis on using DTs to achieve sustainability goals, such as reducing energy consumption, optimizing resource use, and minimizing environmental impact.

.

6.2. Directions for Further Research

In the area of such complex systems, it is crucial to define research priorities with the greatest impact on the development of AI-based DTs. The largest research teams will work on them, and the largest financial resources will be allocated to them. It is important to agree on them and coordinate them in advance so that the efforts of researchers complement each other and do not duplicate each other. The starting point is research into advanced sensor technologies and data collection methods to improve the quality and comprehensiveness of datasets for digital twins. Real-time data processing requires the optimization of algorithms and computational resources, as well as the simulation of entire systems. It is necessary to develop hybrid models that combine artificial intelligence with traditional physics-based simulations to improve accuracy and reliability. Solutions should be scalable; hence, it is necessary to explore scalable AI structures and architectures that can effectively support large-scale and complex systems. This will be helped by establishing standard protocols and frameworks to ensure seamless integration and interoperability between different software tools and platforms (including between current and future transition platforms). Implementation must be cost-effective, so research should explore ways to reduce the costs of developing, deploying, and maintaining AI-based digital twins to make them more accessible. Further development requires the development of AI models that can self-adapt and learn from new data to continually improve their accuracy over time. In current geopolitical and global market conditions, it is important to develop and explore robust cybersecurity strategies to protect data and AI models from cyber threats and ensure data integrity. The shortage of specialists creates a need to create and update educational and training programs to fill the skills gap and equip specialists with the necessary expertise in artificial intelligence and digital twins. The standardization and comparability of solutions increases the need to develop a comprehensive regulatory framework that can keep pace with rapid advances in artificial intelligence and ensure compliance with industry standards. In the case of uncertainty analysis, it becomes necessary to improve methods for quantifying and managing uncertainty in AI models (e.g., via fuzzy logic) in order to increase the reliability of digital twins. Exploring artificial intelligence-based automatic maintenance and updates for digital twins aims to reduce the need for manual intervention, reduce service time, and increase service efficiency. Research into ethical AI practices is needed to ensure that digital twins are developed and used in a way that respects privacy, fairness, and transparency. Promoting interdisciplinary collaboration between AI researchers, engineers, and experts stimulates innovation and the practicality of solutions. For this purpose, it is necessary to establish benchmarks and performance metrics to assess and compare the effectiveness of different AI approaches in generating digital twins [68,69,70,71].

The use of AI in DTs may raise ethical concerns [72,73,74,75,76,77], in particular, leading to significant privacy issues, as the detailed data collected and analyzed may contain sensitive personal information, requiring robust data protection measures to protect individuals’ privacy. AI algorithms powering DTs may inadvertently perpetuate or exacerbate biases present in the data, leading to unfair outcomes and discrimination, especially if the data are not representative or the AI is not carefully monitored and adjusted. Ensuring transparency in AI models is key to ethical accountability. AI decision-making processes in DTs can be opaque, making it difficult to hold actors accountable for AI-based actions and decisions. DTs integrated with AI can become the target of cyber attacks, potentially leading to unauthorized access and misuse of critical systems and data, raising serious ethical concerns about digital infrastructure protection. The implementation of AI in DTs raises concerns, as it may lead to the redundancy of staff currently performing tasks that can be automated. This requires ethical considerations regarding retraining and social impact on affected workers [78,79,80,81,82].

7. Conclusions

AI-based DTs are a partly new tool for monitoring, simulating, and optimizing systems. DTs itself are not a new technology, but AI-based DTs require further and more dedicated research due to their specificity of semi-automatic or automatic creation and the need for supervision. Widespread adoption and mastery of the technology will lead to significant improvements in performance, reliability, and profitability [78,79]. However, this requires a significant upfront investment in transforming traditional approaches to AI-based DTs, training a sufficient group of specialists from various industries who are ready to fully exploit the capabilities of this technology, and working with scientists to develop them [83,84].

The emergence of AI-based DT technology has created unique opportunities to use data transmitted in real time from the physical environment to the digital equivalent of machines, devices, and production lines. AI plays a key role in AI-based digital twin (DT) technology, enabling the creation of highly accurate and adaptive virtual models of physical systems. AI algorithms process vast amounts of data from sensors and other sources, enabling DTs to more accurately reflect real-world system behavior. AI also enhances the predictive capabilities of DTs by identifying patterns and trends in historical and real-time data, enabling more accurate predictions of system behavior. In addition, AI-based models can simulate more different scenarios, providing insight into potential outcomes and helping to optimize performance and prevent failures. AI enables DTs to learn and evolve over time, continually improving their accuracy and relevance as new data are ingested. In this way, AI enables the digital twin to handle complex, nonlinear relationships within the system, enabling the modeling of complex problems that traditional methods struggle to solve. AI also facilitates anomaly detection and diagnostics in DTs, identifying problems earlier and more precisely before they manifest in the physical system. AI helps automate decision making by providing recommendations based on simulations and analyses of the digital twin. AI-based DTs can more easily integrate with other AI systems, enabling holistic monitoring and control across multiple interconnected systems. AI plays a key role in the scalability of digital twins, enabling them to be applied almost automatically to increasingly complex systems with minimal manual intervention and adaptation by human experts. Finally, AI drives the continuous improvement of digital twins, ensuring they remain accurate and valuable tools for system management and optimization in the long term.

Although progress has been made, DTs as a problem-solving paradigm still need to be improved. Adopting our concept proposed in Section 5 instead of traditional approaches can become standard practice, rather than an advanced rescue solution for the most difficult problems observed in engineering, and this will help accelerate this development.