The review followed the PRISMA guidelines, the gold-standard protocol for systematic literature reviews, and searched both the Web of Science and Scopus databases. The team started with a Boolean keyword search that yielded 683 results from Web of Science and 603 from Scopus, then removed duplicates to arrive at 1,064 unique records. Twenty additional papers were harvested from internet searches and reference-list browsing. After screening titles, abstracts, and full texts using the Rayyan web tool — with two reviewers working independently and a third verifying the final selection — only 24 studies survived the gauntlet. That small final sample, the authors note, reflects how young and specialized this research niche really is, not a failure of the search strategy.

What did those 24 studies reveal? The researchers used bibliometric analysis with VOSviewer software to map keyword co-occurrence networks and publication trends, then layered on qualitative content analysis. The keyword network showed that machine learning and additive manufacturing had the strongest link, with link strengths of 19 and 17 respectively, followed by artificial intelligence, deep learning, 3D printing, and 3D concrete printing. Four distinct thematic clusters emerged: AI methods for design generation (dominated by generative adversarial networks and deep learning), AI-driven optimization logic (machine learning, topology optimization, generative design, path planning), assessment and fabrication intelligence (computer vision, buildability, deep GANs), and the broader additive manufacturing context itself.

The first major research area is AI-driven design and form generation. At the conceptual stage, machine learning approaches help architects rapidly evaluate numerous design alternatives against predefined architectural and construction requirements. Generative models such as GANs and variational autoencoders support automated layout generation, producing unique spatial configurations and building designs that align with structural and functional demands. Some researchers have even experimented with large language models: one study showed that ChatGPT could generate codes and polynomial equations from descriptive language prompts, which were then used to build parametric models — a framework that promotes collaboration between architects and AI. Visualization tools like Veras and other AI-enhanced programs can convert simplified digital models into diverse architectural interpretations in minimal time.

Beyond creativity, AI plays a crucial technical role in 3DCP through predictive material modeling. Machine learning algorithms such as XGBoost have been reported to accurately forecast essential mix-design parameters, including slump flow, compressive strength, and overall structural strength, which facilitates more dependable and optimized printability of cement-based materials. Deep learning models can generate novel component geometries while simultaneously evaluating their printability. Semi-supervised models have been used to design micro-lattice and cellular structures for large-scale additive manufacturing, assessing both manufacturability and structural feasibility. Artificial neural networks extend these capabilities into sustainable building design by optimizing construction processes, predicting project performance, estimating costs, and supporting environmentally responsive design decisions — effectively becoming decision-support systems rather than mere predictive tools.

The second research area is optimization of material and structural performance. Here, machine learning has emerged as the dominant AI technique, facilitating data-driven optimization throughout the entire design-to-construction pipeline. Studies report that ML can help reduce material consumption, improve printability, and increase construction efficiency — potentially contributing to more sustainable construction practices. One notable example is a CNN-based framework developed by Ozalp and Aldemir that automatically estimates the structural capacities of 3D-printed concrete walls with optimal material parameters, performing design evaluations autonomously rather than requiring manual calculations. Another study applied explainable ensemble machine learning combined with deep generative adversarial networks to predict bond failure modes and strengths in 3DCP, using data augmentation to overcome limited experimental datasets. A probabilistic machine learning framework developed by Ma and colleagues reliably predicted bond strength and reinforcement embedment requirements with uncertainty estimates, improving both structural safety and construction efficiency.

The third research area is fabrication monitoring and assessment. Because 3DCP uses no formwork, printing defects must be detected before subsequent layers cover them or make the mistake irreversible. Computer vision systems can evaluate temporal changes in layer thickness and surface texture during printing, using image-derived indicators such as entropy standard deviation and maximum layer thickness deformation to assess buildability in real time. Deep learning models can spot defects, anomalies, and process changes by analyzing data from sensors and cameras, helping keep geometry consistent and preventing structural problems. Generative models like variational autoencoders and GANs help simulate defects and improve automated quality control. One particularly striking study introduced a tower-crane-based printing system controlled by deep reinforcement learning, which improves extruder toolpath control while assessing the accuracy and feasibility of large-scale, multi-storey printing — effectively making assessment part of an intelligent control loop rather than a separate inspection stage.

Despite these advances, the review identifies serious challenges that currently limit real-world implementation. At the material level, the concrete mixture must simultaneously satisfy conflicting requirements related to pumpability, extrudability, buildability, shape retention, open time, and early-age strength, with variations in water content, aggregate characteristics, and admixture dosage all playing determining roles. At the fabrication level, dimensional deviations can result from material deformation, inconsistent extrusion, nozzle movement, layer-height variation, and inaccuracies in robotic positioning — and these deviations can accumulate through early layers to influence final geometry and structural performance. Structural challenges include weak or variable interlayer bonding, anisotropic behavior, reinforcement integration, and uncertainty regarding long-term durability. A geometry that is computationally optimized may not necessarily be printable or structurally reliable under construction conditions.

The authors also flag a deeper systemic problem: most of the reviewed evidence comes from simulations, laboratory experiments, and controlled prototypes rather than full-scale construction sites. Assessment accuracy is strongly dependent on the quality and diversity of available data, and few public 3DCP datasets exist. Data augmentation methods like bootstrap resampling and deep generative adversarial networks are being used to make small datasets more robust, but synthetic data raises questions about whether artificially generated samples adequately represent real material variability and construction-site conditions. The review also notes the absence of standardized digital workflows connecting BIM, structural analysis, material data, slicing, robotic control, and quality verification, which restricts practical implementation.

Looking ahead, the researchers outline several promising directions: developing open benchmark datasets that combine material composition, printing parameters, environmental conditions, and observed printing defects; creating closed-loop fabrication control that integrates computer vision and sensor data to automatically adjust extrusion rate, nozzle speed, layer height, and tool paths in real time; building multi-objective design optimization frameworks that simultaneously consider structural performance, material consumption, embodied carbon, operational energy, printability, and construction cost; and connecting architectural models, structural simulations, and robotic control through AI-supported BIM and digital twin workflows. The authors suggest that large language models integrated with BIM platforms may eventually automate G-code generation, reinforcement detailing, and construction planning. If these gaps are addressed, AI-driven 3DCP could reshape architectural design, affordable housing, and sustainable construction — reducing material usage, improving printability, and automating quality control in ways that traditional methods simply cannot match.

Subject of Research: AI applications in 3D concrete printing for architectural design and construction

Article Title: Artificial intelligence driven transformation of 3D concrete printing in architecture: a systematic review

Article References: Artificial intelligence driven transformation of 3D concrete printing in architecture: a systematic review. (n.d.). https://doi.org/10.3389/fbuil.2026.1914666

Image Credits: AI Generated

DOI: 10.3389/fbuil.2026.1914666

Keywords: artificial intelligence, 3D concrete printing, additive manufacturing, machine learning, generative design, computer vision, systematic review, sustainable construction, digital fabrication, architecture, deep learning, construction automation