Machine Learning Approaches for FDM-Based 3D Printing: A Literature Review

Abstract

Three-dimensional (3D) printing has become a widely used manufacturing technology, but predicting the mechanical and physical properties of printed parts remains a critical challenge. In particular, tensile strength, hardness, and surface roughness are essential for assessing product quality and reliability. Addressing this gap requires advanced computational approaches. Machine Learning (ML) algorithms have the potential to enhance automation and provide accurate predictions of product quality in 3D printing. The primary objective of this study is to review, classify, and analyze recent research on the application of ML algorithms for predicting mechanical (tensile strength, hardness) and physical (surface roughness) properties in 3D-printed materials. This review aims to identify current research trends, highlight existing gaps in the literature, and propose potential directions for future investigations in this rapidly evolving interdisciplinary field. For this purpose, a literature review was conducted in the Web of Science database using keywords such as “3D printing”, “machine learning”, “tensile strength”, “hardness”, and “roughness. The review systematically examines the material types, ML algorithms, printing parameters, and testing methods reported in the literature, and the frequency of these parameters is visualized through graphs to illustrate research trends. The findings reveal that ML algorithms achieve high accuracy in predicting tensile strength, hardness, and surface roughness, providing valuable insights for process optimization. However, current research still shows limited evidence for direct improvements in efficiency and error reduction, indicating important directions for future work.

1. Introduction

Three-dimensional (3D) printers allow objects that meet different needs to be designed in a computer environment and then printed. Considering the ease of printing, unique design opportunities, and economic advantages offered by 3D printers, interest in 3D printers is rapidly increasing [1]. Additive Manufacturing (AM), also known as 3D printing, comprises a group of manufacturing techniques that fabricate objects layer by layer to form 3D structures. Key advantages of this method include reducing material waste in the creation of designed structures, enabling rapid prototyping, and allowing the production of parts with complex geometries that would be difficult or impossible to achieve with traditional manufacturing methods [2,3,4]. AM is studied under seven basic principles. These are material extrusion [5], directed energy [6], binder jetting [7], sheet lamination [8], powder bed mixture [9], material jetting [10], and vat photopolymerization [11]. In this list, Fused Deposition Modeling (FDM), a method under material extrusion, is a widely used technique for manufacturing thermoplastic polymers [12,13]. Due to its widespread accessibility in academic research and its ability to generate larger publicly available datasets suitable for ML studies, FDM has become a particular focus of attention. While this review places a specific emphasis on FDM, it also provides a comprehensive and critical examination of ML applications in other major industrial technologies, including vat photopolymerization (SLA), Selective Laser Sintering (SLS), and Directed Energy Deposition (DED). FDM technology, also referred to as Fused Filament Fabrication (FFF), originated in the late 1980s and has seen rapid advancement ever since [14]. FDM is an AM method based on the principle that a thermoplastic filament is taken from a spool and fed through an extruder’s cold and hot ends. The filament is heated above its glass transition or melting temperature at the hot end, creating a semi-molten state. This molten material is then precisely deposited layer by layer onto a build platform using a nozzle with a diameter ranging from 0.3 to 1.0 mm, following the geometry defined in a computer-aided design (CAD) model [15,16]. The cold end pulls the filament from the spool and controls the feed rate. Using the heating chamber and nozzle, the hot end liquefies the filament. The extruded thin, viscous plastic strip adheres to the substrate on which it is laid, solidifies as it cools, and gradually bonds with the previous layer, creating the final three-dimensional part [17]. The process is available in variants adapted to different filament forms (e.g., hot rod/pellet extrusion or cold extrusion of slurries). Process parameters such as nozzle temperature, print bed temperature, print speed, and layer height must be optimized to achieve high-quality parts [18]. While FDM is favored for its low cost, ease of use, and broad material compatibility [3], it is associated with limitations such as shrinkage during cooling, anisotropy in mechanical properties due to differences in interlayer adhesion, and, in some cases, lower dimensional accuracy compared to other AM techniques [19]. Given these characteristics, proper management of process parameters and material selection is critical to achieving the desired mechanical and geometric performance.

It is important to evaluate the performance of the mechanical and physical properties of the materials used in FDM printing [20]. Polylactic acid (PLA) [21,22] and its enhanced variants, PLA Plus, Acrylonitrile Butadiene Styrene (ABS) [23,24], and Polyethylene Terephthalate Glycol (PETG) [25], are widely used thermoplastic polymers in FDM printing. PLA is used in many industrial areas, such as automotive [26], aviation [27], agriculture [28,29], and biomedical [30]. PLA Plus exhibits improved tensile and bending strength compared to standard PLA, which is beneficial for applications requiring more durable and robust parts [31]. In addition, it can be functionally operated with engineering filaments such as Acrylonitrile Styrene Acrylate (ASA), polycarbonate (PC) [32], Polypropylene (PP) [33], Thermoplastic Polyurethane (TPU) [34], Polyamide Nylon (PA) [35], and Thermoplastic Elastomer (TPE) [36]. Additionally, it is feasible to manufacture products using filaments reinforced with carbon fiber [37], including materials such as 316L steel [38] and glass fiber-reinforced polymeric [39].

Many parameters affect the print quality in the FDM production method. Criteria such as durability [40] and linearity [41] can be optimized by choosing the right parameters that affect the print before starting. For example, by modeling the effects of the print parameters on the surface quality and optimizing these parameters, smoother surfaces can be obtained without the need for any post-printing process [42]. The surface quality of 3D-printed parts is important not only for their appearance but also for their functionality and the level of post-processing required. Similarly, mechanical properties such as tensile strength are key indicators of a part’s structural integrity and overall performance. Accurately predicting these characteristics is vital for streamlining the production process and ensuring the reliability of the final product. Traditionally, obtaining the desired surface quality and tensile strength in AM has largely been based on experimental adjustments and iterative testing. These methods typically require extensive testing and repeated adjustments to process parameters, making them both time-consuming and resource-intensive, often without delivering optimal results [43]. It is precisely in this position that the extensive use of ML in AM components is highlighted. ML techniques offer significant advantages for AM in areas such as material analysis and selection, design view, optimization of process parameters, defect detection, real-time monitoring, and sustainability. By leveraging the power of ML capabilities, researchers and industry professionals can significantly increase the scope, breadth, and overall effectiveness of AM [44].

The role of ML, especially in FDM technology, is critical to ensuring the more efficient and optimized reproduction of production parts. ML offers powerful solutions in many areas, such as optimizing print data, detecting surface units and mechanical changes, and identifying possible errors during production. This method provides significant time requirements and budget considerations savings over traditional trial-and-error methods while also improving print quality and reliability. This review focuses on using ML algorithms to predict mechanical and physical test results of materials. The study aims to synthesize this literature and evaluate the materials, parameters, and methods used in existing studies. The potential of ML in the quality and performance optimization of materials has been examined in detail through a literature review. Within the scope of the article, publications were grouped and analyzed according to material type. Publications on commonly used polymer materials such as PLA, ABS, PETG, Polylactic Acid + Carbon Fiber (PLA-CF), and High-Density Polyethylene (HDPE) are presented in Table 1, while studies on less commonly used polymer materials are included in the text.

Several reviews in the literature have compared the applications of ML in 3D printing parameters [45,46,47]. While the study by Zhang et al. [45] focuses on the technical details of ML algorithms and applications such as process optimization and error detection, the reviews by Goh and Rojek et al. [46,47] cover broader and emerging application areas. However, these studies are generally limited to the classification of algorithms and providing an overview. In contrast, our study conducts a comprehensive review of articles retrieved from the WoS database using keywords related to 3D printing parameters and ML, performing statistical and graphical analyses on the selected papers. Specifically, by focusing on the prediction of mechanical and physical test results using ML, it presents a detailed comparison of process parameters, test types, and algorithms. Additionally, through graphs, tables, and standardized charts, the study offers a robust representation of the data, providing a deeper and more original contribution than existing reviews in the literature. In this respect, our study differentiates itself from current reviews in the field of ML and 3D printing parameters and makes a significant contribution to the advancement of the field.

2. Methodology

2.1. Search Strategy

The WoS literature search was conducted using the following keywords: “3D printing”, “machine learning”, “tensile strength”, “hardness”, and “roughness”. In the scope of the review, the selected studies were thoroughly analyzed in terms of the types of materials used, the ML algorithms applied, the printing parameters considered, and the mechanical and physical test methods performed. Moreover, the frequency of these parameters in the literature was visualized through graphs, and their distribution was analyzed in detail. Figure 1 clearly shows that the number of academic studies on using ML to optimize 3D printing parameters has steadily increased over the years, with a particularly sharp rise after 2020, highlighting the growing research interest in this field. In particular, studies focused on tensile strength have increased rapidly since 2022, studies on roughness have increased continuously after 2022, and studies on hardness have shown a lower increase trend. This data shows that academic interest in studies predicting test results with ML of samples printed by creating different combinations of 3D printing parameters has increased, and research in this field is becoming increasingly important. In this context, ML algorithms accelerate process optimization by reducing the costs of experimental studies and contributing to more environmentally friendly production processes.

2.2. Database Selection

The selection of WoS as the primary database for the literature review was made carefully to align with the study’s objectives. WoS is a widely accepted and reputable database that covers leading peer-reviewed journals with high accuracy [48]. Compared to other databases, it offers superior bibliographic detail and record quality [49]. Its records significantly overlap with those of other well-known databases such as Embase, MEDLINE, and Google Scholar, while its indexing process involves expert human oversight rather than being fully automated [48]. Furthermore, WoS successfully reflects the continuously growing volume of scientific publications [50]. Additionally, comparative sampling showed that relevant studies found in other databases, such as IEEE Xplore and PubMed, are largely also included in WoS. Therefore, limiting the review to WoS helped avoid data duplication and enabled a more focused analysis.

2.3. Inclusion and Exclusion Criteria

A total of 90 publications were identified in the WoS database using the defined keywords. Among these, four publications were excluded due to restricted access [51,52,53,54]. Three were excluded as they were conference proceedings. Consequently, 83 peer-reviewed journal articles were included in the scope of this study. All the accessed studies were examined, and the most frequently used polymeric materials are presented in detail in Table 1. The studies focusing on non-polymeric materials are discussed separately under Section 3.3 (Machine Learning Applications in Non-Polymeric 3D-Printed Materials).

Figure 1. Publication statuses on a yearly basis as a result of the publication scan of WoS data with the keywords 3D printing parameters, machine learning, tensile strength, roughness, and hardness [55,56,57].

Figure 1. Publication statuses on a yearly basis as a result of the publication scan of WoS data with the keywords 3D printing parameters, machine learning, tensile strength, roughness, and hardness [55,56,57].

2.4. Study Selection Process

The 83 included articles were systematically analyzed to extract relevant information on the applied materials, ML algorithms, investigated printing parameters, and the mechanical and physical test methods performed. In this process, particular attention was paid to the frequency of these parameters across the literature. Their distribution was visualized through graphs for comparative evaluation. Studies included for detailed analysis in Table 1 were selected based on their focus on the most commonly used polymeric materials in FDM research (e.g., PLA, ABS, PETG, and their composites), as identified during our initial literature search. Studies focusing on non-polymeric materials or less common polymers are reserved for narrative discussion in Section 3.3 to provide a comprehensive overview of the field while maintaining the focus and clarity of the table.

The temporal evolution of the publications, based on yearly distribution, is illustrated in Figure 1. The figure indicates that studies focusing on tensile strength have grown rapidly since 2022, roughness studies have shown continuous growth after 2022, while hardness studies have demonstrated a relatively slower increase. This trend highlights the rising academic interest in ML-based predictions of test results in 3D-printed samples. The study identification and selection process (Figure 2) was adapted in line with PRISMA-style reporting principles, but the present work was structured as a narrative literature review rather than a systematic or scoping review.

3. Literature Review and Analysis of Existing Studies

These publications were analyzed in terms of the type of material used, the algorithms used, the printing parameters, and the tests applied. It was seen that polymer materials were used the most in the examined studies (Figure 3). As shown in the right pie chart of Figure 3, the percentages represent the share of each polymer type relative to the entire dataset, and therefore their sum corresponds to 59%, which is the overall contribution of polymers among all materials. While PLA was used the most among polymer materials, ABS was preferred less than PLA.

PLA is a biodegradable, environmentally friendly thermoplastic polymer widely used in FDM 3D printing due to its low cost, ease of processing, and suitable mechanical properties. Its relatively low melting point and printing temperature facilitate faster and simpler manufacturing, making it ideal for both research and industrial applications [58]. The high use of PLA in the studies is due to the fact that it is an environmentally friendly polymer. In addition, PLA’s low melting point and printing temperatures offer ideal properties for 3D printers. This makes the printing process faster and easier. PLA’s mechanical properties, especially its durability and hardness, allow it to be used in a wide range of applications. In addition, it is suitable for many industrial and consumer-based applications due to its odorless and biodegradable nature. All these factors are the main reasons why PLA is preferred in 3D printing research and applications [59]. The reason why ABS is preferred less, despite its good hardness, is that there are some usage restrictions. It requires higher printing temperatures, which can be more energy-consuming and challenging for some open printers. In addition, it can cause problems with shrinkage and deformation during printing due to high temperatures [60]. Another disadvantage is that it emits an unpleasant odor during printing [61]. This requires users to provide appropriate ventilation. In addition, the fact that ABS material is not environmentally friendly and is not biodegradable is a significant disadvantage in terms of environmental sustainability [62]. For these reasons, the widespread use of ABS is more limited and is generally preferred in special conditions.

3.1. Frequently Optimized Parameters in 3D Printing

In order to establish a clear methodological foundation for ML applications in 3D printing, this section systematically presents the most frequently investigated printing parameters and the corresponding output tests employed in the literature. These parameters, which serve as inputs for ML model training and the associated mechanical or surface quality tests considered as outputs, are essential for understanding how data is structured in the development of predictive models. Furthermore, the statistical distribution of these parameters across the literature (Figure 4 and Figure 5)provides critical insights into their prevalence and contextual usage. Since the performance of ML algorithms is not only determined by the quality of data but also by the extent to which these parameters are addressed in prior research, this section establishes an analytical framework that directly supports the overall purpose of the article by linking parameter structures with ML implementations.

Among the most commonly studied factors influencing print efficiency in FDM-based 3D printing are layer thickness, nozzle temperature, print speed, infill density, and extrusion temperature (Figure 4).

Since FDM technology builds parts layer by layer, each layer plays a critical role in the final output. Key aspects such as the height of each layer and its uniformity are closely monitored. These parameters influence both the quality and mechanical characteristics of the printed part, while also offering important production data, such as the total time needed for fabrication [63]. As layer thickness decreases, more layers are required to complete the print, which in turn increases the overall production time of the part. In FDM, layer thickness typically ranges between 0.17 mm and 0.33 mm. Thinner layers can improve surface finish and detail resolution, but also lead to longer print times, while thicker layers reduce printing time at the cost of surface quality and precision [64]. Infill density refers to the percentage of internal material that is printed within a 3D object. It has a direct effect on the part’s strength, weight, and the time required for printing. A higher infill density typically results in a stronger but heavier and longer-to-print object, while a lower density reduces weight and print time but may compromise structural integrity. Additionally, various infill patterns or styles can influence the overall strength of the print without necessarily increasing the amount of filament used or altering the print’s weight [63]. Print speed refers to how quickly the printer’s nozzle deposits material onto the build surface over time. In general, FDM printers are not well-suited for extremely high-speed printing, as it can lead to defects in quality and layer adhesion. For this reason, more moderate print speeds are commonly recommended. Speeds such as 40 mm/s and 50 mm/s are considered to be within the lower range and are often suggested to maintain print accuracy and surface quality [65]. Print temperature, also known as extruder temperature, refers to the heat setting of the printer’s nozzle that melts and extrudes the filament. Since each material has its own melting point, the ideal print temperature varies accordingly. For instance, ABS typically requires a range between 230 °C and 250 °C, whereas PLA is best printed between 190 °C and 220 °C. To ensure smooth extrusion and optimal part quality, the print temperature should be set at or just above the material’s melting point. This helps avoid issues such as poor layer adhesion or nozzle clogging [66]. The internal structure of the print is shaped by the chosen infill design. It can be found in structures such as triangles, grids, and honeycombs. The raster pattern refers to the path or arrangement of extruded filament lines on the build platform during 3D printing. Structurally, it defines how the material is laid down within the object, influencing its internal configuration and mechanical behavior. This pattern is aligned according to the orientation of the part in the printer, specifically along the three primary build directions: X, Y, and Z. The chosen raster pattern plays a key role in determining the strength, surface finish, and overall stability of the printed component [67]. Figure 6 shows a schematic representation of 3D printing performance.

The geometric model standard dimensions of 3D-printed samples, prepared according to the test conditions for molding and extrusion plastics in ISO 527-2 [68], provide the necessary criteria for accurate testing of these samples. The geometric model standard dimensions of the 3D-printed samples are presented in Figure 7. The application rates of mechanical, physical, and thermal tests applied to the samples prepared according to these standards are shown in Figure 5. Mechanical tests aim to evaluate the durability and performance of the materials and include criteria such as tensile strength, hardness, bending strength, elongation at break, and elastic modulus. In physical tests, characteristics such as surface roughness are analyzed to provide information about the material’s surface quality. Among the tests, tensile strength, hardness, and roughness tests are the most frequently studied and investigated.

Tensile strength tests are performed to assess the maximum stress a material can endure before it fractures or permanently alters. This method is widely applied to understand the mechanical characteristics of various materials, such as metals, polymers, composites, and fabrics. In this test, a specimen is pulled under a steadily increasing tensile load until it ultimately breaks. Throughout the process, the material’s deformation and strain are closely monitored. The data obtained from the test also enables the calculation of important mechanical properties, such as the material’s stiffness, yield point, and strain before failure [43]. Tensile stresses and elasticity modulus programming are calculated by Theorems (1) and (2). σ represents the tensile strength at yield (N/mm² or MPa); A, Section (mm²); F, applied load (kN); E, elasticity modulus (MPa); and $\epsilon$ represents the tension of the sample [70].

$$\sigma ={\displaystyle \frac{F}{A}}$$

(1)

Theorem 1.

Tensile stress calculations.

$$E={\displaystyle \frac{\sigma }{\epsilon }}$$

(2)

Theorem 2.

Modulus of elasticity calculations.

Hardness measurement tests are important for evaluating product performance. It is used in the distribution, classification, and quality control of products. Microfracture parameters are a non-destructive technique for evaluating the mechanical behavior of materials. Polymer options are widely used [71]. Digital Shore D durometer with a 0.1 mm indentation needle is widely used for hardness testing. Hardness measurements are conducted on both the flat and edge areas of the printed samples. This approach helps evaluate the uniformity and consistency of material properties across different areas of the object, offering a more comprehensive understanding of its mechanical performance [72]. An example of the efficiency images is presented in Figure 8.

Surface roughness testing is conducted to evaluate how much a surface deviates from its perfectly smooth or ideal form. This type of measurement is particularly important in engineering and manufacturing, as surface texture can directly impact the functionality, durability, and overall quality of a product. Typically, the process involves guiding a sensor or probe across the surface of the material to detect and record any irregularities. The collected data is then analyzed and expressed through specific parameters such as Ra (average roughness), Rq (root mean square roughness), and Rz (average maximum height of the profile) [43]. Surface roughness can be measured using two main approaches: contact and non-contact methods. Non-contact methods offer several advantages over traditional contact techniques, including lower cost, higher measurement accuracy, and greater efficiency. These methods typically use optical or laser-based sensors to evaluate surface texture without physically touching the material, reducing the risk of damaging delicate surfaces and allowing for faster data collection [73]. Each sample’s surface roughness is measured using a surface profilometer, which scans the surface to capture precise details about its texture and irregularities. The surface roughness is obtained by repeated testing three times in different ways on the same sample with the surface profilometer. The values are then used for the range of the roughness of the surface of the sample, which ensures the durability of the gloss [74].

3.2. Machine Learning Algorithms Used for Parameter Optimization

Predictive modeling is a statistical approach that uses historical data and ML algorithms to predict future outcomes. These models capture complex, nonlinear patterns in manufacturing processes, providing valuable insights for predicting equipment failures, automating quality control, and forecasting demand. Models generally fall into two main categories:

• Classification models, which predict class membership (for example, labeling a part as “defect” or “non-defect”).
• Regression models, which predict a numerical value (for example, predicting a material’s tensile strength in MPa).

ML, which stands out in this case, refers to a set of computational methods that allow systems to learn patterns from data and make predictions without being explicitly programmed. ML methods are generally categorized as supervised, unsupervised, and reinforcement learning [75,76]. Supervised learning, which involves training models on labeled input–output pairs, is commonly applied for tasks such as regression and classification. Popular supervised algorithms include Random Forest (RF) and Support Vector Machines (SVM), known for their robustness and accuracy in handling structured data [77]. On the other hand, Artificial Neural Networks (ANN) and Deep Learning (DL) approaches have gained importance in recent years due to their ability to extract complex, nonlinear relationships from large-scale datasets [78]. Overall, ML provides powerful tools for data-driven decision-making and optimization across a wide range of scientific and engineering domains. In the context of 3D printing parameter optimization, these methods have been increasingly applied.

Among the different types of algorithms examined in the reviewed studies, regression algorithms were the most frequently used, with DL approaches and ANN models following closely behind (Figure 9). ANNs, designed to simulate the structure and function of biological neural systems, enable the analysis of complex data [79].

In the field of materials science, AI helps streamline the development of novel materials and the enhancement of current materials, enabling faster and more efficient material design and refinement [80]. AI approaches, such as ANNs, DL algorithms, and optimization strategies, have the ability to analyze material data and create forecasting models that support the design and optimization of materials and their parameters [81,82,83]. Nikzad et al. carried out an in-depth investigation into the tensile strength of PLA specimens produced through the FDM method. They assembled an extensive dataset comprising 329 samples sourced from 16 scholarly publications. To assess performance, they applied 19 distinct ML algorithms. The authors proposed that future studies could extend this methodology to incorporate a wider variety of materials, such as high-performance polymers, composite materials, and metals, in order to better assess the versatility of ML models. Moreover, they highlighted that integrating additional factors—such as environmental influences and printing patterns—could enhance prediction accuracy for real-world industrial applications. In practical terms, reliable predictive models like those presented in their research can help optimize material behavior and reduce the reliance on trial-and-error methods in manufacturing. This can lead to more efficient product development and tailored production processes in industries such as aerospace, automotive, and biomedical engineering [84].

The fact that only three comprehensive studies were found in the WoS database in a search conducted with five different keywords, namely 3D printing parameters, ML, tensile, hardness, and roughness, shows that there is a significant gap in the literature in this area and that more research should be conducted. The research by Selvan et al. examines the optimization of manufacturing parameters and the prediction of surface quality for PLA material in 3D printing. In their research, samples were produced using an open-source 3D printer exclusively with PLA. A total of 27 specimens underwent tensile and surface roughness testing. The parameters analyzed during printing included infill density, print speed, and layer thickness. For the prediction phase, they employed ANN as the sole ML method. The ANN model achieved an accuracy rate of 98% [43]. Khusheef et al. conducted a study on DL prediction of 3D-printed plastic part properties with thermographic and vibration data. They developed various combinations of parameters, including layer thickness, infill pattern, density, nozzle temperature, print speed, and thermal data. In the study, they printed 27 samples on an open 3D printer using only PLA material. Each sample was printed twice to prove the reliability of the tests. A total of 57 samples were tested for surface roughness, tensile strength, elongation at break, micro surface hardness, and warpage. The collected data was analyzed with the Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) hybrid algorithm. While the best performance was achieved with 99% accuracy in tensile strength, micro-surface hardness and warpage predictions were less accurate than other features [83].

Butt and Mohaghegh introduced an innovative methodology that combines the digital twin approach with ML algorithms to optimize process parameters in the FDM method. They printed 40 samples using PLA material, exploring various combinations of layer height, extrusion temperature, and printing speed. The study focused on analyzing the effects of these parameters on surface roughness, hardness, and tensile strength. The digital twin, developed using ANSYS 2021 Twin Builder, was validated against experimental data and effectively modeled the complex structure of the printing process. By integrating simulation data with experimental results, Convolutional Neural Networks (CNNs) and RF algorithms were employed to establish correlations between input and output parameters. The findings indicated that RF provided higher accuracy rates, ranging from 88% to 95% [85].

Ref. [86] presents the prediction performance of different ML models for tensile strength of 3D-printed parts, as illustrated in Figure 10. The scatter plots compare the actual and predicted values for both training and testing datasets. Among the models evaluated, XGBoost demonstrates the closest agreement with experimental results, indicating its strong potential for predicting mechanical properties. Nevertheless, certain deviations—particularly in the test data—highlight that the models are not yet fully robust. This suggests that while ML-based approaches offer promising accuracy, further refinement and larger datasets are required before they can be reliably adopted for industrial-scale applications.

Table 1 displays the studies conducted on polymer materials, which include details about the material type, printing parameters, tests performed, ML algorithms, sample sizes, and the study results.

Table 1. Machine learning-supported studies in 3D printing technologies.

Studies	ML Algorithms	Printing Parameters	Estimated Output Parameters	Number of Samples	Material Type	Algorithm Success Rates
Nikzad et al., 2025 [84]	CatBoost XGBoost Gradient Boosting Machine Light Gradient Boosting Machine	Nozzle diameter Raster direction Infill density Layer thickness Print speed Nozzle temperature	Ultimate tensile strength	329	PLA	With the integration of the Ensemble method into the CatBoost algorithm, the accuracy of the model increased to R2 = 98.05%.
Ali et al., 2024 [87]	Random Forest Regression	Flow rate Layer thickness Nozzle temperature	Elongation at fracture Tensile strength	27	PLA-APHA Biodegradable blend	The success of the model was determined as 96% accuracy rate.
Hossain et al., 2024 [88]	Polynomial Regression	Extrusion temperature Print speed	Extrusion temperature Tensile strength Elongation at fracture	4	PLA Recycled plastics HDPE TiO2	With the addition of TiO2, durability and insulation properties were improved; in addition, 4.29% elongation at fracture was obtained.
Hossain et al., 2022 [89]	Linear Regression	Nozzle temperature Layer thickness Printing speed	Tensile strength Surface Roughness Thermal stability	6	PLA HDPE Recycled plastic	The ML-based prediction model predicted the nozzle temperature with 99.36% accuracy.
Sharma et al., 2024b [90]	Genetic Algorithm Particle Swarm Optimization Raven Search Optimization Differential Evolution	Coating solution concentration Immersion time Stirrer speed Infill density	Bending Tensile strength	32	PLA	Genetic Algorithm provided the closest predictions to the experimental results and achieved the highest mechanical properties. Differential Evolution and Response Surface Methodology (RSM) approaches also provided successful results.
Sharma et al., 2024a [91]	Gradient Boosting Regression	Infill density Layer height Wall thickness Printing speed	Tensile strength Flexural strength	100	PLA	Gradient Boosting Regression achieved the best accuracy with R2 value over 92% and resulted in minimal RMSE and MAE values.
Tănase et al., 2024 [92]	Random Forest Gradient Boosting AdaBoost Neural Network Support Vector Machine K-Nearest Neighbors Linear Regression	Layer thickness Fill ratio	Young’s modulus, Ultimate tensile strength elongation at break	54	Recycled-PLA	The Random Forest model provided the best predictions with 99% accuracy.
Tandon et al., 2024 [93]	Nonlinear Regression	Fill density Print orientation material type	Tensile strength Elastic modulus	20	ABS PLA PLA + CF	Model accuracy was over 90%.
Kharate et al., 2024 [94]	XGBoost	Biochar content Layer thickness Raster angle Filling pattern and density	Ultimate tensile strength Flexural strength Impact strength	16	PLA composites enhanced with biochar reinforcement	The XGBoost model achieved an accuracy of 96%.
Selvan et al., 2024 [43]	Artificial Neural Network	Infill density Printing speed Layer thickness	Surface roughness Tensile strength	27	PLA	The ANN predicted the surface quality with 98% accuracy.
Marappan et al., 2024 [95]	Support Vector Machine	Infill percentage Print speed and pattern type	Tensile strength Hardness	9	PLA CF composite	The SVM model was found to be successful in predictions with an accuracy rate of 95.65%.
Manola et al., 2024 [96]	Classification Regression Tree	Printing temperature Infill pattern Infill density	Melt flow index Ultimate tensile strength	27	ABS Al- and Cu-reinforced composite	The model showed an accuracy rate of 91.54%.
Ghasemi et al., 2024 [97]	Artificial Neural Network Extreme Gradient Boosting Decision Tree Gradient Boosting Tree Random Forest CatBoost	Layer thickness Nozzle diameter Fill rate Print speed Nozzle temperature Table temperature	Nonlinear bending Torsional behaviors Critical buckling loads Buckling mode changes depending On cell angle Size ratios	1926	PLA biopolymer	The Random Forest model provided the highest success in buckling predictions with MSE = 13.31 and R2 = 99.9%, while CatBoost showed superiority in bending predictions with MSE = 0.00094 and R2 = 99.9%.
Ulkir et al., 2023 [98]	Gaussian Process Regression Support Vector Machine Linear Regression Decision Tree Regressor	Layer thickness Infill volume Table temperature Printing direction	Energy consumption (Wh)	81	ABS	Gaussian Process Regression model provided high accuracy in energy consumption with R2 = 0.99 and MAE = 0.016.
Jain et al., 2024 [99]	Linear, Lasso, Ridge Regression Decision Tree Random Forest Gradient Boost Extreme Gradient Boost Adaptive Boost Regression	Layer thickness Raster orientation Feed rate	Flexural strength	54	MWCNT takviyeli PLA	The XGBR model showed the highest performance in flexural strength with RMSE = 1.776, MAE = 1.366, and R2 = 0.99.
Khusheef et al., 2024 [83]	CNN-LSTM	Layer thickness Infill pattern Infill density Nozzle temperature Print speeds Thermal and IMU data	Surface roughness Tensile strength Elongation at fracture Microsurface hardness Warpage	27	PLA	The best estimate of tensile strength was achieved with 99% accuracy; microsurface hardness and warpage estimates showed lower accuracy than other properties.
Ege et al., 2023 [100]	XGBoost Random Forest Artificial Neural Network Light Gradient Boosting Machine Ordinary Least Squares Regression	Nozzle temperature Printing speed Infill density Layer height	Tensile strength	68	PLA	The XGBoost model provided R2 = 0.98 on training data and R2 = 0.94 on test data.
Ziadia et al., 2023 [101]	XGBoost Random Forest	Printing temperature Layer thickness Printing speed	Tensile strength Young’s modulus Stress at break	54	PLA PLA-CF	The parameters optimized by Genetic Algorithm provided high mechanical properties for PLA-CF material. The best tensile strength was obtained with a temperature of 222.28 °C, a layer thickness of 0.261 mm, and a printing speed of 40.30 mm/s.
Le et al., 2023 [102]	Surrogate Neural Network Model	Layer thickness Printing direction Infill density	Flexure strength Bending stiffness	-	PLA	With the finite element model, the error was 6.13%, the highest strength was provided with a low layer thickness, and the bending stiffness was highest at 0.1 mm layer thickness.
Butt & Mohaghegh, 2023 [85]	CNN Random Forest	Layer height Extrusion temperature Speed	Surface roughness Hardness Tensile strength	40	PLA	The Random Forest classifier provided higher accuracy based on digital twin simulation data, while the CNN model performed lower.
Mishra & Jatti, 2023 [103]	Q- Learning (Reinforcement Learning)	Printing speed Extrusion temperature Layer height Infill percentage	Impact strength Bending strength Tensile strength	31	PLA	The Q-Learning algorithm was successful in parameter optimization and predicted the tensile strength with an error rate of 0.66%, flexural strength with an error rate of 4.37%, and impact strength with an error rate of 10.2%, which is close to the experimental results.
Ogunsanya et al., 2023 [104]	Multilayer Perceptron	Layer thickness Printing direction Extrusion temperature Build temperature printing speed	Dimensional accuracy Porosity Tensile strength	243	PLA	Three neurons, 0.0001 learning rate, and 5000 epochs were determined as the best hyperparameters for the MLP model. The dimensional accuracy of the model was evaluated with RMSE = 1.7, and the optimal learning rate was determined as 0.0001.
Grozav et al., 2023 [105]	Artificial Neural Network	Nozzle temperature Print speed Layer orientation	Tensile strength Stress–strain behavior	48	PLA	The ANN model predicted the tensile strength with 93% accuracy.
Jayasudha et al., 2022 [86]	XGBoost Gradient Boost AdaBoost Random Forest Linear Regression	Extrusion temperature Layer height Shell thickness	Tensile strength	27	PLA	The XGBoost model provided the best performance in tensile strength prediction with 94.6% R2.
Silva et al., 2022 [106]	Artificial Neural Network Genetic Algorithms	Printing speed Fill rate Extrusion temperature Filament thickness Extrusion direction	Tensile strength	149	PLA	The R2 value of the model is over 90%.
Charalampous et al., 2022 [107]	Random Forest Support Vector Regression K-Nearest Neighbors	Layer height Print speed Print temperature	Tensile strength	125	PLA	The KNN model provided the best prediction accuracy with the combination of low layer thickness, moderate printing speed, and 210 °C printing temperature.
Jatti et al., 2022 [108]	Nonlinear Regression	Infill percentage Layer thickness Printing speed Extrusion temperature	Tensile strength Impact strength bending strength	31	PLA	Nonlinear regression predicted the flexural strength with an error rate of 3.474%.
Soundararajan et al., 2025 [109]	Random Forest J48 (Decision Tree) Naive Bayes	Layer height Print speed Nozzle temperature	Surface roughness	42	PLA Plus	Application of Random Forest and J48 algorithms for surface roughness estimation provided accuracy rates above 90% for both models.
Zhu et al., 2024 [74]	Gradient-Boosting Decision Trees 1D-CNN	Layer thickness Scanning speed Material flow Nozzle temperature	Surface roughness	625	PLA	The FE-GBDT model exhibited high performance in surface roughness prediction, providing accuracy above 99%.
Sangwan et al., 2023 [110]	Autoregressive, Regression Models	Layer thickness Infill density Extruder temperature Table temperature Scale	Carbon footprint Printing time Surface roughness	27	PLA	The model was able to predict the remaining life of the nozzle with 96.96% accuracy.
Z. Wang et al., 2024 [81]	Particle Swarm Optimization Genetic Algorithm Multilayer Perceptron Support Vector Regression	Fill density Nozzle temperature Layer thickness	Hardness Material mixture properties	286	PLA TPU PETG ABS	The GA-MLP-3D model provided successful results with an R2 accuracy rate of 87.1%.
Ozdemir et al., 2024 [111]	Random Forest	Nozzle temperature Extrusion flow Layer height	Print quality (weight deviation, inner diameter, surface smoothness)	186	PLA PLA/NC PLA/GNP	The RF demonstrated successful performance with an accuracy rate of 92.8% in estimating the weight deviation.

Table 1. Machine learning-supported studies in 3D printing technologies.

Studies	ML Algorithms	Printing Parameters	Estimated Output Parameters	Number of Samples	Material Type	Algorithm Success Rates
Nikzad et al., 2025 [84]	CatBoost XGBoost Gradient Boosting Machine Light Gradient Boosting Machine	Nozzle diameter Raster direction Infill density Layer thickness Print speed Nozzle temperature	Ultimate tensile strength	329	PLA	With the integration of the Ensemble method into the CatBoost algorithm, the accuracy of the model increased to R2 = 98.05%.
Ali et al., 2024 [87]	Random Forest Regression	Flow rate Layer thickness Nozzle temperature	Elongation at fracture Tensile strength	27	PLA-APHA Biodegradable blend	The success of the model was determined as 96% accuracy rate.
Hossain et al., 2024 [88]	Polynomial Regression	Extrusion temperature Print speed	Extrusion temperature Tensile strength Elongation at fracture	4	PLA Recycled plastics HDPE TiO2	With the addition of TiO2, durability and insulation properties were improved; in addition, 4.29% elongation at fracture was obtained.
Hossain et al., 2022 [89]	Linear Regression	Nozzle temperature Layer thickness Printing speed	Tensile strength Surface Roughness Thermal stability	6	PLA HDPE Recycled plastic	The ML-based prediction model predicted the nozzle temperature with 99.36% accuracy.
Sharma et al., 2024b [90]	Genetic Algorithm Particle Swarm Optimization Raven Search Optimization Differential Evolution	Coating solution concentration Immersion time Stirrer speed Infill density	Bending Tensile strength	32	PLA	Genetic Algorithm provided the closest predictions to the experimental results and achieved the highest mechanical properties. Differential Evolution and Response Surface Methodology (RSM) approaches also provided successful results.
Sharma et al., 2024a [91]	Gradient Boosting Regression	Infill density Layer height Wall thickness Printing speed	Tensile strength Flexural strength	100	PLA	Gradient Boosting Regression achieved the best accuracy with R2 value over 92% and resulted in minimal RMSE and MAE values.
Tănase et al., 2024 [92]	Random Forest Gradient Boosting AdaBoost Neural Network Support Vector Machine K-Nearest Neighbors Linear Regression	Layer thickness Fill ratio	Young’s modulus, Ultimate tensile strength elongation at break	54	Recycled-PLA	The Random Forest model provided the best predictions with 99% accuracy.
Tandon et al., 2024 [93]	Nonlinear Regression	Fill density Print orientation material type	Tensile strength Elastic modulus	20	ABS PLA PLA + CF	Model accuracy was over 90%.
Kharate et al., 2024 [94]	XGBoost	Biochar content Layer thickness Raster angle Filling pattern and density	Ultimate tensile strength Flexural strength Impact strength	16	PLA composites enhanced with biochar reinforcement	The XGBoost model achieved an accuracy of 96%.
Selvan et al., 2024 [43]	Artificial Neural Network	Infill density Printing speed Layer thickness	Surface roughness Tensile strength	27	PLA	The ANN predicted the surface quality with 98% accuracy.
Marappan et al., 2024 [95]	Support Vector Machine	Infill percentage Print speed and pattern type	Tensile strength Hardness	9	PLA CF composite	The SVM model was found to be successful in predictions with an accuracy rate of 95.65%.
Manola et al., 2024 [96]	Classification Regression Tree	Printing temperature Infill pattern Infill density	Melt flow index Ultimate tensile strength	27	ABS Al- and Cu-reinforced composite	The model showed an accuracy rate of 91.54%.
Ghasemi et al., 2024 [97]	Artificial Neural Network Extreme Gradient Boosting Decision Tree Gradient Boosting Tree Random Forest CatBoost	Layer thickness Nozzle diameter Fill rate Print speed Nozzle temperature Table temperature	Nonlinear bending Torsional behaviors Critical buckling loads Buckling mode changes depending On cell angle Size ratios	1926	PLA biopolymer	The Random Forest model provided the highest success in buckling predictions with MSE = 13.31 and R2 = 99.9%, while CatBoost showed superiority in bending predictions with MSE = 0.00094 and R2 = 99.9%.
Ulkir et al., 2023 [98]	Gaussian Process Regression Support Vector Machine Linear Regression Decision Tree Regressor	Layer thickness Infill volume Table temperature Printing direction	Energy consumption (Wh)	81	ABS	Gaussian Process Regression model provided high accuracy in energy consumption with R2 = 0.99 and MAE = 0.016.
Jain et al., 2024 [99]	Linear, Lasso, Ridge Regression Decision Tree Random Forest Gradient Boost Extreme Gradient Boost Adaptive Boost Regression	Layer thickness Raster orientation Feed rate	Flexural strength	54	MWCNT takviyeli PLA	The XGBR model showed the highest performance in flexural strength with RMSE = 1.776, MAE = 1.366, and R2 = 0.99.
Khusheef et al., 2024 [83]	CNN-LSTM	Layer thickness Infill pattern Infill density Nozzle temperature Print speeds Thermal and IMU data	Surface roughness Tensile strength Elongation at fracture Microsurface hardness Warpage	27	PLA	The best estimate of tensile strength was achieved with 99% accuracy; microsurface hardness and warpage estimates showed lower accuracy than other properties.
Ege et al., 2023 [100]	XGBoost Random Forest Artificial Neural Network Light Gradient Boosting Machine Ordinary Least Squares Regression	Nozzle temperature Printing speed Infill density Layer height	Tensile strength	68	PLA	The XGBoost model provided R2 = 0.98 on training data and R2 = 0.94 on test data.
Ziadia et al., 2023 [101]	XGBoost Random Forest	Printing temperature Layer thickness Printing speed	Tensile strength Young’s modulus Stress at break	54	PLA PLA-CF	The parameters optimized by Genetic Algorithm provided high mechanical properties for PLA-CF material. The best tensile strength was obtained with a temperature of 222.28 °C, a layer thickness of 0.261 mm, and a printing speed of 40.30 mm/s.
Le et al., 2023 [102]	Surrogate Neural Network Model	Layer thickness Printing direction Infill density	Flexure strength Bending stiffness	-	PLA	With the finite element model, the error was 6.13%, the highest strength was provided with a low layer thickness, and the bending stiffness was highest at 0.1 mm layer thickness.
Butt & Mohaghegh, 2023 [85]	CNN Random Forest	Layer height Extrusion temperature Speed	Surface roughness Hardness Tensile strength	40	PLA	The Random Forest classifier provided higher accuracy based on digital twin simulation data, while the CNN model performed lower.
Mishra & Jatti, 2023 [103]	Q- Learning (Reinforcement Learning)	Printing speed Extrusion temperature Layer height Infill percentage	Impact strength Bending strength Tensile strength	31	PLA	The Q-Learning algorithm was successful in parameter optimization and predicted the tensile strength with an error rate of 0.66%, flexural strength with an error rate of 4.37%, and impact strength with an error rate of 10.2%, which is close to the experimental results.
Ogunsanya et al., 2023 [104]	Multilayer Perceptron	Layer thickness Printing direction Extrusion temperature Build temperature printing speed	Dimensional accuracy Porosity Tensile strength	243	PLA	Three neurons, 0.0001 learning rate, and 5000 epochs were determined as the best hyperparameters for the MLP model. The dimensional accuracy of the model was evaluated with RMSE = 1.7, and the optimal learning rate was determined as 0.0001.
Grozav et al., 2023 [105]	Artificial Neural Network	Nozzle temperature Print speed Layer orientation	Tensile strength Stress–strain behavior	48	PLA	The ANN model predicted the tensile strength with 93% accuracy.
Jayasudha et al., 2022 [86]	XGBoost Gradient Boost AdaBoost Random Forest Linear Regression	Extrusion temperature Layer height Shell thickness	Tensile strength	27	PLA	The XGBoost model provided the best performance in tensile strength prediction with 94.6% R2.
Silva et al., 2022 [106]	Artificial Neural Network Genetic Algorithms	Printing speed Fill rate Extrusion temperature Filament thickness Extrusion direction	Tensile strength	149	PLA	The R2 value of the model is over 90%.
Charalampous et al., 2022 [107]	Random Forest Support Vector Regression K-Nearest Neighbors	Layer height Print speed Print temperature	Tensile strength	125	PLA	The KNN model provided the best prediction accuracy with the combination of low layer thickness, moderate printing speed, and 210 °C printing temperature.
Jatti et al., 2022 [108]	Nonlinear Regression	Infill percentage Layer thickness Printing speed Extrusion temperature	Tensile strength Impact strength bending strength	31	PLA	Nonlinear regression predicted the flexural strength with an error rate of 3.474%.
Soundararajan et al., 2025 [109]	Random Forest J48 (Decision Tree) Naive Bayes	Layer height Print speed Nozzle temperature	Surface roughness	42	PLA Plus	Application of Random Forest and J48 algorithms for surface roughness estimation provided accuracy rates above 90% for both models.
Zhu et al., 2024 [74]	Gradient-Boosting Decision Trees 1D-CNN	Layer thickness Scanning speed Material flow Nozzle temperature	Surface roughness	625	PLA	The FE-GBDT model exhibited high performance in surface roughness prediction, providing accuracy above 99%.
Sangwan et al., 2023 [110]	Autoregressive, Regression Models	Layer thickness Infill density Extruder temperature Table temperature Scale	Carbon footprint Printing time Surface roughness	27	PLA	The model was able to predict the remaining life of the nozzle with 96.96% accuracy.
Z. Wang et al., 2024 [81]	Particle Swarm Optimization Genetic Algorithm Multilayer Perceptron Support Vector Regression	Fill density Nozzle temperature Layer thickness	Hardness Material mixture properties	286	PLA TPU PETG ABS	The GA-MLP-3D model provided successful results with an R2 accuracy rate of 87.1%.
Ozdemir et al., 2024 [111]	Random Forest	Nozzle temperature Extrusion flow Layer height	Print quality (weight deviation, inner diameter, surface smoothness)	186	PLA PLA/NC PLA/GNP	The RF demonstrated successful performance with an accuracy rate of 92.8% in estimating the weight deviation.

3.3. Machine Learning Applications in Non-Polymeric 3D-Printed Materials

In addition to the studies presented in Table 1, studies other than PLA, PETG, HDPE, ABS, and additive PLA were examined. Khan et al. used Portland Cement, Fly Ash, Silica Fume, GGBFS, and Recycled Aggregates as materials in their study. The ANN, RF, and SVM methods were compared with ML methods. Nozzle speed, layer height, sand–binder ratio, and extrusion pressure were evaluated as printing parameters. Using these parameters, compressive strength, tensile strength, flexural strength, and bond strength were estimated with ML methods. ML models were found to be effective in improving printing quality and optimizing material usage in 3D concrete printing. They stated that ecological impacts can be reduced by 30–40% with the use of new materials [112]. In the study conducted by Ege and Arıcı, carboxymethyl cellulose, gelatin, and graphene oxide composite hydrogels were used. Ultimate tensile strength and strain at break were estimated using the XGBoost regression method with Shapley Additive Explanations analysis. In this study conducted on 15 samples, the modulus and ultimate tensile strength increased with N-hydroxysuccinimide and 3-dimethylaminopropyl carbodiimide, while flexibility decreased; carboxymethyl cellulose increased flexibility and decreased strength. Flexibility and strength increased with the addition of up to 1% graphene oxide, but higher rates of graphene oxide caused a significant decrease in mechanical properties [113]. Prada Parra et al. studied carbon, Kevlar, and glass fiber-reinforced composites with a nylon matrix. In the study conducted on 127 samples, the researchers estimated the elastic modulus and tensile strength using parameters such as fiber content, fiber type, filler density, fiber placement angle, and the CatBoost and KNN algorithms. CatBoost estimated the elastic modulus with 98% accuracy. The model was suggested as a cost-reducing tool in mechanical property estimation [114]. In their study, Sattari et al. estimated the tensile strength, toughness, and glass transition temperature using Gaussian processes, RF, and monomer weight ratios, and physical–chemical parameters on samples produced using 127 thermoplastic resins. The printing failure rate was reduced to 3%, and the glass transition temperature out-of-range rate was reduced to 17%. Pareto optimal solutions were provided for tensile strength and toughness [115]. In their study, Luo et al. created a data set with 126 samples produced using Ti-6Al-4V material. They performed the estimation of tensile strength and elongation at break with the DCNN-ResNet101 architecture using laser power, scanning speed, and photodiode data. With ResNet101, tensile strength was estimated at 98.7% and elongation at break with 93.1% accuracy [116]. Other studies in this area have investigated various materials, such as 3D-printed concrete [117], Fe–Ni–Ti–Al steel [118], PDMS matrix with glass microfibers and hollow glass particles [119], Ti-6Al-4V alloy [120], AZ31 magnesium alloy [121], 316L stainless steel [122,123], HDPE foam [124], and fiber-reinforced polymers produced via FDM [125].

In addition to polymer-based studies, the effectiveness of ML in predicting the mechanical properties of non-polymeric materials has also been demonstrated. For instance, Zhang et al. developed an ANN model to describe the correlation between texture characteristics and tensile properties of AZ31 magnesium alloy. The model exhibited high predictive accuracy, with R² values exceeding 0.97 for yield strength, ultimate tensile strength, and elongation. Figure 11 illustrates the regression performance of the ANN in the training set, where the predicted and experimental values show excellent agreement [121].

In the study where surface roughness was evaluated, pectin fraction, CNC ratio, crosslinking density parameters, RSM, and linear regression were insufficient to predict the printing ability with 15 samples of polysaccharides. The RF model provided 88% accuracy with only rheological measurements [126]. Using resin material, layer thickness, filling density, printing angle, exposure time, and lifting speed printing parameters were evaluated on 32 samples as input. With these parameters, surface roughness was estimated with PSO and GA. The RSM-PSO model provided a minimum surface roughness of 0.2042 microns. PSO was found to be 4.84% more effective than GA [127]. In other studies on roughness, AlSi10Mg [128,129], steel [73], CoCr alloy [130], silver nanoparticles [131], PMMA, Ni-based powders [132], stainless steel [133], and 316L stainless steel [134] materials were used.

4. Results and Discussion

This review examines the effectiveness and potential of ML algorithms in predicting the mechanical and physical test results of materials. In recent years, with the increase in data-driven approaches in the manufacturing sector, the use of ML techniques has accelerated, and these methods have begun to stand out as an alternative to traditional experimental approaches. Literature reviews indicate that ML techniques play a significant role in enhancing the quality and performance of various polymer materials used in AM, including PLA, ABS, HDPE, PETG, and TPU. These approaches are widely used to optimize printing parameters, predict mechanical properties, and improve overall process efficiency, making them valuable tools for advancing material performance and reliability. ML algorithms reduce the number of experimental trials by improving manufacturing processes, reducing error rates, and minimizing process variability. Thus, manufacturing processes become more predictable, and costs are reduced by increasing automation in quality control processes.

Traditionally, optimization of printing parameters such as nozzle temperature, speed, and layer thickness has been carried out through experimental trials. While these approaches provide valuable insights and establish process baselines, they are often costly, time-consuming, and limited in their generalizability across different printers and materials.

Future research should increase the accuracy and generalization ability of algorithms by using more material types and manufacturing parameters and improving the quality of the data used. For this, model performance can be increased by creating comprehensive and high-accuracy data sets. Furthermore, incorporating sensors to monitor data throughout the printing process could facilitate the advancement of more adaptable systems. Considering environmental and sustainability factors will increase the potential for research in this area to provide social benefits. In particular, the integration of recycled materials into production processes can contribute to reducing environmental impacts. ML-based optimization techniques are considered an effective solution for both minimizing material waste and reducing carbon footprints.

The findings of this review show that ML techniques are moving from a supportive role to a core component of AM workflows. Compared with conventional trial-and-error experimentation, ML can shorten optimization cycles, reduce resource consumption, and improve the automation of production while enabling accurate prediction of mechanical and physical properties (e.g., tensile strength, hardness, surface roughness). These observations align with recent ML-centered studies in AM and related subfields.

In comparison to experimental optimization, ML-based strategies offer broader parameter exploration with fewer trials. For example, Selvan et al. [43] achieved 98% accuracy in predicting tensile strength of PLA samples using ANN, while Khusheef et al. [83] reported 99% accuracy with CNN-LSTM models combining process monitoring data. Such findings demonstrate that ML not only improves predictive accuracy but also accelerates optimization cycles and reduces material waste, positioning it as a powerful complement to traditional trial-and-error methods.

In particular, supervised learning remains dominant: decision trees and RF, SVM, and ANN frequently achieve strong performance in predicting test outcomes and in process parameter optimization. Recent examples include polymer and composite systems modeled via ANN-based and hybrid optimization frameworks, as well as roughness prediction in metals using RF and DL. The reviewed literature also reveals important distinctions in why certain ML algorithms outperform others in specific prediction tasks. The underlying reason lies in the nature of the data and the architecture of the algorithms. As shown in Table 1, the majority of studies employ structured, tabular data consisting of discrete numerical printing parameters such as temperature, speed, and layer thickness. For this type of input, tree-based ensemble methods like RF and Gradient Boosting are particularly powerful because they are less prone to overfitting with small datasets, handle nonlinear interactions effectively, and offer robust performance. By contrast, CNNs are architecturally designed for unstructured, spatial data such as images. Their strength lies in extracting features directly from pixels, which makes them especially effective for analyzing melt-pool images or detecting surface defects from camera feeds. For example, Butt & Mohaghegh [85] reported that RF outperforms CNNs, often relying on tabular digital twin data, where RF is naturally more suitable. Thus, algorithm superiority is not absolute but context-dependent: RF’s stronger performance reflects the alignment of model design with the data modality rather than an inherent weakness of CNNs. Future work should therefore align ML model selection with the type of data available (e.g., RF for parameter tables, CNNs for image streams).

A critical finding of this review is that “current research still shows limited evidence of direct improvements in efficiency and error reduction.” To clarify this point, it is useful to distinguish between ML for prediction and real-time, closed-loop control. Most studies reviewed employ ML primarily as a predictive tool: they demonstrate that, given a set of printing parameters, the resulting part properties (e.g., tensile strength) can be accurately forecast. This reduces the need for extensive trial-and-error experiments and indirectly enhances efficiency by narrowing the parameter search space before printing begins. By contrast, direct improvements in efficiency and error reduction arise from systems capable of detecting defects as they occur and correcting them in real time. Such approaches require closed-loop control, where ML algorithms analyze live sensor or camera data during printing and dynamically adjust process parameters such as material flow or nozzle speed. Only a limited number of studies have implemented such systems to date, highlighting why direct evidence remains scarce.

Beyond FDM, ML applications have also been critically explored in other major AM technologies. In SLA, ML has been applied to optimize orientation and exposure time to improve mechanical properties and surface roughness, while also predicting resin properties such as viscosity and tensile strength, thus accelerating material development [135,136,137,138]. In Powder Bed Fusion (SLS/PBF-LB/M), CNN-based approaches are widely used for in situ defect detection from powder bed images, achieving accuracies above 95% and enabling real-time quality assurance [139]. In DED, ML models link process parameters such as laser power and feed rate to droplet morphology, temperature fields, and residual stresses, reducing reliance on costly physics-based simulations. Physics-Informed Machine Learning (PIML) further improves predictive accuracy with less data by incorporating physical laws into the learning process [140,141,142]. At the nanoscale, techniques such as nanoimprint lithography and femtosecond laser-based microfabrication increasingly rely on ML for process control and defect detection, where manual inspection is infeasible [143,144]. Finally, advanced studies on metals—including tantalum via PBF-LB/M and ultrasonic vibration-assisted laser cladding (UVALC)—highlight the challenges of high processing temperatures and residual stresses, where ML-driven optimization is critical for achieving reliable and reproducible outcomes [145,146]. This broader perspective underscores that ML not only transforms FDM but also drives innovation across diverse AM technologies, contributing to parameter optimization, defect detection, and material development.

At the same time, advanced paradigms such as transfer learning and deep learning are emerging but remain comparatively under-explored across materials and printer types. Recent work demonstrates transfer learning from in situ process data to mechanical properties in laser powder bed fusion (Ti-6Al-4V), indicating a promising path to generalizable models that can reduce data demands when moving to new materials or machines. Another promising direction is the tighter integration of ML with information-rich signals and surface features acquired during or after printing. Such approaches support adaptive control, surface quality prediction, and, ultimately, Industry 4.0-ready smart manufacturing pipelines.

The corpus also highlights the importance of covering non-polymeric systems to ensure generalization. For example, fatigue life prediction for laser AM stainless steel and parameter optimization for 316L stainless steel illustrate how ML extends beyond polymers; surface roughness modeling in aluminum alloys further underscores applicability to metals. These non-polymer studies complement polymer-focused summaries (Table 1) and reflect the broader scope discussed in Section 3.3.

Overall, while experimental optimization remains essential for baseline validation, ML approaches substantially enhance efficiency, reproducibility, and predictive capability. A combined strategy—where experimental trials provide the foundation for training and validating ML models, and ML then guides further parameter refinement—offers the most effective pathway toward intelligent and adaptive AM workflows.

Overall, trends synthesized in this review (Figure 1) indicate a marked increase after 2022 in ML-assisted studies, with tensile strength and surface roughness attracting intense interest. To accelerate progress, future research should (i) develop generalizable models that transfer across printers, materials, and environments (e.g., via transfer learning); (ii) couple ML with real-time data streams and information-rich features for adaptive control; and (iii) report protocols and datasets with greater consistency to enhance reproducibility and cross-study comparability.

5. Conclusions

In this review, the effectiveness and potential of ML algorithms in predicting the mechanical and physical properties of materials used in 3D printing technologies have been comprehensively examined. With the increase in data-driven approaches in the manufacturing sector, ML techniques stand out as a strong alternative to traditional experimental methods. Beyond prediction accuracy, ML-assisted approaches offer clear advantages by reducing experimental costs, accelerating parameter selection, enabling systematic exploration of design spaces, and ultimately improving reproducibility and shortening the development cycle in AM research.

The literature shows that ML algorithms play a strong role in predicting final product quality and mechanical properties. While many studies remain prediction-oriented, some research also demonstrates the potential of ML to support process efficiency and defect reduction when combined with parameter optimization and in situ monitoring. Therefore, ML should be seen primarily as a predictive tool with promising, yet still emerging, implications for improving the printing process itself. Recent trends further highlight emerging directions such as transfer learning across different filament types, inverse modeling to determine optimal process parameters for targeted mechanical properties, and integration of ML with in situ monitoring systems for real-time defect detection. These advances suggest that ML gradually evolves beyond pure prediction towards a more comprehensive role in process optimization and adaptive control.

Future research is expected to be supported by experimental studies to test the applicability of the algorithms to a wider range of materials and printer types and to progress towards reducing uncertainties in the process by developing hybrid models. In addition, combining ML with optimization algorithms such as Bayesian optimization and genetic algorithms may accelerate the identification of optimal process windows. Establishing benchmark datasets and open repositories will also be crucial for reproducibility and broader adoption in the field. At the same time, ensuring standardization in preliminary preparation processes such as data collection, labeling, and processing will also accelerate the integration of these technologies into the industry. As a result, integrating ML algorithms into 3D printing is a harbinger of a new era in manufacturing technologies and a step towards more intelligent, efficient, and adaptive production systems.