Review and Insights Toward Cognitive Digital Twins in Pavement Assets for Construction 5.0

Abstract

With the movement of the construction industry towards Construction 5.0, Digital Twin (DT) has emerged in recent years as a pivotal and comprehensive management tool for predictive strategies for infrastructure assets. However, its effective adoption and conceptual implementation remain limited in this domain. Current review works focused on applications and potentials of DT in general infrastructures. This review focuses on interpreting DT’s conceptual foundation in the flexible pavement asset context, including core components, considerations, and methodologies. Existing pavement DT implementations are evaluated to uncover their strengths, limitations, and potential for improvement. Based on a systematic review, this study proposes a comprehensive cognitive DT framework for pavement management. It explores the extent of enhanced decision-making and a large-scale collaborative DT environment. This study also identifies current and emerging challenges and enablers, as well as highlights future research directions to advance DT implementation and support its alignment with the transformative goals of Construction 5.0.

1. Introduction

The construction industry exhibits lower productivity than other sectors, primarily due to factors such as the slow adoption of technologies [1,2]. According to a recent industrial survey report, construction was identified as one of the least digitised industries considering assets, usage, and workers [3]. Industry 4.0 technologies have contributed to the digital transformation of many sectors, including construction. Industries 2.0 and 3.0 focused on mechanisation and automation of manufacturing processes, respectively, whereas the fourth industrial revolution relied on digitalisation, automation, big data, and innovative processes [4]. Industry 4.0’s recent focus is building interconnected environment applications to operate and maintain assets efficiently [5]. However, the proposed evolution to Industry 5.0 is based on combining collaborative human and machine intelligence. This aims for more informed decisions and integrated sustainable systems [4,6]. Furthermore, the connection between the physical and digital worlds is moving beyond current digital enablement and toward digital control [7]. This vision integrates the creativity of humans with intelligent machines and digital systems to collaborate productively [8,9]. For the construction industry, moving toward construction 5.0 involves incorporating advanced technologies to address the complexities of current infrastructure management, i.e., road pavement assets. This includes autonomous collection systems, realistic asset state prediction, and intelligent decision support systems. Currently, Construction 4.0 aims to enhance industry productivity and enable cooperation with assets by the help of technologies [10,11]. These technologies include Digital Twins (DTs), Artificial intelligence (AI), Internet of Things (IoT), Building Information Modelling (BIM), and Geographical Information Systems (GIS). Its applications expand across different phases of the construction assets, including operation and maintenance [6]. AI enables predictive analytics and recognition techniques, and IoT provides real-time data collection through facilitated communication between devices and cloud networks. Moreover, BIM involves modelling information that enriches 3D models for standardised lifecycle data management, and GIS supports spatial and temporal data analysis and visualisation. DTs are a virtual representation of a physical entity that enables real-time monitoring and lifecycle analysis of its mirrored entity [12,13,14]. However, DT leverages the combined capabilities of different technologies for comprehensive and enhanced data-informed management. Construction 4.0 and 5.0 within a broad industrial revolution timeline are presented in Figure 1 below.

The global DT market is expected to have 50% compound annual growth, reaching 184.5 USD billion by 2030 [15]. In this market, built environments, including intelligent buildings, smart cities, energy, etc., form around 20% of the overall value [16]. Smart city DT solutions alone will reach $5.9 billion by 2029, and most IoT-based platforms will achieve DT capability [17]. Despite this significant growth in the DT market, the growth of the DT concept, and digitalisation in the construction industry, the development and implementation of mature DTs remain limited and in their early stages [18,19,20,21]. Although examples are being developed in some applications, it is still in the prototype stage [20,21,22,23] and, in some areas, lacks the complete concept of DT adoption [21,24]. For instance, most examples of asset management are limited to information collection or virtual models, with less attention paid to road assets [25,26]. However, the barriers towards its efficient adoption in the construction industry include a lack and fragmented comprehension of DT concepts [24], absence of clear vision, and standardised DT implementation [27,28], which results in interoperability, data integration, and management issues [20]. Moreover, the limited prototypes and holistic frameworks, as well as the complexity of developing and maintaining DTs, have created more barriers. This is also due to a shortage of specialised digital skills in the construction industry [19,20,25].

For road pavement infrastructure management, DTs could enable predictive maintenance, proactive management of resources, and extended asset life [20,25,26]. Despite its numerous advantages, it is not yet well-developed and lacks full design considerations and implementation.

Existing road pavement infrastructure encounters several challenges, from adapting to growing demands and priorities of maintaining and improving service to expanding its lifespan. Pavement assets are the main contributors to energy consumption and emissions [29]. Its management is critical for the safe and efficient movement of users and the economy, and it must be managed efficiently [30]. Its operation and maintenance phase is usually a point of concern due to the cruciality of any possible improvement [31]. The following sections present an overview of flexible pavement management and the emerging solution of DTs.

1.1. Flexible Pavement Assets

Across industries, different maintenance strategies, planning, and implementation approaches are used. These approaches are corrective, reactive, preventive, condition-based, predictive, and prescriptive. Corrective approaches implement actions only when a part of the system fails. Preventive focuses on actions at a pre-defined set of times [32]. Condition-based assessments rely on evidence of asset deterioration or deviation from normal service levels [33]. The predictive approach predicts ahead of time whether the asset is going towards damage [34]. DTs can improve different interventional methods, from current preventive and reactive strategies to predictive and cognitive approaches, such as predictive and prescriptive maintenance. Figure 2 shows the developed industrial and technological evolution of maintenance strategies.

The flexible pavement asset consists of multiple structural layers. Its evaluation focuses on structural and functional aspects and is impacted by operational and environmental factors. Structural refers to the pavement’s ability to withstand service conditions, such as resistance to deformation and cracking under continuous cycles of loads. This is assessed through field and laboratory testing and mechanistic modelling. Onsite structural evaluation is based on data derived from non-destructive testing (NDT) methods such as the Falling Weight Deflectometer (FWD) and traffic speed deflectometer [35]. In addition, inspections and measurements of structural cracking, rutting, or pothole distress are used as metrics for structural integrity, as they reflect subsurface failures. Ground penetration radar (GPR) is another onsite effective method for detecting pavement subsurface failures. The functional performance refers to the pavement’s surface smoothness and conditions, presenting ride quality and roughness, i.e., surface cracks, potholes, and skid resistance. The severity and intensity of these conditions can be detected by visual inspection activities or through measuring methods, i.e., ride quality testing methods. On the other hand, structural and functional factors interact in terms of their development mechanisms, i.e., excessive rutting is a factor that causes crack formation. Moreover, environmental conditions, i.e., moisture and temperature, drive material interaction within pavement layer performance. This can cause or accelerate damage development under different operational loads, primarily due to the asphalt material’s viscoelastic and plastic material mechanics. For example, pavement cracking and deformation are influenced by load patterns, loading speed, temperature differentials, material properties, aging, thermal stress, moisture, and wheel–surface contact stresses [36,37,38]. Performance and driving factors are crucial for effective pavement state prediction and management. However, in pavement maintenance, interventional decisions use predefined thresholds, such as quality metrics, to trigger maintenance. This is often after significant damage has occurred [30]. Maintenance decisions also depend on the broader context and factors, including network conditions, environmental factors, treatment costs, service levels, and user risks.

1.2. Pavement Management Systems

The pavement management system (PMS) was introduced in the early 1970s by the American Association of State Highway and Transportation Officials (AASHTO) [39]. PMS is “a set of defined procedures for collecting, analysing, maintaining, and reporting pavement data to assist the decision-makers in finding optimum strategies for maintaining pavements”. In 1989, the Federal Highway Administration (FHWA) required related agencies to implement PMS by 1993 [40]. Over time, various organisations developed systems using different performance indicators and data standards. This results in PMS being a broad term for various systems. Some examples of these systems are PMS, PAVER, Pavement Information and Needs System (PINS), and Rural Infrastructure and Pavement Management System (RIPPS). For instance, road, bridge and culvert data were incorporated into a holistic system called the highway asset management system (HAMS) [41]. The U.S. Army Corps developed the PAVER system and adopted the pavement condition index (PCI) as a quality metric [42]. The RIPPS focuses on managing pavement for low-volume rural roads. Maintenance management systems (MMSs) are subsystems that focus on the function and effectiveness of the maintenance activities [43]. However, PMSs and MMSs are the most used systems in highway management [44].

Performance rating indices are used to rank the health of pavement, such as PCI, International Roughness Index (IRI), Present Serviceability Index (PSI), and Present Serviceability Rating (PSR) [45]. These rating indicators are measured or derived based on road segment evaluation processes using inspecting and testing survey data. However, IRI is the most employed within PMS [42]. The overall process of PMS involves data collection, evaluation and prediction to identify current or future deterioration trends.

1.2.1. Data Collection

Data collection is commonly based on manual pavement surface inspections, which are costly, labour-intensive, time-consuming, prone to errors, and reliant on expert evaluation [44,46,47]. Surveys involve onsite manual measurements and collect photos to report pavement conditions. This extends intervals of updating data and reduces the efficiency of maintenance planning and forecasting its status [42]. Therefore, some agencies have adopted specialised vehicles equipped with sensors, such as lasers, cameras, and deflectometers. However, this resulted in expensive processes and equipment [48]. Moreover, this process resulted in issues like data inconsistency and disruption of traffic flow [47] and, furthermore, the inability to cover the extended pavement area due to time and budget constraints [49], in addition to the limitation of using these expensive methods for relatively small municipalities due to insufficient budgets [50].

Recently, data collection has undergone a significant evolution with new tools and low-cost automated methods [51]. Image processing and AI detection methods for pavement images, such as using deep learning (DL) algorithms, have shown promising results [46,50,52,53]. These tools have improved the surface data collection performance of PMS [48]. Furthermore, the Machine Learning (ML) evolution has elevated processes to further severity measurements for various defect types, as in [54,55]. Moreover, there is a growing use of smartphones and unmanned aerial vehicles (UAVs) to collect pavement data. However, different subsystems and processes were developed separately, and PMSs also developed various subsystems used to support road management. Therefore, the data types and standards variability posed a limitation in the integration process, which resulted in a fragmented management approach [56]. PMSs have no centric dataset and lack smart capabilities in managing data [2]. Currently, to manage road asset conditions spatially and prioritise their interventions, most agencies have adopted global positioning systems (GPS) and GIS [57].

1.2.2. Pavement Performance Modelling

Prediction modelling is used to plan maintenance and optimise preventive maintenance schedules through existing statistical and empirical models. Models used are either developed for pavement design or using survey data to estimate the asset deterioration trends statically. Generally, prediction modelling can be classified as data-driven and physics-based models. Physics-based models, known as purely mechanistic models, rely on experimental structural response data such as stress and strain. Data-driven models include deterministic, empirical or statistical models, i.e., regression models. In addition, Mechanistic-empirical (ME) models incorporate some experimental response parameters in statistical form. Data-driven models rely on mathematical relationships between variables [58]. The modern form is ML models. On the other hand, Probabilistic models, such as Bayesian and Markov, are models that involve probability prediction to account for uncertainty [58]. They need extensive computational resources, work with limited data, and lack quantitative correlation [59].

Statistical models rely on time-independent data processing, which limits their prediction due to the incremental form of pavement deterioration [60]. These traditional models have different well-known limitations, such as data sensitivity, requiring structured and high-quality data, relying on assumptions for data preparation [61]. Moreover, they cannot account for complex and highly nonlinear relationships, i.e., pavement material interaction physics [62], in addition to randomness issues and limited prediction accuracy when a more extensive dataset or number of predictor variables is used [61,63,64].

The nonlinearity of ML models outperforms other methods in terms of accuracy and big datasets. It can predict patterns and deal effectively with multiple factors and randomness issues [51]. However, model knowledge is capped to a limited training dataset and requires high-quality training data. Moreover, it is prone to overfitting as it performs well on training data while poorly generalising to new data. It also lacks interpretability [62], which leads transport agencies to adopt other deterministic approaches due to ML model challenges for decision-makers [51,65]. For instance, Austroads, an Australasian head organisation of road transport agencies, in their last deterioration models reports, stated that most road agency members prefer to use deterministic models such as ME models over ML [65,66]. They noted that the ME model is more stable and reflects the explanatory deterioration of pavement. Moreover, it can be incorporated into the existing PMS [65]. Although ML provides robust models, overfitting and its limited transferability to other data, along with the black-box interpretability and complex implementation, are the main issues of its adoption in PMSs [59,67]. However, ME models have led to the advancement in pavement design [58], and they are more accurate than pure empirical models. The data-driven models often lack the physical interaction of the underlying pavement mechanics, which limits their accuracy in complex pavement systems [67]. However, ME models are data-driven models developed and validated through a one-time validation process and require continuous calibration [68]. Austroads deterioration models for thin pavements are based on a multi-variate nonlinear ME model [65]. The models were developed from data between 1994 to 2018. The fit of models (R²) was only 0.23 for the roughness [67] and 0.5 for the rutting [65]. They indicated the need for calibration for local use conditions.

Using design ME models for analysing existing in-operation pavement involves critical assumptions, as these models were developed from pavement data in different environmental conditions and converted mixed vehicle classes as input [69]. This, in addition to relying on assumptions about the loads the asset handles, which can differ significantly from actual in-situ conditions [70], can lead to misrepresentation of asset performance. In pavement design, ME models use various traffic inputs for future scenario analysis. This can cause contradictions between existing and designed asset performance. Additionally, using these models for predictions during operation relies on estimated traffic loads and constant operation loads and patterns assumptions over the pavement’s service life [70]. This often overlooks traffic patterns, operational changes, and various overloading scenarios. Traffic loading is a critical factor for pavement response, and it can change significantly over time. For instance, the FHWA technical report states that substantial traffic volume changes occurred over time at specific sites, often due to major road construction projects or events that alter traffic patterns [71].

1.3. Digital Twin for Pavement Management

Due to the limitations of data-driven models in capturing complex degradation and the lack of actual operational conditions, it is suggested that blending the strengths of these models could enhance accuracy [59]. The ME, as a simple blend, outperforms previous methods in pavement design [58]. The current PMS struggles with data integration, and limited predictive performance hinders predictive maintenance adoption [2,51,72]. Fragmented processes hinder efficient management and timely interventions, highlighting the need for a technology-driven PMS [51]. The DT concept uses multi-modelling with multi-sources lifecycle data, including live asset data, to improve modelling and integrate processes [2,73]. Predictive methods can enhance interventions, cut emissions, and improve operation and maintenance planning [26,74,75], necessitating frameworks for DT [20,21,24].

However, accelerated digitalisation and unstandardised DT adoption have led to misconceptions in some construction applications. Some technology integration works incorrectly labelled their efforts as DT, presenting it as a trendy term for traditional monitoring or modelling. This misrepresentation hinders the full implementation of true DTs, resulting in low maturity levels of DTs [76,77], along with unrealistic expectations of DT [78], and lack of focus on its purpose [19]. Opoku et al. [79] reviewed drivers of DT adoption in construction, highlighting enhanced predictive maintenance as a key success factor, demonstrating the expected added value in DT development.

Recent advancements in sensing and technologies for pavement have led to the emerging development of real-time monitoring, such as non-destructive methods, i.e., [35,80,81], and recent imagery-based AI defect detection, in addition to instrumented pavement through embedded sensing technologies, as in [82,83,84]. This allows dynamic evaluation of the underlying structural state of the asset for monitoring applications and capturing pavement response, i.e., stress, strain, moisture, and temperature. DTs and these real-time monitoring, i.e., Structural Health Monitoring (SHM), share various similarities. However, SHM provides periodical data collection or real-time monitoring for insights throughout parameter change reflection. Its advancement can enable mature implementation techniques for DTs [25]. DT integrates data processing and optimises employed sensing and monitoring capabilities for broader and more efficient data queries [20]. It expands SHM to simulate and predict changes for the long term, considering the entire life cycle and examining what-if scenario analysis [25,85]. These capabilities of DT are seen as an extension of the SHM and an expansion of its functions [19]. DT’s uniqueness lies in interconnecting previously unconnected systems. This implementation required focus and holistic understanding for successful pavement DT that effectively incorporated suitable and mature monitoring methods.

Recent Reviews of DTs in Transportation

Conceptual DT architecture, the three-dimensional framework, was initially presented in three main components by Grieves [86]: a physical entity, a virtual entity, and a connection between them. However, with the expansion of DT adoption and the need to align with recent technology and applications, the five-dimensional (5D) conceptual DT model was proposed by Tao in 2018 [87]. The new conceptual model added data and service components for broader applicability. The 5D model was referenced in DT development for the mesostructure of asphalt mixture material [88], where virtual moulding of material samples was applied.

However, in transportation engineering, the previous reviews of DT focused on its general applications for maintenance [89], applications in road and railway networks [26], and broad technologies used for infrastructure life phases [90]. Other transportation reviews were conducted for BIM and DTs in the overall digitalisation process [91], autonomous vehicles and transportation operations [92], transportation safety, mobility, and environmental impact [93]. Furthermore, the very recent reviews still focus on applications of DTs for transportation infrastructure management [20], and evaluate general levels of adoption of all transport sectors [94].

For road pavement, in particular, a review focused entirely on data acquisition systems and only for bridge roads [95], general DT applications in road pavement with no framework provided [96], and pavement sensing technologies for DT development focused on sensing methods, including the asset construction phase, with no technical framework provided [25]. Reviews on the transportation asset DTs are summarised in

Table 1. Recent reviews on DTs in transportation engineering.

Work	Year	Review Focus
[96]	2024	Explore DT applications in road pavement management.
[93]	2024	Transportation DTs for safety and mobility applications and architecture for a transportation DT (TDT).
[20]	2024	Overall Digital Twins of transportation infrastructure management.
[92]	2024	DT-ITS core considerations, progress, services development of DT-ITS concerning architectures.
[94]	2024	Transportation Operation and Maintenance adoption and general asset management framework.
[97]	2023	BIM applications in civil engineering using new Information and Communication Technologies (ICTs).
[24]	2023	Transportation infrastructure management DT concept definition, life cycle application, and technology.
[98]	2022	Predictive maintenance transportation motor vehicles.
[26]	2022	DT rail and road infrastructure networks.
[91]	2022	BIM and digital twins in the digitalisation of transportation.
[95]	2022	Digital twins of road bridges inspection.
[90]	2022	Underground infrastructure construction and Operation & Maintenance (O&M) on locating and mapping technology.
[89]	2020	General DT applications for maintenance.

below.

Nevertheless, none of these works explored the structure and requirements of DT architecture, particularly for flexible pavement assets. Although one presented a framework [94], it remains a broad framework focused on general transportation assets, particularly regarding fleet and logistics aspects. Moreover, the work evaluated road transport DTs, focusing on applications such as autonomous vehicles, simulating vehicle operations, and operation safety. Furthermore, the developed framework is broad and generic. Therefore, it is unsuitable for pavement asset use, especially given flexible pavement assets’ complex nature and management requirements. Thus, relying on this can pose challenges in implementation if asset-specific aspects are overlooked. Potential critical oversights and effects in pavement context are analysed in

Table 2. Flexible pavement critical aspects and potential challenges of the generalised framework adoption.

Aspect	General Framework Possible Assumptions	Asphalt Pavement Asset Criticality	Challenge/Effect
Modelling material behaviour	Not available (analytical only)/Elastic assumptions	Account for asphalt responses (i.e., viscoelastic/plastic response, time &temperature dependency, stress dependency response)	Unrealistic behaviour, inaccurate deterioration, overlooking temperature/load type or pattern)
Lifecycle/Simulation data input	Constant loads of operational/environmental data	Sensor data preparation (time-based loading, distributed temperature, overloading, etc.)	Inaccurate distress prediction (i.e., deformation), degradation rates, scenario analysis (i.e., event-impact scenario)
Integrated predictive model techniques	One modelling approach	Combine underlying physics of asset material response and facilitate	Enhanced accuracy and less computational cost allow faster and long-term predictions.
Continuous model updating and validation	One-off model development and training	Periodic data-driven update (simulation progress, survey data, sensor response for validation)	Updated knowledge of deterioration mechanisms, long-term prediction
Sensor data	Place on critical points/represent the structure	Rely on physics-based virtual sensors, validation study sections, incorporate users/mobile feeder	Unapplicable due to linear horizontal nature/network level

. A DT framework for practical adoption remains underdeveloped.

Current works neglect DT structure and design specific to asset needs for a comprehensive framework. Additionally, no enhanced cognitive DT framework or the potential of new construction 5.0 advancements have been proposed. Literature shows DT applications and potentials without attention to demonstrating their value for specific assets. This gap hinders pavement management in adopting DTs.

Therefore, in assessing the current proposed DT works for road pavement throughout the DT concept and structure, core considerations will uncover the current state of knowledge, outline the requirement for the actual implementation of cognitive pavement DT, and outline a framework that integrates possible solutions. This will also identify current gaps and direct future research. This research raised the following questions:

• What are the components and critical considerations of digital twins (DTs) in the pavement assets; their current maturity, and the gaps for cognitive DT capabilities?
• What are the required framework elements, processes, and methods for developing predictive cognitive DT for pavement that enable enhanced decision-making and support the Construction 5.0 paradigm?

The main objective of this research is to review the concept of DT in other industries, where it originated, and examine the literature on pavement asset DTs. This will clarify the current conceptual ambiguities of DT and the core requirements in this asset context. This research leverages the existing literature on DTs related to pavements. It discusses the context of flexible road pavements, mainly focusing on DT definition, the interconnected relationships among its required components, and the considerations needed for pavement assets. It will assess the developed pavement DTs to establish the foundation for the advanced adoption of cognitive DTs. Subsequently, this study will propose a holistic framework, outline the necessary considerations of potential decision systems for cognitive level DTs, propose conceptual collaborative DTs for the Construction 5.0 vision, and identify gaps and future needs for its comprehensive implementation.

1.4. Research Structure

The structure of this paper is divided into three sections. The first section presents the introduction, pavement management and gap, and methodology. Section 2 includes key features and considerations of the DT concept and discusses and analyses the developed pavement DT works. Section 3 addresses key points of cognitive pavement DT requirements, proposes a DT framework, highlights potential solutions and challenges, and summarises research gaps and future directions. The overall paper design is illustrated in Figure 3.

1.5. Methodology

This narrative systematic review presents core research questions and specific objectives. The methodology leveraged multiple databases such as Scopus and Google Scholar to extract articles related to the subject areas. Additionally, this work focused on good-quality peer-reviewed articles, conferences, seminars, and books. The search was performed in the databases between January and November 2024. The review encompasses various papers in the last ten years that introduced and explored the core concept and implementation of DTs across different industries, where the DT concept initially developed. This broader perspective helped establish a foundational understanding of DTs before narrowing the focus to pavement DTs works.

For the pavement-related DT works search, the fields “title”, “abstract”, and “keywords” were used as “Digital Twin” AND “road pavement” OR “flexible pavement” OR “pavement maintenance”. To avoid literature that focused on transportation operations such as mobility and automated vehicles, generic words alone, such as “road” and “infrastructure”, were not used. The search was limited to works done between 2018 and November 2024, as DT applications in the civil and construction industry are relatively recent, and no DT or works related to road pavement were found in prior time.

The resulting papers passed a screening process considering the function of their titles. This was followed by an assessment of the abstract to filter works beyond the focus of this paper. The introduction section was reviewed for those not reflecting a clear scope in their work titles or abstracts. In this process, the focus of the filtering stage considered proposed DT works for road pavement, particularly flexible pavement assets only. In addition, this included works developed for various asset managing aspects such as data collection and sensing, evaluation, modelling, or decision-making, while proposed as DT-based. This allowed for a focused analysis of relevant information that addressed the research needs and aligned with the research objectives. The initial search resulted in 28 papers on DT for pavement management articles. The 28 papers were filtered out based on review criteria into 22 final papers, shown in Figure 4. The selection methodology is presented in Figure 5.

2. Digital Twins and Pavement Management Digital Twins

2.1. Digital Twins Concept

The DT concept was established in 2003 when the base for NASA product lifecycle management (PLM) was introduced in terms of virtual and physical spaces [86,99]. Afterwards, NASA, in 2010, published a technology roadmap [100], which provided a technical definition and revealed its key features [101]. Industries such as aerospace and manufacturing have adopted the concept and contributed to the design of the framework of DT in their sector. Therefore, the applications originated and were designed to meet their sector’s standards and needs.

Different definitions were introduced of DTs. DTs are complex simulations built upon historical and real-time data designed to replicate the condition of a physical object [102]. They are also defined as a representation of a physical product that utilises data from the physical element or system to mimic real-world behaviour in the corresponding virtual counterpart [103]. The virtual space is reinforced by modelling and simulation tools, sensor data, and databases to enable more integrated and remote management. Another common definition indicates that DTs incorporate various physics, probabilistic analyses, and scales [2]. In most definitions, the simulations revealed the ability to analyse what-if scenarios of asset response under specific conditions. This is enabled by the connection and information exchange between components [101]. The benefit of a twin is achieved using actual operational data and integration of multiple data sources, as well as data exchange between the digital and physical spaces. The conceptual DT concept is illustrated in Figure 6.

2.2. Civil Assets DT Core Considerations

DT core characteristics include the physical product lifecycle data of the asset, synchronisation of data [104], types of data connections between physical and virtual spaces (unidirectional or bidirectional flow) [105], visualisation, use of sensor data, and advanced analytical or simulation techniques [106]. In the context of construction industry assets, Oditallah et al. [74] defined the DT as “a virtual copy of a physical asset, process, or system of assets that comprise the twin purpose-related lifecycle data, reasonable data exchange and synchronisation, analytical or simulation models, and intelligent techniques to model real-world conditions and responses of the physical entity”. This definition highlights the adoption of DT characteristics that fit the purpose of twinning. It emphasises purpose-related lifecycle data, applicable data exchange and synchronisation, and modelling real-world data or conditions. The key characteristics of DTs are summarised in Figure 7.

2.2.1. Connectivity

DT requires a data link to connect its digital and physical counterparts, facilitating data transfer between entities. The nature and frequency of connectivity depend on the specific application or service. A DT utilises the best available data, sensor readings, and other pertinent information to replicate its physical counterpart accurately. Connectivity encompasses the data transfer rate (e.g., real-time, near real-time, or periodic) and the connection type between the linked environments, whether unidirectional or bidirectional. Connectivity requirements need to align with data updates to specific objectives of the developed twin [26]. For instance, in autonomous vehicle operations, DTs require real-time, bidirectional connectivity to provide immediate feedback, influencing vehicle performance in response to environmental and operational changes. In contrast, for pavement health monitoring, real-time data and bidirectional connectivity are not as important. In this context, DTs primarily evaluate asset conditions and provide maintenance recommendations based on periodic assessments, rather than real-time interventions. Therefore, the emphasis is on long-term asset management and its responses instead of immediate operational changes, with actual data from the physical asset being more significant than real-time updates.

Bidirectional connectivity allows virtual actions, such as activating or deactivating physical components, to impact the physical asset. This is typically in mechanical or electrical systems such as production lines, where actuators can be controlled. For passive or non-mechanical assets, as in structural elements, the existence of bidirectional connections is often unnecessary. However, the feedback from the virtual space to the physical system may involve rescheduling maintenance or triggering manual interventions. Feedback and decision-making can affect the physical environment by either automated decisions or manual responses [107].

2.2.2. Modelling

Modelling and simulation are fundamental elements of DT systems. It can enable detailed visualisations, what-if scenario analysis, and solution validation. As a unique feature of DTs, accurate and continuously growing prediction models form the foundation for enhanced decision-making. Analytical modelling encompasses statistical models, including ML models, physics-based numerical models, and other intelligent techniques. However, the value lies in integrating multiple approaches and diverse data sources, particularly operational data from the physical asset and lifecycle data. Hybrid modelling, combining both approaches, is often used to leverage the advantages of each. Physics-based modelling can capture materials or internal structures in complex conditions to provide high predictive accuracy.

2.2.3. Data

Data representing the whole lifecycle of an asset are essential for linking assets and their operational activities for enhanced decision-making [108]. To support the DT’s objectives, data such as geometrical and material properties related to the DT process can be incorporated [109]. In addition, operational and lifecycle data can be used for various processes within the DT. Material data can be used to model and understand the behaviour of asset response and affect informed decisions such as treatment type selection. Data on asset conditions, historical interventions, and design can also validate predicted conditions in the DT, and, furthermore, establish relationships between status change rates and operational data.

A DT typically consists of several data subsystems, such as asset inspection data, to process and reflect the asset’s condition over time. Sensor data are used directly for monitoring or modelled to derive additional parameters for modelling, simulation, and visualisation. For example, raw vibration data may be used to derive internal forces that are more relevant for DT prediction purposes. The use and management of such data must align with the designed DT’s functions.

2.2.4. Interpretation, Services, and Feedback

DTs can interpret the results in the context of the service required. This could include alerts and notifications triggered by predefined thresholds of predicted state. Current condition assessments using collected raw data or processed outputs can also be considered in the provided service. Decision support systems (DSS) can analyse the outcomes of related DT components and trigger recommendations based on monitored or predicted parameters. Some DSS and advanced knowledge-based systems incorporate advanced ML models or reinforcement learning to optimise planning decisions [110,111]. However, integrating DSS creates an intelligent layer and elevates the cognitive level of the DT framework. This enhances reliability and efficiency for managed assets and enables collaboration between machines and operators, reinforcing the vision of Construction 5.0 and elevating the maturity toward a truly cognitive DT. General Electric highlighted that a knowledge base system component incorporates data analysis and domain expertise and optimises DTs with industry best practices.

The visualisation rate of state updates is directly reflected by the data transfer rate and modelling or analysis practices. On the other hand, the required type and frequency of visualisations depend on the application’s goal and purpose, as well as the needed system interactivity. Visualisation plays a role in data and output of modelling and simulation, such as investigating trends, identifying potential issues, and evaluating results in various scenarios. This is useful in complex systems where real-world activities are modelled over shorter timeframes. Real-time visualisation is essential when immediate responses are required. For instance, in traffic management or autonomous vehicle systems, high-frequency data updates require real-time responses and instant decisions. In contrast, health monitoring for pavement shows gradual increment changes and does not require continuous real-time updates. The periodic assessments can be adequate, allowing slower modelling and visualisation rates for long-term assessment systems. However, accuracy in visualising asset degradation can be more critical in this case. Precise feedback obtained from DTs not only benefits its performance but also reinforces the asset lifecycle outcomes for future asset development.

Overall, as connectivity determines the type of connection and the data rate, the visualisation and modelling requirements, along with feedback mechanisms, influence the connectivity design. Considering the core function and dependency in the DT structure. The role, relationships and criticality between its components are essential. The required service and the DT’s primary purpose form the DT pyramid’s foundation. This determines the needed modelling and simulation types, along with the necessary connectivity and visualisation. This structure specifies what data will be generated, how they will be utilised for the intended purpose, and how DT outcomes will be interpreted and integrated. Thus, Figure 8, shown below, describes the DT’s base, which guides the dependency of other functional components.

2.3. Road Pavement DTs Discussion

The research on pavement DTs focuses on sensors for automating surface data collection and analysing defects. A group of works integrates these data into static 3D models using BIM and other spatial visualisations, while others apply real-time sensor data for monitoring and decision-making. Additionally, some studies aim to enhance virtual asset presentations and a few focus on updating 3D models to reflect surface condition changes for improved asset representation.

2.3.1. Pavement Surface Prediction DTs

Yu et al. [112] proposed DT to predict the highway pavement performance and improve its current preventive maintenance. Sensors were used to collect pavement surface current conditions. The prediction model learning process used ANN, random forest (RF), ridge regression (RR), and support vector regression (SVR). This involved utilising asset-related data, such as maintenance data and traffic flow. Data were visualised spatially using a BIM 3D model based on predicted data for support planning maintenance.

Similarly, Consilvio et al. [113] provided architecture for the pavement DT, in which computer vision-based algorithms were utilised to evaluate pavement condition collection. Both Consilvio and Yu [112] involved other existing asset data, such as maintenance records, performance indicators, and BIM model visualising for decision support. Consilvio used an AI-based clustering method to filter road sections based on the severity of their condition. However, these works overlooked the accounting for the structural aspects of pavement or real operational data contribution, and the focus was confined to surface conditions only. Although ML predictive capabilities to forecast health metrics were used, this use has been long-standing; for instance, Marcelino et al. [60] employed a Boosting-based learning algorithm (TrAdaBoost), while other work compared five ML techniques [114]. Consequently, DT works could not address the added value of DT over the traditional ML models used. Nevertheless, reflecting predicted performance indices on a static 3D model does not constitute an entire DT. DTs should offer the expected enhanced precision and more cognitive service than traditional ones.

As a vision for futuristic road performance modelling, a proposed conceptual framework for DTs is presented by Kaliske et al. [115]. Their work proposed interactions of roads, tyres and vehicles as innovative technologies for vehicle–tyre interaction. The work introduced a potential contribution for tyre–road sensor data in monitoring road damage or changes in friction conditions. Although the work highlights the physics-based simulation data, their conceptual work proposal focused on the vehicle–tyre simulation perspectives only and road asset was not discussed. Ficara et al. [116] deployed a system that relies on edge cloud services to detect pavement anomalies. The work used statistical analysis based on the surface conditions and collected data from the crew and road users’ vehicles. Road user video data were processed based on AI screening. The work suggests improvement in planning the manual collection of data based on the road user’s preliminary data.

2.3.2. Sensing and Monitoring DTs

Barisic et al. [117] proposed a DT focusing solely on the thermal state of asphalt pavement. Specifically, temperature data were collected over the years using several instrumented highway sections and modelled for continuous condition monitoring. Another study used sensor data to propose a strain monitoring framework based on strip sensors under vehicle loads, which is used to assess the developed road pavement distress [118]. Data of loading variation were tested to assist, validate, and predict the measured strain scenarios using ML models. These works considered limited sensor data, neglecting other factors that involve pavement performance. Steyn et al. [119] used sensors for environmental monitoring data, asset imagery data, and camera data for vehicle counting. Point cloud data were derived from images of the demonstrated local road DT, in addition to road surface texture and temperature. The authors reported that acquiring the environmental parameters for monitoring supports management in the maintenance decisions of the roads. However, the role of collected real-time data and the use of environmental data were not discussed in their work. In addition, the work focused on collecting sensor data and no prediction techniques were used. Similarly, Meža et al. [120] proposed a DT of pavement for modified material structure and their work used point cloud scan data for road pavement model generation and visualisation. Data from temperature and moisture sensors, pressure pads, strain and deformation sensors were all mapped in a digital BIM model using the Common Data Environment (CDE) platform. The work reported that integrating the asset model with sensor data provided insight into the overall structural health of the monitored road. However, the data were mapped into the digital model for monitoring, and no modelling to predict structural measures was presented.

Digital Twin Box is a group of tools that was deployed in work to introduce DT for monitoring [121]. It comprises a GPS device, 360 camera, and IoT sensors for collecting data such as humidity and temperature. The work collected video data for the identification and tracking of objects in the road environment. The work considered DT development by linking real-time data to accessible cloud services for monitoring services. As in the previous works [117,118,119,120], these efforts focused on integrating sensor data into digital models as a support decision tool. Although sensor data integration is essential in DT, this part alone cannot comprehensively reflect the entire physical asset state. These works have not shown how these data contributed to the developed system. Instead, they solely presented IoT data on static digital models rather than establishing predictive analyses of asset deterioration based on actual leveraged data.

2.3.3. Surface Defect Detection, Visualisation, and As-Built Digitisation

For data collection automation, ML was frequently implemented for defect detection. Wang et al. [122] used imagery data for automated detection of distress and to assess the current condition of pavements. In their work, the ML training process involved multiple integrated data to build an enhanced detection model, which was described as a DT. Similarly, Sierra et al. [123] presented a road pavement twin based on the reality capture model. UAV images were used to detect current irregular surfaces and damages to the pavement and to detect surface defects in the captured model. In another work, laser scans were used to collect point cloud data and visualise the current pavement surface [124]. The collected 3D point cloud data were processed into a model and then used for surface-level flatness analysis and detection of defects. Collecting the current condition for pavement management is a promising aspect for feeding DT models, as in [122,123,124]. However, these data reflect the point-of-time conditions and are limited to surface status only. Therefore, it has contributed to automated defect classification. Its novelty lies in predicting current situations of surface conditions rather than predicting future asset status.

To manage the collected data for decision-making, a work implemented a DT for pavement based on BIM and GIS [125]. In their work, pavement surface data were collected using sensors, and the data were converted into a performance index. Road sections were modelled in BIM and georeferenced to GIS data. Similarly, Bosurgi et al. [126] used a BIM platform to manage the condition and quality information of pavement surveys. Another work proposed a framework for PMS in a systematic workflow containing data acquisition to support decision-making [72]. The framework shown in Figure 9 outlines automated road distress detection and 3D distress quantification by embedding data information into GIS. This work also aimed to use defect data on spatial platforms for collaboration as in previously mentioned works. Road agencies usually adopt spatial presentation and GIS for enhanced data management. These mentioned works proposed high-level frameworks that lack enhanced or advanced modelling and simulation for what-if scenarios.

Other works focused on the modelling of existing pavement assets. These works aimed to detect defects using ML, create 3D models, and reflect the conditions detected in the developed model. For instance, Cao et al. [127] developed an interactive system to present pavement cracks in 3D visualisation, where the crack boundary is detected and extracted based on a 3D crack edge feature algorithm. Data were acquired based on camera and laser scan sensors. The twin here is presented to replicate the existing shape and type of defects on the 3D model. This allowed the visualisation of pavement surface defects to reflect the conditions of road assets and help in decision-making and maintenance planning. Wang et al. [122] constructed a 3D model based on field-captured reality data and used deep-object detection algorithms for pavement distress detection. The authors integrated data by leveraging a lightweight engine used to graphically represent five distresses. Their proposed framework is shown in Figure 10. Cao’s interactive system [127], along with this work, are promising enablers for visualisation aspects in DT, presenting updated visualisation capability of updated site data; however, it is not an entire DT system.

Furthermore, D’Amico et al. [128] used the existing road pavement condition database to develop a BIM model for that asset. The model was replicated based on two surveys carried out in their study. The collected data included point clouds, GPR scans and GPS records. The author used parametric elements in BIM that can adapt their model features to the information collected for road conditions. The processed data were integrated and adapted as 3D polylines extracted from the surveys, as shown in Figure 11. The proposed method used data to visualise the current road surface conditions. Bertolini et al. [129] introduced an approach that involves scanning the pavement surface to identify irregularities that may correspond to specific types of pavement distress. GPR surveys were utilised to generate multiple datasets, aiding in the detection of subsurface pavement failures. They followed a method based on a grid system for surface-level defects and 3D voxels for the subsurface. Although no prediction of the future status of pavement assets was achieved in these works [128,129], it is a promising method for visualising functionalities in DT. As DT is supposed to replicate the defect on the 3D model to reflect its status and future status, these approaches potentially benefit DT.

Other works on pavement DTs focused on creating the digital model of pavement assets. This also allowed the building of the pavement model surface, including its deformations and irregularities in terms of pavement layers. Forming the 3D model of an existing road can establish a DT base. This can indeed, along with previously mentioned works, be useful for visualisation in DTs. For instance, Fox-Ivey et al. [130] proposed a DT model using a 3D scanning and positioning system for inspection and damage detection. To develop a digital model of an existing asset, Jiang et al. [131] mapped collected data to build a digital surface model (DSM). Then, fitting processes for horizontal and vertical alignment and cross-section generation were implemented (Figure 12). Similarly, Pan et al. [132] utilised a 3D point cloud segmentation process to acquire semantic information using 3D deep learning models. The process involved clustering the point cloud and alignments fitting based on polynomial approximation. The points extracted and segmentation of road surfaces were converted to detailed components using two effective deep learning models, KPConv and Superpoint Transformer. The fitted curves were then used to separate road surface points into lanes, shoulders, and central medians. The authors reported that the work offers the base model of physical assets, which potentially can be expanded to complete DTs. An illustration of the process is shown in Figure 13.

2.3.4. Developed Pavement DTs Summary

The discussed works have used various tools and approaches in their studies.

Table 3. Developed DT methodologies.

Work	Methodologies
	Data Acquisition	Model Type/Development/Approaches	Platform/Integration Strategies/Tools	Interaction Type/Output/Feedback
[112]	Vehicle-line scan, Accelerometer	ML-Prediction: multi-ANN, XGBoost, etc.	BIM software (Autodesk Revit)/CDE, Visual programming	3D model-Roughness IRI index visualisation
[117]	Set-temperature sensors (profile), weather station sensors	Statistical prediction, numerical analysis/ validation	Analytical/numerical software	Real-time temperature prediction
[118]	Strain sensor	Physics-based model	Numerical analysis software (LS-DYNA)	Strain, load prediction
[119]	Smartphone, UAV camera, CVV camera, infrared, LiDAR, temperature	Depth map modelling, multi-view stereo (MVSNet)	Cloud IoT platform	3D model/real-time data streaming
[120]	Embedded (temperature, humidity, displacement, pressure, stress, strain)	3D BIM model	Centralised data management, visualisation and collaboration platform	3D model/real-time data streaming
[121]	Camera/Video, GPS, Gyroscope	Object detection model (SSD MobileNet V2)	Open-source cloud computing	Real-time data streaming, object detection
[127]	LiDAR, Camera	3D Object crack segmentation (CNN), 3D feature extraction	Python-ML platform (TensorFlow software)	As-is 3D crack visualisation
[122]	UAV-Camera	Object Detection (cracks, potholes) based on (YOLOv5)	Photogrammetry/mapping software, Game engine, real-time 3D platform	As-is distress visualisation
[123]	UAV-Camera	Object segmentation crack (U-Net, VGG-16)	Reality modelling (ContextCapture software)	Distress detection
[124]	3D LiDAR, GIS	Point cloud processing, parametric modelling	Bridge/road software (Civil 3D), plugin-integrated visualisation (Leica’s toolset for Revit), 3D Surveying software (Cyclone 3DR)	As-built 3D model/virtual assessment (flatness or distortion)
[125]	Camera, GPS camera	Spatial data analytics	BIM modelling, GIS software	3D model GIS integration
[126]	NA	3D object -inspection, lifecycle data integration	BIM Infra software	3D Model-based data
[116]	Smartphone sensors/camera/GPS, edge computing	NA	Cloud-based storage/services/query platform	Condition indicators visualisation
[72]	Camera, LiDAR, GIS	Image processing, data analysis	GIS, point cloud computing software	Distress detection, quality index visualisation
[113]	UAV camera, GPS, Light Detection and Ranging, survey data	Parametric modelling, computer vision-based algorithms, AI clustering decision optimisation	Cloud computing platform. Game engine, real-time 3D platform	Mapped/visualised data support decision system
[128]	GPR, Light Detection and Ranging, Laser Profiler	3D data processing, Parametric modelling	Design/visualisation software (civil, infra)	As-is 3D pavement model/distress visualisation
[129]	Equip vehicle-laser Measuring System, GPR surveys	Point cloud processing/reasoning/Cells and voxels classification, visualisation	Specialised point cloud computing software	As-is Surface reproducing/subsurface distress visualisation
[130]	GPS/GNSS, 3D scan-integrated inertial measurement units (IMUs)-	Geo-referencing, 3D points cloud processing	LDTM software, Cad viewer	As-is 3D surface modelling
[131]	Existing Map database, aerial images, derived point cloud	3D data processing	3D modelling, point cloud processing software, visual programming	3D model element reconstruction/digitising
[132]	3D Point cloud data	Deep learning-segmentation, elements fitting	Point cloud processing software, Algorithm-Python	3D Design element reconstruction/digitising

summarises and presents the data acquisition methods, approaches used, platforms or integration strategies utilised, and types of output feedback.

The analysed literature reveals that current implementations often fail to fully adopt the DT concept. This is seen in the limited inclusion of in-service behaviour of modelled assets and the absence of enhanced predictive techniques over already existing ones. This hinders the realisation of DT’s full benefits. The proposed frameworks mainly focus on the automated detection and extraction of defects, considering the functional aspects based on surface conditions. Although the proposed work is not the entire DT, it has valuable contributions toward its full development.

The literature has different viewpoints proposing the conceptual structure of DTs for pavement management. Nevertheless, the predictive and analytical capabilities of the DT concept remain key components. This includes advanced decision support system (DSS) models, which have not been fully integrated into the DT framework to propose cognitive and reliable decisions and continuously learn and share knowledge for asset management. The developed DT works in the literature have been classified into different categories based on their focus and contribution area in

Table 4. Developed DT’s focus and contribution areas.

Work	Category
	Current Condition Detection	3D Model Creation/Visualisation	Support Monitoring Data	Analytical Prediction/Decision Support
[116]				X
[112]	X
[72]				X
[128]	X	X
[115]			X
[117]
[124]		X	X
[119]			X
[113]	X			X
[122]	X
[127]	X	X
[131]		X
[122]	X	X
[130]		X
[123]	X
[120]		X	X
[126]				X
[118]			X
[125]				X
[129]	X	X

X denotes that the feature is applicable to the corresponding work.

. In this table, ’X’ is used to indicate which focus category is associated with each work.

However, cognitive DT implementation is intended not only to facilitate services related to current and future asset status assessment; it further optimises the accuracy of the overall reflected performance. Figure 14 illustrates the expected services of Pavement DT.

3. Advancing Toward Cognitive Pavement DT

The framework design aims to bridge the asset needs from an operational point of view to provide a starting point for effective pavement DT development. The following discussion includes considering actual operational conditions, integrating multi-modelling methods, and defining visualising aspects in the context of DT structure and flexible pavement needs. This would advance the developed framework toward full DT implementation for construction 5.0 environments.

3.1. Operational Data, Asset Databases and Performance Modelling

Actual operational data and incorporated material modelling in DT are key for enhancing the prediction of pavement actual performance. Service condition data, such as temperature and operational data, i.e., load frequencies, speed, and overloading, are essential in performance degradation. However, sensors and weather station data can be leveraged and calibrated based on formulas and models, such as models developed in [133,134,135]. This will allow DT to account for in-service assets’ actual temperature and moisture levels. Moreover, other existing pavement lifecycle data can be acquired using asset databases, such as LTPP, which can facilitate actual data for analysis and prediction. The existing databases, i.e., PMS data, involve the asset structure data, i.e., layer thicknesses and material test data under various environmental conditions. This would yield insightful contributions to the enhanced prediction of DTs. Regarding operational loading data, Weigh-In-Motion (WIM) data, which is sensor data, classifies stream traffic loads and presents the weight of vehicles, loading speed, actual repetition of each type, and other detailed traffic. These data can be modelled in time series form and leveraged to present specific site operations characteristics. When sensors are absent, they can also undergo metamodeling approaches to estimate other network pavement loads. This connection between the actual site environment and the operational scenario in the virtual space can enable predictive DT. Sensor systems can rely on WIM or similar camera-based systems and embedded sensors for temperature or leveraging weather station data within the asset zone. These data can be employed in DT modelling, prediction, and decision selection processes.

Other data, such as existing maintenance records and detected surface condition data, can be integrated into DT using data inquiry and management methods, i.e., SQL databases. Surface condition automated methods, i.e., cameras and laser sensors, can be used for frequent inspections, where acquired data can be classified based on AI detection and stored for further prediction operations. Recent works proposed systems adaptable to vehicles to be used as collectors. These potentials are supported by edge computing technologies that use cloud models and access various data and computing resources. It is possible that future integration of autonomous vehicles can help to capture road surface data over time and analyse them based on edge computing. This includes sensing technologies developed that can be leveraged for structural testing through advanced non-distractive methods. These data, i.e., deflection responses, can be acquired from road vehicle users to indicate the structural health of pavement and used in calibration and validation processes in the DT modelling component.

One of DT’s challenges is determining data updating and connectivity aspects. The structural health of infrastructure assets changes slowly. Data collection is periodically based on the rate of deterioration and frequency of decisions needed. Therefore, periodical use of the continuously collected data can be used to update the periodical modelling. This can be done by segregating and redistributing the newly available operational data to fit the modelling frequency.

In terms of predictive modelling, data-driven models within DT, such as ML models, can predict pavement temperature at various depths using other environmental factors and weather sensors. Moreover, at the pavement performance modelling stage, data such as traffic, axle load, etc., which are collected from WIM and integrated asset databases, require data modelling in DT functionalities. For instance, predicting traffic flow at a given time of the day or under extraordinary event changes for what-if scenario analyses. For performance prediction, the physics-based models can contribute to forming enhanced hybrid models such as surrogate models. Physics-based models can incorporate material characteristics for response prediction. This considers operational and environmental conditions leveraged from the physical pavement to forecast degradation based on the physics law of the pavement, i.e., viscoelastic response. Continuous prediction results can be incorporated into ML models to facilitate faster and generalised predictions of pavement for decisions. This overall process involves the lifecycle data for accurate prediction and predictive DT. Furthermore, NDT devices are used to gather structural performance data and can be adapted to validate and optimise the performance models in DT. For instance, various works, as in [35,136,137,138,139,140,141], used ML to establish a model based on simulation data for structural analysis using NDT data and optimise its performance. These surrogate models open a window for further innovative modelling in DTs.

3.2. Intelligent Decision-Making Support Systems

Enhanced performance prediction in the DT approach allows a precise decision-making process. Furthermore, integrated DSS can automate plan prediction based on advanced knowledge models and optimise decisions based on incorporated industry best practices. This decision model can leverage performance modelling outputs, other system and asset inputs, asset operation and maintenance policies, and best practice considerations to inform optimised decisions. Its cognitive feature can adopt budget limitations, intervention factors, treatment costs, etc., [142]. This is essential to close the DT feedback loop and achieve a cognitive level. ML models in DSS can guide, support and optimise decisions and elevate the practice to a prescriptive maintenance approach [143], which predicts and prescribes maintenance actions. For instance, reinforcement learning is frequently used to optimise planning decisions [110,111,144]. This allows the cognitive DT to be optimised for continuous learning over time. In this context, various works developed decision models: Georgios et al. [145] used a decision tree to provide repair strategy, defect cause and treatment based on distress type, severity, and budget; Philip and AlJassmi [146] adopted Bayesian Belief Networks (BBN) to produce optimal sustainable decisions; and Abu Dabous et al. [147] proposed multi-criteria decision analysis (MCDA) based on multi-quantitative evaluation, while another work adopted a risk assessment BBN model trained to expand its knowledge domain over time [148].

Considering Construction 5.0 principles and focus, DSS integration is a significant step in aligning the future cognitive DTs with its goals. In addition to advancing DT maturity, it combines automation with skilled human oversight. This can enhance the intelligence, connectivity, and collaborative aspects for the next evolution in construction. DSS in a DT can interpret and present complex data in a more accessible way to support teamwork in digital systems for optimised outcomes. Furthermore, the continuous learning ability of reinforcement learning helps knowledge-based systems adapt. These systems can plan, facilitate and automate routine decisions, which allows expert humans to focus more on critical and strategic decisions. In addition, the recent generative AI has a promising role in enabling workers to gain real-time insights. This facilitates complex DT functionalities and translates insights into plain language, bringing human creativity and machine cognition closer.

3.3. Visualisation & Data Integration

The DT visualisation interface can help to clearly understand the actual state as visual information about defects. This plays a fundamental role in the route causing the deterioration mechanisms. Visualisation in the pavement asset context can be divided into macro and micro levels. The macro level refers to the visualised conditions in terms of performance indices. This data visualisation helps spatial-based investigation for planning interventions. Micro visualisation can involve pavement structure geometry changes, surface, and appearance degradation, reflecting the nature and severity of defects. It can be used to determine deterioration mechanisms and their interaction to support decisions. Although some developed DT works presented in the previous discussions are promising methods for micro-scale visualisation, further visualisation methods for DTs are required.

Operable platforms and tools that handle Application Programming Interfaces (API) can facilitate data exchange between different systems, leveraging various data sources and model results among multiple applications and platforms, and GIS adds a spatial context to asset visualisation. BIM models also provide 3D adaptable models rich with information. Simulation model visualisation is also essential for generating and updating visualised assets. Designed DT could involve multiple types of visualisations for short and long-term decisions.

3.4. DT Platforms

Industrial platforms are enablers for DTs and serve as ecosystems for various industries’ applications. Examples include the Siemens Xcelerator Platform, Microsoft Azure DTs, and IBM DT Exchange. These cloud-based platforms provide tools to create comprehensive, integrated digital environments. However, other DT-enabled platforms were explicitly developed for the construction industry, such as platforms or CDEs. For instance, Bentley’s iTwin platform supports reality capture to create a digital context and integrates iTwin IoT for analysing sensor data. Autodesk Tandem focuses on the construction and operational phases of assets. Similarly, Trimble Connect CDE provides DT capabilities by integrating with field data and offering real-time project analytics.

These platforms are centralised environments to store and manage all relevant asset data. This includes the development of dashboard visualisation, date information, cloud-based simulations, predictive models, and other data sources. These technologies and processes enable DT to control the integration of multi-modelling and analytics approaches based on various cloud-based services and data storage throughout APIs. Thus, a well-designed visualisation helps understand current and projected asset conditions. Furthermore, it allows integrated workflows with GIS and DSS systems to ensure comprehensive decision-making.

3.5. Integrated & Collaborative DTs

An integrated collaborative DT is a system of interconnected DTs communicating and sharing data to support more comprehensive decisions. In this concept, each DT holds its functions and purposes, such as pavement management DT, traffic management DT, road users service DT, and autonomous vehicle DT. However, they also share essential data across the network as needed. When requested through data protocols, data from one twin can directly benefit or inform the processes of another. For instance, a DT of road traffic can share data about traffic flow and density, helping pavement DT with periodical asset assessment. This includes transferring data from one twin, i.e., traffic DT, for what-if analysis in pavement DT. Similarly, pavement DT can feed road traffic DT with any predicted hazards that could affect the traffic in the future, helping adjust routes based on current or future conditions. Furthermore, an autonomous vehicle twin system can provide pavement DT with surface-related data, tyre contact data, and defect location. Combining such structural data-sharing and protocols in a collaborative environment can be systematically achieved. However, sharing data protocols and security issues can be an issue in its adoption. Figure 15 illustrates conceptual data-sharing actions between several DT systems, and Figure 16 illustrates pavement DT within an integrated DT collaborative environment.

The proposed federated DT system, Twins Chain, a possible term, can also improve resource efficiency by reducing the redundancy of duplicating data collection or analysis efforts by an isolated DT system. This would save computing power, data storage, and cloud service interruptions. This collaborative approach also aligns with the Construction 5.0 vision for sustainable, intelligent systems for adaptive management of construction resources. It also provides a conceptual step toward standardisation for centric transport DTs, allowing shared access and collaboration across disciplines to realise the concept of smart city DTs.

3.6. Proposed Cognitive Pavement DT Framework

Prior analysis and discussion sections put forward a map considering all required data, possible models, and support components of a cognitive DT for pavement management. This outlines the most effective architecture framework for road pavement based on DT structure and pavement-developed systems. The conceptual framework in Figure 17 consists of a skeleton equipped with different tools and technologies.

The system utilises data from multiple sources, including sensors on user vehicles, existing databases, and surveying technologies, to collect physical asset condition data. It employs cloud-edge services for data detection, analysis, and classification, which are used in different functional units within the proposed system. The framework also incorporates data preparation and management processes within dedicated data collection and management layers.

The proposed DT framework integrates multisource data, including processed data on operational, environmental, and surface conditions. A processing unit organises operational data into time-series sets (e.g., traffic load, asset temperature, and material properties) for physics-based and ML model predictions. These models use a multi-modelling approach with surrogate models, enabling advanced analysis, what-if scenarios, and visualisation. This framework can potentially bridge the existing gap in the literature concerning discussed works. It incorporates the DT concept’s multi-modelling feature to improve pavement systems’ predictive modelling. This enhancement facilitates superior forecasting and enables more informed decision-making. Thus, it highlights the additional value of the DT approach in comparison to traditional practices.

A DSS combines predicted and existing asset data to inform and optimise planning, incorporating classified surface conditions from processed field data. The framework relies on cloud-based dashboards, CDEs, or DT platforms for seamless integration and user interaction. This dynamic, cognitive DT evolves by updating surrogate models and DSS knowledge. This can be leveraged based on reinforced learning DSS, improving predictions as new data become available, and enhancing decisions as more properties and practices are provided throughout the asset’s service life. In contrast to the discussed foundational level of previously developed DTs, comprehending processes within the presented pavement framework encompasses all required management components. This includes data collection, detection of current conditions, and an enhanced approach to future state predictions. This framework integrates actual asset operational data and proposes sensing pathways, in addition to intelligent decision-making systems with continuous learning features inherent in the DT concept, allowing prediction models and decision experts to grow for ongoing improvements in DT performance. Such advancements will effectively bridge the current adoption gap and align with future industry 5.0 requirements during implementation.

However, the proposed architecture has potential challenges. Data integration (e.g., multi-sensor data) and managing diverse data types (e.g., sensors, databases, surveys) impose technical difficulties. In addition, the suggested continuous real-time physics-based and ML models require significant processing power. This could demand excessive computational power and cap its implementation. Furthermore, model calibration for ensuring surrogate models and DSS adapt to changes can be a challenge. These, in addition to high digital skills, are required in its full implementation.

Implementing the proposed DT framework in real-world projects will validate model accuracy and potentially refine system integration. It will also identify further practical challenges and assess how well the DT framework adapts to real-world conditions.

3.7. Research Gaps and Future Recommendations

This work comprehensively reviewed, analysed and discussed possible solutions and technologies for a holistic framework of pavement DTs. To advance to practical implementation and operate the proposed frameworks, real-world case implementations to validate and refine system architecture and focus on facilitating the technical challenges are needed in the future. Some of the key research gaps and future research recommendations are as follows:

• To advance the current underdevelopment practice of DTs, implementation should align with the proposed DT concept. The structural aspect of the asset and actual operational conditions need to be incorporated. Furthermore, the predictions and decision processes of pavement DTs need improvements. Simulation is a core feature of the DT fundamentals. However, it is not mentioned in most developed works. Methods for integrating modelling and simulation for advanced analytical modelling and what-if scenario functionalities require further research. Physics-based simulations, such as the finite element method, and other model updating and optimisation techniques outlined in the proposed framework require further investigations. In addition, real case implementation as proof of concept is required for physics-based methods and ML hybrid integration. This considers real asset operational and lifecycle data to overcome current limitations and achieve improved prediction in growing DT. This work recommends further exploring and assessing these hybrid simulation methods incorporating real-life data.
• Visualisation tools are critical in predicting complex systems for user-friendly use, interpreting findings and making informed decisions through intelligent systems. It is a conjunction of all data types, from simulation, modelling, and decisions to asset state visualisation as proposed in the proposed framework. A few works have developed virtual models that replicate assets and visualise current defects and surface conditions, which are proposed in the literature. However, there is a need for methods to reflect further predicted data and conditions for the future state of the asset, in addition to developing visualisation components for users within DT platforms to align with full DT capabilities.
• DT depends on leveraging remote actual data from the physical asset. Recent NDT methods, i.e., [80], used 3D-digital image correlation to the in situ testing based on vehicle tyre load deflections, or a Laser dynamic deflectometer (LDD) to capture the deflection as the vehicle moves [81]. A work presented tyre–pavement interaction to provide data potentially helping road damage analysis [115]. However, these approaches imply collecting actual data and asset responses based on on-road vehicles, as highlighted in the asset response sensor data in the proposed framework. This significant research direction on the future of non-destructive sensing enables the enhanced asset status prediction of future DT implementation.
• Cognitive DT implementation involves a complex and wide range of data parameters and information. This can also challenge the traditional decision-making processes due to their complexity and numerous dependencies. Therefore, optimising these decisions would require advanced self-learning DSS for managing, interpreting, and optimising detailed and complex datasets. Consequently, more sophisticated DSSs are required to facilitate the effective implementation of DTs and adapt for Construction 5.0.
• The reliance of DTs on existing asset data highlights the critical importance of structured databases, such as LTPP. However, this presents other challenges due to the lack of digitised data records in integrating data within DTs. The solution for data operability, security, and existing database use requires developing innovative methods to facilitate their implementation. Therefore, interoperability and data security solutions must be addressed, and cloud service solutions in integrated diverse data for modelling, storing, and visualising need to be explored. This will help move toward a sustainable digital built-in shared environment across multiple sectors.
• In the current literature, there is an absence of works investigating the cost and environmental considerations of DT implementation for asset management. This includes human–machine relations, trust, connection, and sustainability factors within infrastructure asset DTs for Construction 5.0 readiness. In addition, it is essential to understand the factors that influence the adaptation and adoption of DT systems. These required contributions will impact the successful deployment of DTs.

4. Conclusions

This research provides a crucial step in developing a comprehensive DT framework capable of meeting DT requirements, enhancing status prediction, and incorporating growing knowledge DSS to achieve cognitive systems aligned with the Construction 5.0 principles. This paper reviewed the concept of DT structure, outlined critical considerations, and assessed existing DT implementations for pavement assets. The review spanned various aspects, highlighted current gaps, and proposed asset-specific needs and required components with recent advancements that could facilitate DT implementation. The redeveloped cognitive DT framework will foster its implementation in the management of pavement infrastructure, including integration of decision systems and prediction improvement from advanced technologies and profound DT concept realisation to enhance the overall DT efficiency. Research gaps to fully implement the proposed framework were also presented.

The review revealed that the existing literature on pavement DTs lacks a full DT concept presentation, which includes enhanced predictive capabilities, efficient use of actual operational data, intelligent DSSs, and the incorporation of what-if scenario modelling. Most studies emphasise surface condition data with a significant focus on digital models or assessing present conditions rather than extending to fully predictive DTs.

The proposed framework in this work addresses the gap between current practice and the true concept of DTs. This was achieved by addressing the added value of DTs over the previously developed practices. It adopted the full DT concept’s features, incorporating a multi-modelling approach and actual physical asset-handled data to improve predictive modelling, and comprised comprehensive components, including specific asset needs and growing predictive capabilities. This was through advanced DSS incorporation to facilitate data interpretation and growing models to expand knowledge and practices of actions, adding the cognition layer to the current developed DT maturity. In addition, this work proposed collaborative DTs for optimising and standardising smart city-scale DT implementation potentials. This enhances decision-making and optimises planning strategies while aligning with Industry 5.0. This research contributes to the design of DTs for pavement infrastructure, enabling predictive and cognitive systems. While existing systems do not fully address the concept of true DTs, the proposed cognitive DT implementation needs to overcome some identified gaps. However, implementing DTs remains challenging due to their complexity and lack of standardised processes. Moreover, data interoperability related to multi-sensor and source data fusion and security poses significant challenges, particularly for large-scale DT applications.