Computer Vision and Machine Learning Approaches for Defect Detection in 3D-Printed Cementitious Materials: A Systematic Review

Abstract

3D printing is evolving at a fast pace in both the manufacturing and construction sectors. These advancements can greatly benefit these industries. However, the 3D printing of concrete structures presents some challenges due to defects in the 3D concrete printed elements. Hence, this study systematically reviews Artificial Intelligence (AI)-driven techniques, such as Computer Vision and Machine Learning, to identify surface defects that can occur in 3D-printed cementitious material structures. The adopted methodology was the PRISMA statement with the aim of reporting the systematic review and meta-analysis. Two well-known databases, Web of Science and Scopus, were utilised for data extraction of articles published during the past 10 years, between 2014 and May 2025. The initial search provided 110 articles, both conference and journal papers; after screening, only 11 were left for the final review assessment. The smaller number of the final articles shows that much work is still needed in this area. It has been observed that various computer vision and machine learning-based methodologies were employed to classify defects in 3D concrete printed structures. Deep learning algorithms, such as YOLO and RT-DETR, were featured as the most efficient in real-time defect detection and quality monitoring. It was also observed that real-time monitoring systems attached to 3D printers help in reducing the material wastage, which is essential to meet the sustainable goals. However, more work is still required to underline the defects of 3D-printed cementitious material, probably with the involvement of AI image processing tools and techniques. This can help to automate the defects in 3D-printed structures, and by this, the productivity could be enhanced.

1. Introduction

3D printing, also known as additive manufacturing, has transformed several sectors by providing complex products and projects with remarkable efficiency. This technology creates three-dimensional objects layer by layer from computer models, utilising materials such as plastics, metals, ceramics, and even several building material composites [1]. The major industries that have taken a lead in 3D printing utilisation for prototyping, tooling, and creating final products are aerospace, healthcare, and automotive [2]. For instance, the aerospace industry utilises 3D printing to make lightweight components that could be used in aircraft or associated forums, while the medical industry uses it to create patient-specific implants and prosthetics, helping them save time with more proficient output [3]. 3D printing has become a foundation of modern engineering due to its adaptability and cost-effectiveness, which drives innovation and reduces delays that bring this technology to the top over the traditional ways of manufacturing [4,5].

3D printing is also an emerging technology in the construction industry, providing solutions to labour shortages, high costs, safety concerns, and environmental problems, while ensuring efficient productivity [6,7]. The technology enables rapid construction of building components and structures using cement, mortar, concrete, polymers, and even recycled materials. This also supports reducing the depletion of natural resources by minimising waste usage and recycling the available materials, ultimately helping to reduce the carbon footprint in comparison to traditional methods of construction. Several researchers have demonstrated potential savings in construction time, labour, and cost of 3D-printed homes compared to traditional construction [8,9]. To illustrate, in 2021, a 3D-printed locality was created in Texas, demonstrating the feasibility of cost-effective housing solutions. The use of 3D printing has also proved attractive for architects to experiment with complex designs that are challenging or impossible to realise using traditional construction methods [10,11].

Despite the various benefits for the construction industry, 3D printing does not come without its problems [12]. One of the key concerns is the structural integrity of 3D-printed structures, as layer-by-layer deposition can create relatively porous and weak interphases, rendering them vulnerable to delamination, cracking, ingress of water and deleterious materials, and creating a higher risk of structural failure [13]. The properties of 3D-printed materials can change based on the mixture design of the ink material, open time, printing rate, and ambient and curing conditions. Moreover, changes in water-to-cement ratio (w/c) in the concrete mixture can reduce strength and durability. All these factors influence the void formation and porosity in the 3D-printed concrete [14,15]. While recycled or alternative materials are environmentally helpful, their performance can be unpredictable. This irregularity in the mixture and materials’ properties makes it difficult to achieve architectural and structural standards [15,16].

The processing of a 3D-printed structure faces two key challenges: sustaining anticipated performance, including surface quality, and upholding accurate set dimensions. To meet the standards and eliminate the irregularities in the 3D-printed concrete, it is essential to perform the surface finishing after the layer-by-layer extrusion [17,18]. The long-term performance of 3D-printed objects under environmental conditions is another concern. Variations in temperature, moisture exposure, and seismic activity are all significant factors that 3D-printed structures must survive, but the long-term behaviour of printed materials in such settings is still unidentified and requires the foremost attention [19]. For instance, 3D-printed concrete may have different thermal expansion properties in comparison to regular concrete, which leads to cracks or deformations over time [20]. The durability of printed structures in harsh environments or seismic regions is an unresolved subject, demanding further training and testing [21,22,23]. Moreover, these flaws emphasise the need for continued research to optimise printing materials and procedures as well as quality control methods to monitor the performance of 3D printing construction [24,25].

One of the main hindrances that restricts 3D printing in construction is its regulation and standardisation. The current building laws and regulations mostly rely on traditional building methods and materials, making it a hurdle to authenticate 3D-printed structures. The lack of standardised testing techniques for 3D-printed materials and components obstructs achieving safety and quality standards [26]. Addressing these complications will require collaboration among industrial players, researchers, and regulatory government agencies to develop new norms and frameworks specific to additive manufacturing. The research on 3D-printed structures is not new; however, it has been enormously heightened over time. As per the information gathered from the Scopus database [27,28], the first publication related to 3D printing appeared in 1931; however, there were not much more studies until the 1950s. Afterwards, the idea was immersed, and the publication trends also boosted. Figure 1 shows the number of published articles from the year 2000 to May 2025. It can be seen that currently researchers are highly focused on conducting a thorough investigation into the work related to 3D printing. Over time, a linear increase can be observed, showing the interest and potential in this field.

Figure 1. 3D printing-related publications as per Scopus database (2000–May 2025).

The identification of damage in concrete is of primary importance, and different techniques have been developed for the non-invasive assessment of elements and structures. For 3DCP, it is important to use real-time monitoring capability for in situ quality assurance during construction. A range of traditional non-destructive testing techniques has been developed to detect defects in conventional concrete structures, as summarised in Table 1. The principles, detection capabilities, limitations, and suitability of representative methods for 3DCP were also compared. For example, conventional techniques, such as Ultrasonic Pulse Velocity (UPV), Impact-Echo (IE), and Ground Penetrating Radar (GPR), are primarily designed to detect internal defects, including voids. GPR has a strong potential to identify interlayer debonding and internal voids in 3DCP due to its sensitivity to dielectric discontinuities. However, UPV and IE are often compromised by strong wave scattering and the complex geometries of printed components, which reduce measurement reliability. Infrared Thermography (IT) can serve as a useful preliminary mapping tool for near-surface defects, but it is highly sensitive to environmental conditions and surface emissivity. Simple hardness-based approaches, such as the Rebound Hammer (RH), are not recommended due to their limited accuracy on rough and heterogeneous printed surfaces.

Table 1. Traditional techniques to detect defects in concrete elements.

Reference	Method	Principle	Internal Defects	Surfaces Defects	Real-Time Monitoring	Limitations	3DCP Suitability
[29,30]	UPV	Measures wave velocity through concrete	Yes-(limited)	N/A	N/A	Strong scattering at interlayer interfaces; less effective in complex shapes	Limited (affected by 3D-printed layering and porosity)
[31]	IE	Stress wave reflections detect internal anomalies	Yes	N/A	N/A	Requires expertise to interpret data	Good (detect interlayer debonding)
[32]	GPR	EM waves reflect off interfaces	Yes	N/A	N/A	Signal attenuation in high moisture; geometry complicates interpretation	Highly Suitable (effective for layer debonding, voids)
[33]	IT	Surface temperature variations reveal anomalies	Yes	N/A	N/A	Affected by environment and surface conditions	Suitable (good for preliminary mapping)
[34,35]	AE	Monitors stress waves from active cracking	Limited	Yes (indirect)	Yes	Cannot easily localise defects without complementary methods	Useful (for monitoring crack evolution during loading)
[36]	RH	Measures surface hardness rebound	N/A	Yes (rough)	N/A	Surface roughness and heterogeneity reduce accuracy	Not Recommended (surface variability in 3D printing)
[37]	Half-cell potential	Measures electrochemical potential for corrosion	Limited	N/A	Limited	Less useful for non-reinforced or early-age structures	Limited (unless reinforced elements are included)
[38,39]	Radiography	Radiation penetrates material and creates images	Yes	N/A	N/A	Safety issues, costly equipment	Possible (rarely practical on-site)
[40,41]	LS/DIC	Captures surface geometry and deformation	N/A	Yes	Yes	Only measures surface, requires line of sight	Highly Suitable (excellent for layer deformation/cracking)
[42]	FOS	Embedded fibres monitor strain, temperature	Yes	N/A	Yes	Installation complexity, must be embedded during printing	Ideal (if pre-installed)

Not applicable: N/A.

For surface assessment and deformation tracking, Laser Scanning and Digital Image Correlation (LS/DIC) offer high-resolution measurement of layer geometry and crack evolution. These optical methods are especially suitable for 3DCP since they can be deployed during or immediately after printing to quantify surface defects and dimensional deviations. For real-time monitoring, a limited subset of techniques can provide continuous measurement capabilities. Fibre Optic Sensors (FOS), if integrated within the printing process, can enable direct monitoring of strain, temperature, and early-age cracking evolution with high sensitivity and temporal resolution. Similarly, Acoustic Emission (AE) can detect active cracking events but requires additional localisation methods to accurately characterise damage zones. LS/DIC can also be implemented in near-real-time modes to track surface deformation continuously, provided that line-of-sight access is maintained.

Overall, the comparison of the various techniques highlights that the majority of the well-established methods that are effective for post-process inspection lack inherent real-time monitoring capability that is critical for insitu quality assurance during 3DCP. This underscores the need for further integration of embedded sensing technologies and advanced surface measurement systems to enhance defect detection and ensure structural integrity throughout the printing and curing process.

While 3D printing holds a massive potential for the construction industry, resolving these defects and challenges is crucial for its successful implementation. Advances in material science, printer technology, and quality control processes, along with the development of industry standards, play a significant role in unlocking the full potential of 3D printing in construction. It is apparent that 3D printing is the future and a need for the construction industry; however, with a massive volume of construction, there is a high possibility of getting defects in the structural elements. Hence, this review intends to focus on the studies that are related to the surface defects detection in 3D-printed structures, mainly AI-driven defects analysis in 3D-printed cementitious materials or 3D concrete printing (3DCP). The scope of the review covers the 3D printing structures constructed using cementitious materials. While related themes such as quality monitoring, process optimisation, repair, and visualisation, etc., are discussed within the reviewed literature, this review primarily focuses on AI-driven computer vision and machine learning approaches for surface defect detection in 3DCP. The aforementioned themes are discussed only to the extent that they directly relate to the defect detection objective. It is worth mentioning that several review studies are available discussing the parameters, characteristics, and workability of the 3D printing structures in construction [43,44,45,46]; no such review is available discussing or highlighting the surface defects in 3DCP that require utmost attention. The outcome of this review paper is beneficial for additive manufacturing processing in construction to detect and minimise surface defects, which can significantly enhance 3D printing productivity and performance.

2. Methodology

The methodology of this review is based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, which follows the four phases and is recommended by several researchers for review processing [47]. This review was not prospectively registered. In the first phase, the articles are gathered based on a keyword combination and applied limitations from selective databases. In the second phase, three types of screening of the articles are made (i.e., duplication checks within and among the databases, and relevancy of the articles), first based on the title, and then the abstract. The process follows a set flow so that irrelevant articles get eliminated by the screening at each stage. In the third phase, the eligibility of the articles is evaluated by performing a full article screening. In this process, any article which is not relevant to the scope of review gets eliminated. In the final phase (i.e., the inclusion of the articles), a detailed assessment of the final extracted articles is made to get useful information and highlight the future gaps. The detailed PRISMA flow diagram followed in this review can be seen in Figure 2, while further details are discussed below.

Figure 2. PRISMA flow diagram.

2.1. Research Strategy

This review, based on the PRISMA statement, used the designed method to retrieve the information from the databases based on the defined scope of this study. The Web of Science (WoS) and Scopus databases were selected to extract the articles. The scope of the review was focused on the identification of defects through AI-driven techniques in cementitious 3D printing structures.

2.1.1. Selection Criteria

The data mining aimed at selecting technical papers in journals and conference proceedings from the last 10 years (2014 to 23 May 2025) that are published in English. The keyword arrangement in both WoS and Scopus databases was the same, giving an initial 42 and 68 articles, a total of 110. The following combinations were chosen to extract the most relevant articles: (“3D printing” OR “3D-printing” OR “3D” OR “Additive Manufacturing”) AND “Defects” AND (“Artificial Intelligence” OR “Machine Learning” OR “Computer Vision” OR “Photogrammetry” OR “Image Processing”) AND (“Concrete” OR “Cement” OR “Mortar”). The focus of this review was to retrieve the information from the technical studies; hence, the review papers were not considered.

2.1.2. Quality Assessment

The quality assessment was directed by predefined inclusion and exclusion criteria. Only those articles were considered which specifically addressed 3D-printed cementitious materials or 3DCP and employed AI-driven techniques for defect detection. The 110 articles were further assessed for quality and relevance. Initially, 26 duplicated articles were eliminated from the database. The articles common to both WoS and Scopus were retained under the WoS category and removed from Scopus. After this assessment, a title-based screening was performed, eliminating 23 articles from WoS and 22 from Scopus. The third screening was based on the abstracts, where 8 articles from WoS and 16 articles from Scopus were eliminated, leaving a total of 15 articles combined for further processing. These 15 articles were further assessed based on the full text reading, where 5 articles were eliminated as they did not fulfil the review’s scope. This gives the final 10 articles for further assessment. However, before proceeding further, a snowball technique was applied to the reference lists of these 10 articles that passed the full-text screening process. This process identified one additional article that had not been retrieved through the original WoS and Scopus database searches but fulfilled the inclusion criteria upon assessment. Hence, the article was included in the final review, bringing the total number of articles for final review assessment to 11. The detailed quality assessment processing of the data mining can be seen in Table 2.

Table 2. Screening for quality assessment.

Screening	Articles Retained		Articles Eliminated
	WoS	Scopus
Initial Search	42	68	0
After Duplication	42	42	26
Title Screening	19	20	45
Abstract Screening	11	4	24
Full Article Screening	8	2	5
Snowball Technique *	1		-
Total	11		-
			100

* The article identified via the Snowball technique was sourced from the reference lists of the 10 retained articles and was not part of the original WoS/Scopus search results.

2.2. Outline of Extracted Articles

2.2.1. Distribution of Articles

Following the PRISMA protocol, there were a total of 11 articles left for the final review assessment. Table 3 shows the distribution of the collected conference papers and journal papers against each year in both databases, while Figure 3 shows the total number of publications in each year with the trend over time. In 2024, a gradual increase can be seen in the publication trend, which indicates further potential in the studied area.

Table 3. Details of publication in each database.

Year	Conference Paper		Journal Paper
	WoS	Scopus	WoS	Scopus
2022	0	0	3	1
2023	0	1	0	0
2024	1	0	4	0
2025	0	0	0	1

Figure 3. Publication trend over time.

2.2.2. Keyword Co-Occurrence

In a systematic review, the keyword analysis is essential as it demonstrates the interconnectivity of the work conducted by the researchers. In this case, VOSviewer (version 1.6.20) was utilised to perform the keyword analysis, showing the connectivity of the keywords in the gathered 11 articles. Figure 4 shows the network visualisation of keyword co-occurrence, where the minimum occurrence of keywords was considered as 1 for 85 keywords. The significant clusters reported were “computer vision” and “3d concrete printing”. Table 4 shows the top five keyword co-occurrences, where it can be observed that the highest occurrence was for “computer vision” with a total link strength of 66, followed by “3d concrete printing” with a total link strength of 57, and finally followed by “3-d printing”, “3d-printing”, and “concretes”, where each were reported as having link strengths of 37.

Figure 4. Network visualisation of keywords.

Table 4. Top 5 keyword co-occurrences with link strength.

S. No	Keyword	Occurrence	Total Link Strength
1	computer vision	6	66
2	3d concrete printing	5	57
3	3-d printing	2	37
4	3d-printing	2	37
5	concretes	2	37

3. Interpretation of Article

The final 11 articles were evaluated and divided into a couple of categories to better understand the importance of 3D printing and its associated defects in construction. Table 5 shows the overall summary of these articles, focusing on the identified defect type, along with the method/technology, key findings, and focused area. The articles were distributed into two focused areas, mainly: “Defect Detection, Repair, and Quality Monitoring” and “Process Optimisation and Visualisation”.

Table 5. Summary of final extracted articles.

S. No	Author	Year	Journal/Conference	Defect Type Addressed	Method/Technology	Key Findings	Focused Area
1	Senthilnathan and Raphael [48]	2022	Sustainability	Surface defects	Computer vision LiDAR CNN	Robotic system detects and repairs defects using automated methods	Defect Detection, Repair, and Quality Monitoring
2	Rill-García, Dokladalova [49]	2022	Additive Manufacturing	Printing defects and anomalies	Computer vision	Automatic detection and localisation of anomalies
3	Liu, Yang [50]	2022	International Journal of Pavement Engineering	Pavement cracks	FDM-based 3D printer Image processing	Automated crack sealing with bitumen enhances road maintenance
4	Senthilnathan and Raphael [51]	2023	40th International Symposium on Automation and Robotics in Construction	Layer inconsistency	Computer vision Histogram of Oriented Gradients (HOG)	Proposes a non-intrusive quality monitoring technique
5	Gu, Li [52]	2024	Frontiers of Structural and Civil Engineering	General structural defects	Computer vision LiDAR CNN	Proposes a robotic repair system using 3D printing for maintenance
6	Jhun, Lee [53]	2024	IEEE Access	Filament defects	2D profile sensing Computer vision	Real-time filament shape monitoring for consistent extrusion
7	Zhao, Wang [54]	2024	Construction and Building Materials	Printing defects and anomalies	You Only Live Once (YOLO)v7 Deep Learning	High-accuracy real-time defect detection and prevention
8	Zhao, Sun [55]	2025	Automation in Construction	Printing anomalies	Transformer Networks RT-DETR	Transformer-based defect detection reduces material waste
9	Mechtcherine, Van Tittelboom [56]	2022	Cement and Concrete Research	Structural inhomogeneity	Destructive Testing Non-destructive Testing	Discusses anisotropy, voids, and standards for 3DCP quality testing	Process Optimisation and Visualisation
10	Zimmermann, Griego and Flatt [57]	2024	RILEM International Conference on Concrete and Digital Fabrication	Surface irregularities	AR Machine Learning Digital Twin	AR utilisation for enhanced defect visualisation via Microsoft HoloLens
11	Silva, Wagner [58]	2024	Inventions	Layer morphology issues	Machine vision ML	Automated layer width regulation for print consistency

3.1. Defect Detection, Repair, and Quality Monitoring-Related Studies

The identification of damage in concrete is of primary importance, and different techniques have been developed for the non-invasive assessment of elements and structures. 3D printing requires improved attention to both quality management and structural reliability due to observed time constraints and unreliable material dispersion during printing operations. Table 6 summarises the articles related to the “Defect Detection, Repair, and Quality Monitoring” categories under the headings of Objective, Key Challenges, Testing Methods, Quality Control Methods, Structural Considerations, and Way Forward. Further discussion has been provided below.

Table 6. Summary of defect detection, repair, and quality monitoring-related studies.

S. No	Author	Objective	Key Challenges	Testing Methods	Quality Control Methods	Structural Considerations	Way Forward
1	Senthilnathan and Raphael [48]	Develop a non-intrusive, computer vision-based quality assessment method for 3D-printed concrete structures	Inconsistency in print quality, workability loss over time, deformities, and discontinuities in printed layers	Two-bin Linear Binary Pattern (LBP) textural analysis Entropy calculation Image data collection using a 2D camera	Entropy-based thresholding to detect defects, real-time monitoring, and corrective actions during printing	Layer thickness reduction Structural stability due to the weight of successive layers	Extend the practice to real-time monitoring, improve accuracy and integrate corrective actions during printing
2	Rill-García, Dokladalova [49]	Develop a computer vision-based detection of anomalies for 3DCP	In-line characterisation of the last printed layer	Image Acquisition Interlayer line and layer Segmentation Local geometry and texture characterisation	Single RGB Camera Computer vision	Segmented interlayer lines measure multiple geometrical properties	Extend 3DCP monitoring towards a closed-loop control
3	Liu, Yang [50]	Integrate computer vision and path planning algorithms for crack detection and sealing Examine the feasibility of 3D printing technology usage for road maintenance	Instability in sealant applications causes splashing and overfilling Pump stability issues affecting nozzle head movement CNC machine errors, including wear and tear, geometrical errors, and thermal variations Slow image processing speed	Laboratory testing on concrete slab specimens with artificially induced cracks MATLAB-based image processing algorithms Greedy path planning algorithm for crack sealing Modified Anycubic Prusa I3 MK2 printer with 0.1mm precision	Image pre-processing and enhancement techniques Noise elimination algorithms Calibration of workspace using pixel-to-distance ratio	Modified printer structure with a separate crack-sealing platform Replacement of the hot-end extruder with a specialised nozzle head Three-dimensional movement capability for uneven surfaces	Development of more robust crack detection algorithms Enhancement of platform structure and components
4	Senthilnathan and Raphael [51]	Develop an autonomous quality monitoring method for 3DCP using computer vision-based texture extraction techniques	Presence of voids, cracks, and defects in printed layers Real-time monitoring of quality during the printing process	Modified Histogram of Oriented Gradients (HOG) algorithm Weighted and normalised weighted texture metrics for analysis	Continuous monitoring using computer vision Temporal textural variation study to assess buildability properties	Focus on extrudability and buildability Correlation between workability, moisture content, and surface texture	Automation of layer cropping using image segmentation Further validation of the HOG algorithm for different concrete mixes
5	Gu, Li [52]	Develop a computer vision-based robotic system for automated concrete defect detection and repair using 3D printing technology	Low automation in concrete defect repair Manual inspection limitations Time-consuming and labour-intensive processes	Grouping of LiDAR and depth camera YOLOv7 neural network for defect detection	10-fold cross-validation Cosine similarity analysis Evaluation of repair effect through appearance and mechanical properties	Employ precise robotic arm movement Sensor fusion for accurate defect location Coordinate system transformation	Enhance system automation, reduce human intervention, and simplify operational procedures
6	Jhun, Lee [53]	Develop a real-time geometric quality monitoring system for extruded filaments in 3DCP	Variability in concrete properties, High costs of failures Limitations of existing sensors Requirement for comprehensive cross-sectional data	2D laser profile sensing Image processing Clustering methods Linear regression	Real-time, non-contact, non-destructive automated inline testing	Physical parameters and extrusion rate of concrete filaments	Integration with deep learning Material property predictions Closed-loop quality control
7	Zhao, Wang [54]	Develop a real-time, accurate defect detection system for 3DCP using deep learning and computer vision	Conventional observation limitations Real-time performance Accuracy of existing automated methods	YOLOv7 algorithm Data augmentation	Real-time monitoring Automatic defect detection Measurement using enhanced YOLOv7	Breakpoints, voids, cracks, pits, and deformation due to uneven extrusion	Expand defect datasets On-site performance evaluation Integration with BIM systems
8	Zhao, Sun [55]	Develop a real-time, high-accuracy defect monitoring system for 3DCP using transformer networks	Limited data availability High computational power requirements Necessity for real-time performance	RT-DETR algorithm Data augmentation Transfer learning Loss function optimisation	Real-time, automatic defect detection and measurement using enhanced RT-DETR through YOLOv8 and YOLOv7	Breakpoints, voids, pits, and deformation due to rapid drying and fast hydration	Expand defect datasets Assess detector adaptability to different materials Define adjustable printer parameters

Senthilnathan and Raphael [48] developed a computer vision method to track 3D-printed concrete structure defects that occur during manufacturing by identifying printing process-related deformities. A textural analysis based on a two-bin Linear Binary Pattern (LBP) was adopted. This technique enables the measurement of surface texture variations in the extruded layers. Furthermore, the printed layer images were captured by a 2D camera that enables entropy calculation to evaluate texture patterns. The real-time monitoring enables users to amend extrusion speeds, which leads to better print quality and less material waste. Rill-García, Dokladalova [49] developed a computer vision-based method for the detection of anomalies of 3DCP. With the help of a fixed camera at the nozzle, image acquisition was performed. With a F-score (harmonic mean of precision and recall) of 91%, the developed strategy traces and identifies the printed layers distinctly.

Liu, Yang [50] demonstrated a technical approach for automated pavement crack sealing through a modified 3D printing method. The utilised system incorporates a precision-adjusted Anycubic Prusa I3 MK2 printer, computer vision, and path planning algorithms within its structure. The study identified important factors for sustaining high-quality standards of 3D-printed construction structures by emphasising both material placement accuracy and structural system stability. It was observed that the same difficulties with splashing, inconsistent material flow patterns, and the problem of overfilling exist at both construction and 3D printing at scale. The water level sensors and calibration systems used in the study were found to enable automated quality controls that deliver vital knowledge about preventing defects in 3D-printed construction structures. Hence, the construction-scale 3D printing requires improved attention to both quality management and structural reliability due to observed time constraints and unreliable material dispersion [50].

In another study conducted by Senthilnathan and Raphael [51], a visionary approach was adopted for automated 3DCP structure quality inspection through visual defect identification of voids and cracks. For this purpose, a modified Histogram of Oriented Gradients (HOG) algorithm with the capability to detect minor and major texture variations across horizontal, vertical, and diagonal directions was utilised. The results showed that the normalised weighted texture values effectively detect defective layers within the printed sections as they inversely relate to workability changes throughout time. The study also highlights real-time computer vision feedback control systems that can operate printing parameters, including extrusion speed, to enhance print results and lower material consumption.

The repair of concrete structures can be done with a robotic system based on computer vision and 3D printing methods. The system operates through a combination of LiDAR and depth cameras, along with the YOLOv7 convolutional neural network, to automatically identify concrete defects and perform repairs. Through simultaneous localisation and mapping (SLAM) functionality, the system achieves the ability to produce models of cylindrical surfaces while obtaining exact geometric parameters. Through coordinate system transformation, the robotic arm obtains precise movements to position itself at the repair site [52]. Jhun, Lee [53] also studied the real-time geometric quality monitoring system and focused on the problems caused by inconsistent concrete properties and high expenses that lead to printing failure. The 2D laser profile sensing system and a rotatable nozzle were utilised with the involvement of image processing algorithms. It was observed that the developed sensor provides affordability, along with precise design elements for 3DCP, in addition to being lightweight, surpassing commercial sensors. The analysis proved that the system can implement real-time monitoring in construction sites to ensure safe and high-quality construction of 3D-printed structures.

Zhao, Wang [54] employed deep learning and computer vision techniques to identify the defects in 3DCP. The deep learning-based data augmentation enables the system to create accurate virtual data that substitutes training dataset information for better measurements. The study showed the importance of real-time monitoring as it helps in defect prevention that can occur due to breakpoints, voids, and cracks in the 3DCP structure. In addition, the study also evaluates the defect rates corresponding to printer control parameters to help operators gain better printer management abilities. In another study, Zhao, Sun [55] worked on a real-time high-accuracy defect detection system for 3DCP through the application of the RT-DETR transformer-based detector. The study utilised deep learning data enhancement and image management techniques. It was evident that the superior performance and optimised computation speed of 72 Frames Per Second (FPS) became possible for the enhanced RT-DETR algorithm through different strategic approaches, including transfer learning and loss function optimisation, which produced results with mAP50 of 98.1% and mAP50–95 of 68.0%. In addition, it was revealed that the system provides operators with real-time defect detection along with measurement of breakpoints and pits with voids, which allows for adjustments to printer settings to lower waste losses and decrease structural hazards.

3.2. Process Optimisation and Visualisation-Related Studies

For better structural integrity and scalability, researchers have adopted the process optimisation and visualisation in 3DCP. Table 7 shows the summary of the articles related to process optimisation and visualisation, which provides the information under the categories of Objective, Key Challenges, Testing Methods, Quality Control Methods, Structural Considerations, and Way Forward for each relevant study.

Table 7. Summary of process optimisation and visualisation-related studies.

S. No	Author	Objective	Key Challenges	Testing Methods	Quality Control Methods	Structural Considerations	Way Forward
1	Mechtcherine, Van Tittelboom [56]	Examines characterisation and quality control methodologies for 3DCP	Lack of formwork and compaction affects mechanical properties Difficulty in applying conventional testing methods to printed concrete	Destructive Testing: Compressive strength Flexural strength Interlayer bond strength Shrinkage test Durability test Non-Destructive Testing: Geometry assessment Layer defects Shrinkage measurement Early age monitoring Photogrammetry 3D laser scanning	Comparison of specimen preparation techniques (casting vs. full-scale printing vs. laboratory setup)	Influence of printing direction on strength and durability Material characterisation for interlayer bonding Value curing conditions on mechanical performance	Standardised testing protocols for printed concrete Long-term durability studies and field applications Automated quality control using AI and computer vision
2	Zimmermann, Griego and Flatt [57]	Develop an augmented reality visualisation and machine learning-driven approach for inspecting 3DCP structures Create a digital twin for better assessment of defects	Trouble in achieving ideal control of extrusion material properties Limited trust and adoption of 3DCP structures Issues with filament tearing due to mismatching nozzle/extrusion speeds or nozzle clogging	Camera mounted on printing nozzle for top-down perspective monitoring Convolutional Neural Networks (CNN) for defect detection Development of a software tool using Unity and Mixed Reality Toolkit Testing with Microsoft HoloLens 2 for AR visualisation	Real-time monitoring system using nozzle-mounted cameras Machine learning-driven image analysis for defect detection Digital twin creation with defect labelling and positioning Automated placement of a digital twin onto concrete structures	Impact of filament tearing on reinforcing steel bars Risk of concrete outflow and steel bar exposure Need for tracking defect positions and analysing implications Subsequent layer sealing for structural integrity	Development of more reliable documentation procedures Improvement of user-friendly inspection software Enhancement of automatic digital twin superimposition Integration of extended reality technology in 3DCP Research on interdisciplinary aspects, including robotics and computer vision
3	Silva, Wagner [58]	Develop a real-time precision control system for 3DCP using machine vision and learning algorithms to regulate the layer	Frequent manual adjustments leading to errors Volatile material characteristics of concrete Real-time monitoring and controlling layer width Limited dataset for machine learning training	Numerical simulations using OpenFOAM to generate synthetic datasets Machine learning models (XGBoost) trained on synthetic data Real-time monitoring using a depth camera and rotary mechanism	Real-time layer width measurement using computer vision Automated adjustments of robot motion speed based on ML predictions	Layer width consistency for structural integrity Material flow and motion speed impact on layer morphology Simulation of concrete rheology using the Herschel–Bulkley model	Improve the rotary mechanism for better camera alignment Validate the system in real-world printing scenarios

The additive manufacturing of 3DCP results in a layered structure compared to conventional concrete casting methods because of increased component anisotropy and inhomogeneity. Such properties require adjustments to established quality control methods. The assessment methods for printed concrete quality include strength measurements through compression tests and flexural tests, together with interlayer bond tests and shrinkage studies. The Digital Image Correlation (DIC), combined with photogrammetry and 3D laser scanning, can be used for in situ quality assessment. The defects within 3DCP occur naturally because of the additive nature of layers, where quality control measures, combining advanced monitoring systems, controlled materials, and printing parameters, can help in reducing defects affecting structural integrity and long-term performance [56].

Zimmermann, Griego and Flatt [57] established an innovative solution to manage 3DCP quality control problems by utilising augmented reality and machine learning approaches. The process consists of three steps, where the cameras are placed on nozzles to track printing operations while implementing convolutional neural networks to analyse collected data until these findings become digital twins. The system identifies and analyses filament-tearing defects as they affect structural integrity, mainly in places with reinforcing steel bars. The proposed solution offers Microsoft HoloLens 2 as an augmented reality interface that displays defect details directly on top of physical structures. The initial trials proved that the digital twin projection onto concrete structures is possible, yet more enhancements should focus on self-placement methods and user-friendly interfaces.

Silva, Wagner [58] optimised 3DCP through machine learning and real-time monitoring systems integration. The primary objective was to automate the control of printed layer width by predicting the necessary motion speed of the extruder based on various printing parameters. The study also employed numerical simulations to generate a synthetic dataset, which was then employed to train an XGBoost model. The model achieved an accuracy of 87% in predicting robot speed. The application of computer vision was shown to allow real-time monitoring and alterations during the printing process, which ensures high-quality outcomes [46]. Based on the findings, it was suggested that further research should emphasise expanding the dataset to accommodate different concrete mixes and evaluate the framework’s effectiveness in harsh environments.

3.3. Discussion

From the 11 extracted relevant articles, various aspects concerning defect identification, quality control, and maintenance of 3DCP were carefully examined. Defect identification was a critical part of this investigation. Many studies involved robotic solutions, along with computer vision and machine learning techniques for automated structural integration. Researchers tend to examine automated solutions for defect repair by modifying 3D printers to conduct crack sealing operations and maintenance. These efforts aim to improve durability with limited physical involvement.

Several studies employ quality monitoring systems that utilise deep learning techniques along with real-time sensing and profile scanning equipment to uncover the defects. It was observed that the combination of augmented reality with digital twins can deliver high-quality visual inspection. In terms of process control investigation, it was observed that the real-time changes in printer settings (extrusion rate, nozzle height, print speed, temperature, vibration or compaction controls, etc.) can lead to improved layer placement and decreased degree of defects.

While comparing the reviewed studies, a clear methodological progression is evident, from classical texture-based algorithms such as LBP and HOG towards transformer-based deep learning models, with a corresponding improvement in detection accuracy and processing speed. In terms of defect categories, surface voids, cracks, and breakpoints were the most commonly addressed, while filament irregularities and layer inconsistencies received comparatively less attention. Regarding reported metrics, performance varied considerably, from an F-score of 91% in camera-based segmentation approaches to mAP50 of 98.1% in transformer-based models, reflecting the influence of both algorithm sophistication and dataset quality on detection reliability. On the other hand, active sensing technologies such as LiDAR and laser profiling offer richer geometric data but introduce higher system complexity, whereas passive RGB camera-based approaches remain more accessible and deployable in real construction environments. It is worth mentioning that across both thematic groups, real-time monitoring capability emerges as the most critical performance differentiator, with deep learning-based systems demonstrating clear advantages. However, a shared limitation across nearly all studies is the reliance on small, context-specific datasets, which constrains model generalisation and cross-environment applicability.

Various studies examined dynamic monitoring systems that control both extruder velocity and printing outlet position, while detecting defects in the structure. The implementation of automated response loops is shown to lead to better print reliability, which minimises material waste and strengthens structural integrity. The implementation of robotics, SLAM technology, and LiDAR sensors in 3DCP processes can lead to more accurate defect identification and maintenance without human involvement, which can ultimately enhance productivity.

The work on defect identification of 3DCP is still limited, and there is a lot of potential for further development to resolve challenges in 3DCP. Most challenges persist because developers have to deal with real-time performance issues along with scalability problems and construction environmental conditions. Hence, at present, there is a need for larger defective material databases while developing an automatic defect system for enhancing 3DCP quality monitoring capabilities.

Regardless of the promising results demonstrated in the reviewed studies, several barriers hinder the field adoption of AI-driven defect detection systems in 3DCP. One such barrier is limited dataset availability, as most studies relied on small, project-specific datasets that restrict model generalisation to different concrete mixes and printing conditions. Another barrier is the varying light conditions on construction sites that pose difficulties for camera-based systems, as inconsistent illumination directly affects image quality and detection reliability. Moreover, the material variability in cementitious mixtures further complicates defect classification, since changes in mix composition alter surface texture and appearance, making it difficult for trained models to maintain consistent performance. Another important barrier is the sensor placement, as the mounting of cameras or laser sensors on moving nozzles introduces vibration, causing occlusion and alignment issues during printing. Furthermore, the absence of standardised benchmark datasets and evaluation protocols across studies makes direct performance comparison difficult, limiting the ability to objectively assess and validate detection systems for broader industry deployment.

Based on the overall assessment of the final studies, a few loopholes were identified that could lead to a new research direction for the improvement of 3D-printed structures in construction:

• The dataset for 3DCP defects remains limited. Development of standardised benchmark datasets with diverse defect categories, concrete mixes, and environmental conditions is essential for better model training, testing, and cross-study performance comparison.
• The research on AI-based defect detection technology for 3DCP requires the development of multimodal sensing frameworks that integrate camera-based, LiDAR, and laser profiling systems to improve detection robustness across varying site conditions.
• A crucial advancement would involve the development of closed-loop control systems that combine real-time AI defect detection with automated printer parameter adjustment, enabling immediate corrective action during printing.
• Real-time AI systems can minimise printing defects by detecting them in advance, thus allowing immediate corrective measures that stop structure degradation. However, standardised validation procedures and benchmark protocols must be established to enable objective evaluation and broader industry deployment of such systems.

4. Conclusions

The purpose of this systematic review of AI-driven defect assessment was to evaluate the defects in 3D-printed cementitious material structures. In this case, the PRISMA statement was adopted to extract the technical articles (Journals and Conferences) from the Web of Science and Scopus databases by considering the papers from 2014 to 23 May 2025. The initial search provided 110 articles, and after the final screening, it was reduced to 11 studies. Although the final screening provided a smaller number of articles, it was still worth reviewing, as there is a need to work further in this area to avoid the defects in 3DCP. As per reviewed studies, the modern quality assessment of 3D concrete printing occurs through the integration of computer vision, deep learning, and automated systems. A deep learning system confirmed its position as a robust solution since it enables both quantitative evaluations and instant detection of defects. In addition, it was observed that the execution of real-time monitoring systems with automated corrective actions can provide 3DCP with more reliability because material waste can decrease while the construction processes can produce improved results. However, before executing a large-scale implementation of 3DCP, the industry must address existing implementation limitations, which include high computational requirements along with restricted data access and system adaptation in real-time. This review provides an insight into the 3DCP and gives a way forward to the researchers to address the challenges for its better implementation.