Among the reviewed studies, the most prevalent approach for defect detection involved the use of AI-enhanced algorithms to analyze images of the AM process in real time. Although various image analysis methods have been explored, Convolutional Neural Networks (CNNs) emerged as the most widely implemented technique. A CNN is a deep learning architecture that extracts features from input images through multiple processing layers, enabling effective analysis of complex visual data [12]. The application of AI-based image analysis provides a level of consistency and reliability in defect detection that is often unattainable through manual inspection.

CNNs are most effective when trained using large datasets, which form the basis for their detection capabilities. An example of this is found in a study testing premade CNN sub-models for defect detection in Selective Laser Sintering (SLS). In this study, CNNs were utilized to process complex data structures, such as colored images. The process involved capturing video of the SLS procedure at two frames per second, then processing each frame in real-time to detect defects such as foreign bodies, part edges, powder accumulation, and powder trenches. To ensure the images could be processed accurately, they were first cropped to remove black pixels and unwanted areas, which is a necessary step despite its time-consuming nature. The study evaluated two models: Visual Geometry Group with sixteen layers (VGG16), and Xception, a newer model with thirty-two layers designed for computational efficiency. After training on SLS data, VGG16 outperformed Xception in all tests, achieving a peak accuracy of 97.1% and precision of 96.3% in defect detection. Additionally, VGG16 has less layers and can analyze images with less data, whereas Xception may have trouble with learning sufficiently without large datasets because of its network complexity [13]. However, when compared to models such as EfficientNetV2B0 and You Only Look Once (YOLO) version five, which were specifically tested for surface defect detection in the Powder Bed Fusion (PBF) process, both CNN models proved insufficient. EfficientNetV2B0 achieved 90% accuracy in detecting surface defects, 99% overall defect detection accuracy, and reduced training time. Similarly, YOLOv5 produced comparable results but had the added advantage of combining multiple images from different camera perspectives, which enhanced its ability to analyze content more comprehensively. This feature enabled YOLOv5 to detect defects within one second, proving to be more efficient than its competitors [14]. Additionally, updated versions of YOLOv5, YOLOv8, and YOLOv11 were evaluated for defect detection. YOLOv8 was employed to analyze infrared thermal images. Defects detected by infrared cameras often present as high heat signatures. These heat signatures are processed through three subsequent stages to clean and categorize the defects, further training the YOLOv8 network. The system achieved an impressive 99.3% accuracy in detecting the trained defects with a dataset smaller than 1,000 images [15]. YOLOv11 was evaluated using 2.58 million data parameters to train the model to detect scratches, cracks, warping, and material inconsistencies. The software was able to accurately classify defects across multiple materials and output precise locations, identifying where the defects reside [16].

Another CNN model used to scan layers AM processes for defect identification is U-Net. This model processes an image by first compressing it to focus on anomalies, followed by upsampling to restore the image to its original resolution. Through this process, the algorithm only analyzes regions of the image that were flagged to contain an anomaly, ultimately processing less data to improve computational speed [17]. These features make U-Net a highly suitable model for defect detection. U-Net was employed to detect defects in Selective Laser Melting (SLM) such as melt pools and demonstrated several unique advantages over the previously discussed models. The U-Net model could produce highly detailed images of abnormally melted pools, whereas in other models, variations in light and contrast interfered with image quality, preventing full data analysis. Furthermore, the U-Net model required less runtime and fewer parameters to identify defects, thereby minimizing training time and computational power necessary for real-time analysis within less than a second [18].

Another method to optimize algorithmic speed involves masking the image so only the desired geometry is analyzed, followed by converting the image to grayscale to eliminate the need for color processing. Using this approach, the AI algorithm tested achieved an accuracy of 97.5% [19]. By reducing image resolution and removing color information, the algorithm can operate significantly faster, making it better suited to deliver real-time results.

Additionally, two other methods were compared by analyzing powder bed fusion printing on a layer-by-layer basis. The VGG19 and SmallNet models were both trained using the same PBF image dataset to ensure accuracy in comparison. While VGG19 is a large network like VGG16 previously discussed, SmallNet is comparatively small. VGG19 performed slightly better with a 91% defect detection accuracy; however, because of its large networks size, it required 1,300,000 training images. Due to SmallNet’s small network size, it significantly reduced computational cost with similar accuracy within one percent. Considering both factors, SmallNet was determined to be the more favorable option due to its comparable accuracy and greater efficiency in processing images with minimal computational power. Moreover, a technique known as segmentation was applied to both models to further reduce analysis time. This method automatically assigns a label to each pixel within an image, classifying it into one of several predefined categories. Segmentation was then used to determine which areas of the image required examination by the algorithm, effectively filtering the data and reducing the area of interest by 40% before defect detection was initiated. The segmentation process was performed every five layers, optimizing both efficiency and accuracy [20].

Another approach, distinct from image-based detection, is sensor-based detection. These sensors play a crucial role in providing real-time feedback during the AM process, thereby improving overall print quality. Several types of sensors have been employed for this purpose, including accelerometer sensors to monitor printing nozzle movement, vibration sensors to detect anomalies in the printing process, acoustic sensors to identify breaks or unusual machine behavior, and temperature sensors to gather information on layer temperature. Different combinations of these sensors can be used to detect distinct types of defects; however, such detection is only possible with the integration of AI. By training these sensors using deep learning algorithms, manufacturers can obtain real-time insights not only into the visible aspects of the printing process but also into multiple mechanical and thermal conditions of the machine itself. This allows for monitoring of both the manufactured part and the operational state of the equipment, unlike traditional image-based detection systems [27].

An example of sensor-based detection involves using a CNN to analyze vibration data captured by sensors. To train the AI model, the sensor outputs data in the form of a matrix, which is then processed and categorized by the CNN to distinguish between normal and abnormal behavior. In one study, accelerometers were attached to both the print bed and the print nozzle, each transmitting signals every twenty milliseconds. This setup successfully detected a crack in the bottom layer of the part and identified surface roughness along one of the part edges. Although this method proved feasible, it required considerable processing time for each reading, which limited its real-time detection capabilities [28].

An alternative method involves pairing multiple sensors— such as vibration, temperature, and humidity sensors—to obtain more comprehensive data on printing conditions. In this setup, the vibration sensor provided readings every 0.1 seconds, while the temperature and humidity sensors sent readings at one-second intervals. Instead of using a CNN for data analysis, two Recurrent Neural Networks (RNNs) were employed [29]. RNNs process data similarly to CNNs but possess memory in the form of hidden layers that retain information from previous inputs, as illustrated in Fig. 5. This characteristic makes RNNs highly effective for identifying temporal dependencies within sensor data [30]. When tested, two RNN architectures—a Long Short-Term Memory (LSTM) network and a Gated Recurrent Unit (GRU)—were evaluated using vibration data in stereolithography (SLA) and digital light processing (DLP) processes. After training on approximately 500,000 data entries, both models achieved around 80% accuracy in defect detection. While the LSTM model slightly outperformed the GRU, both were constrained by the time required for data processing—approximately five minutes per cycle. This delay makes real-time defect detection difficult, limiting the models’ practical application in active monitoring systems [29].

Another method for sensor-based defect detection involves monitoring acoustic waves during the AM process. Whereas vibration sensors capture direct mechanical vibrations from the part or machine, acoustic sensors detect vibration waves traveling through the air. Some acoustic sensors are active, meaning they generate their own sound waves, which travel through the printed part to identify internal pores and defects by analyzing the reflected signals [1]. In one study, acoustic and vibration data were collected simultaneously to enhance defect detection accuracy. These datasets were analyzed using CNN, RNN, and other model types across various AM processes and machine configurations. Results showed that the CNN and RNN models consistently outperformed other approaches, accurately detecting defects in both the printed part and the machine’s mechanical systems. They were able to analyze nozzle flow and identify potential misalignments, predicting the likelihood of defects before they became visually apparent. Although these models demonstrated strong real-time predictive capability, they were computationally intensive, which limits their feasibility for manufacturing applications [31].

Table 1 summarizes and provides key comparisons of each AI model used for defect detection addressed in this paper.

Understanding the correlation between each input and output parameter, and adjusting the parameters accordingly is important in all AM processes. The optimization process improves the quality of the output and increases productivity, enabling the automation of AM processes. This optimization is essential for achieving higher efficiency, better product reliability, and consistent manufacturing outcomes. Using a NN trained on input and output data, it is possible to optimize input parameters.

Pawel Mieszczanek developed a closed-loop control system for 3D Melt Electrowriting (MEW) using a Feedforward Neural Network (FNN) [32]. A telecentric vision system combined with automated experiments rapidly collected input-output data characterizing the MEW parameter space. FNN with twelve hidden-layer neurons was trained to suggest optimal input parameters for achieving user-specified outputs. The resulting closed-loop control system significantly reduces output errors in real-time. Fig. 6 shows how the complete closed-loop control system operates. The system achieved the desired outputs—such as fiber diameter, jet angle, and Taylor cone area— within 5% accuracy, enabling automation and real-time quality control of the complex MEW process [32].

Mehran Ghasempour-Mouziraji developed an independent Artificial Neural Network (ANN) optimized for predicting residual stress and displacement in Directed Energy Deposition (DED) processes using SS316 [35]. Data was generated through finite-element simulations of 20-layer deposition, varying laser power, scanning speed, and laser beam radius. After systematic architecture optimization, independent ANNs were trained and integrated with Non-dominated Sorting Genetic Algorithms (NSGA) to identify optimal process parameters [35]. This approach reduces process development time by replacing iterative simulations, enabling real-time quality prediction, and improving DED process efficiency.

Beyond these approaches, various optimization studies utilizing diverse NN architectures and Data processing techniques have been conducted. Shubham Chaudhry developed and compared two data-driven reduced-order models for predicting thermal and mechanical behavior in SLM [36]. The POD-ANN model uses Proper Orthogonal Decomposition (POD) to compress data into a set of representative modes, then trains an ANN to learn input-output relationships. The CAE-MLP model employs Convolutional Autoencoders (CAE) to automatically compress data into a latent space, connected to inputs via a Multilayer Perceptron (MLP). Both models achieved rapid and accurate predictions of process-induced deformation, displacement, and stress fields. Table 2 shows the error levels of each method according to the experimental results [36].

Harshal Nejkar developed an Attention-Enhanced Neural Network (AENN) framework to predict tensile strength of Polylactic Acid (PLA) specimens in Fused Deposition Modeling (FDM) and optimize process parameters [37]. The model learns relationships between three input parameters—printing speed, layer thickness, and nozzle temperature—while using Monte Carlo dropout for uncertainty quantification. Firefly and JAYA metaheuristic algorithms applied to the AENN model converged to identical optimal input values for maximum tensile strength, demonstrating the framework’s effectiveness and applicability to other processes and materials [37]. Svyatoslav Korneev developed an AI system that addresses manufacturing uncertainties by predicting actual droplet shapes [38]. To avoid expensive multiphysics simulations, the approach combines k-Nearest Neighbors (kNN) for data augmentation and Monte Carlo methods to generate CNN training data. The trained CNN extracts the manufacturing uncertainty kernel and predicts final shapes using Graphics Processing Unit (GPU) acceleration [38]. This approach enables faster predictions than multiphysics models and allows process optimization based on realistic fabrication outcomes. Jieyang Peng developed a Vision Transformer (ViT) framework with supervised contrastive learning to optimize three input parameters—flow rate, feed rate, and hot end temperature—based on real-time images [39]. Features extracted from images are mapped into feature space via Multilayer Perceptron layers, and a contrastive loss function ensures similar features are classified together. Training is conducted on 946,000 preprocessed images. This label-consistent learning mechanism enables the model to detect subtle differences, and after fine-tuning, ViT model achieves superior three-parameter optimization performance. Table 3 shows that the ViT model achieves higher accuracy than the previous models. Through real-time evaluation, parameters with errors are detected early, thereby reducing the defect occurrence rate. Utilizing this approach, optimal parameter combinations can be identified to minimize defects [39].

Mojtaba Mozaffar developed a method to predict temperature distribution in real-time using an RNN [40]. A large dataset was generated using an in-house finite element code, GAMMA, with input parameters including laser power, scan speed, toolpath strategy, geometric size, and shapes. Based on this dataset, the RNN was trained to predict two-dimensional temperature profiles according to the given input parameters. In this study, an RNN structure, which is stacked with the GRU formulation, was used to comprehensively learn hidden correlations among the data. This model is trained on 250,000 Finite Element Method (FEM) data. While the inference time remains fast, the framework is primarily designed for offline thermal analysis rather than real-time closed-loop process control. Using the trained RNN, the temperature distribution at any desired location in the DED process can be predicted, demonstrating the potential of RNNs for modeling complex behaviors in AM processes [40].

There is research that optimizes laser power using diverse types of NNs in DED processes. Yi-Ping Chen developed Multi-step Model Predictive Control (MPC), which is combined with a digital twin system to a DED process, computing predictions and optimized control sequences in real time [41]. A Time-series Dense Encoder (TiDE) NN was integrated with MPC to enable multi-step ahead prediction. This framework was trained on 640,277 time-series segments. The TiDE-integrated MPC accelerated the prediction process and demonstrated accurate temperature tracking to anticipate potential porosity defects, thereby generating optimized laser power profiles [41]. Vispi Karkaria developed a digital framework for optimizing input parameters in DED processes by applying Bayesian Inference and Bayesian Optimization (BO) approaches [42]. Initially, a surrogate model was developed using machine learning that integrates LSTM with Bayesian Inference. Unlike conventional LSTM, the incorporation of Bayesian Inference enables the inference of Bayesian posterior distributions and the measurement of predictive uncertainty. This surrogate model was utilized to predict real-time temperatures at different spatial locations during DED manufacturing. The accuracy of the surrogate model was validated. Based on the predicted temperature, Bayesian Optimization-based Time Series Process Optimization (BOTSPO) was employed to provide an optimized laser power profile that ensures the manufactured component meets the required mechanical properties [42]. Riddhiman Raut used Graph Neural Networks (GNNs) to predict temperature distributions in LPBF processes [43]. GNNs represent mesh elements as graph nodes and use message-passing mechanisms to capture local and global relationships. The Single-Laser GNN (SL-GNN) with four hidden layers achieved high accuracy, with a Mean Absolute Percentage Error (MAPE) of 3.77% for temperature prediction and 7.6% maximum error for temperature peaks. The Multi-Laser GNN (ML-GNN) with additional hidden layers showed lower accuracy but improved performance through hyperparameter optimization. The approach demonstrates near-linear computational scalability to graphs containing up to millions of nodes and edges. However, inference time scale is several seconds for the largest graphs, and this limits the framework’s applicability to real-time prediction. Both models demonstrated applicability across different machines and materials for process optimization [43]. Yunze Wang developed a model using Invertible Neural Networks (INN) in the variable material Screw-Based Material Extrusion (S-MEX) process to predict flow rate according to input parameters and present process optimization [44]. The INN not only predicts outputs based on learned data but also possesses an inverse function that suggests appropriate process parameters when desired outputs are specified. Data was collected by adjusting parameters such as material composition, screw speed, and extrusion temperature. A strategy for maintaining a consistent flow rate was developed using the trained INN, and quality improvements were confirmed through the optimized manufacturing process. This study demonstrated performance enhancement and the potential of the S-MEX process based on INN that learned the relationship between input parameters and outputs [44]. Xiao Shang developed the Accurate Inverse process optimization framework in laser Directed Energy Deposition (AIDED), which integrates machine learning models and genetic algorithms to optimize DED manufacturing [45]. The framework extracts melt pool contours from images and processes them using Principal Component Analysis (PCA). MLP NNs were trained to predict melt pool characteristics, achieving high accuracy (R² = 0.995 for area, 0.966 for angle). A genetic algorithm then optimized two objectives—higher print speed and smaller track width—by generating diverse input parameter combinations and evaluating them using the trained MLPs. The framework successfully determined optimal laser power, scan speed, and powder feed rate values, with prediction errors of 1.75% for width and 12.04% for height in multi-layer configurations. Transfer learning enhanced applicability across different materials [45].

In addition to traditional manufacturing contexts, research has explored the applicability of AI technologies to various specialized and emerging applications of AM processes. Youssef Abdalla researched a deep learning model for assessing printability when producing drug formulations using SLS in the pharmaceutical field [46]. Based on the trained deep learning model, the printability of formulations can be determined, and by appropriately adjusting input variables such as molecular structure and particle size, process optimization becomes possible. An Ensemble NN composed of five layers was developed and compared with other NNs, demonstrating its superiority. The Ensemble NN trained using molecular structure (Morgan fingerprint) as input variables to the first NN achieved the highest accuracy at 90%. Fig. 7 presents the results of the comparative analysis of the performance between the Ensemble NN and others. This study is the first to evaluate the application of NNs in the pharmaceutical field, demonstrating the feasibility of personalized medicine production through AM methods [46].

Amulya Shetty developed a trained Backpropagation Artificial Neural Network (BPANN) model capable of predicting displacement and strain of artificial bone in response to changes in the mechanical environment [47]. BPANN is an NN using a backpropagation algorithm that calculates the error between predicted results and actual values from initial input values, adjusts weights, and progressively reduces errors. The BPANN was trained based on a total of nine experimental datasets, three geometric structures (Lidinoid, Diamond, and Gyroid) with thicknesses of 1, 1.5, and 2 mm and loads of 3, 6, and 9 kN. The generated model predicted deformation with high accuracy in unknown designs, and using this model demonstrated the potential for process optimization [47]. In AM, using mixed materials, predicting material behavior, and controlling quality through appropriate parameter optimization are important factors.

Matsive Ali developed a real-time adaptive control framework applicable to smart AM processes [48]. A NN was trained using the Soft Actor-Critic (SAC) method, one of the Reinforcement Learning (RL) algorithms, and this trained NN was utilized to control the robotic arm used in AM operations. Since direct application of RL algorithms to manufacturing processes can cause equipment damage and production defects, a digital twin approach was applied to create a virtual robotic arm synchronized in real-time. Fig. 8 shows the results of constructing a virtual robot arm system. An RL algorithm was trained in a digital twin environment using Hierarchical Reward Structure (HRS) and transfer learning for robotic arm AM optimization. The trained RL network autonomously controls the robotic arm to produce user-defined outputs, enabling optimal control in complex AM processes without human intervention [48].

Traditional NNs have been trained on input and output data from a single printer to predict manufacturing processes. In the case of 3D printer farms, manufacturing facilities that use multiple 3D printers for mass production, different errors occur across individual printers, necessitating NNs trained on integrated datasets. Benjamin Standfield investigated geometric deformation prediction in 3D printer farms using 3D CNNs with federated learning [49]. Three types of tooth geometries were printed on different printers, scanned, and processed through noise removal, alignment, and voxel-based segmentation. The study compared centralized and decentralized (federated) approaches for predicting and compensating deformation. While centralized and federated approaches showed negligible practical differences in performance, federated averaging achieved a 94.75% reduction in network bandwidth compared to transmitting entire datasets, demonstrating efficient deformation compensation across 3D printer farms. Fig. 9 shows the overall methodology of prediction and compensation [49].

Predicting the shape and quality of manufactured parts in AM processes presents significant challenges. Traditional numerical simulation methods require extensive computational time, while prediction through deep NNs demand large datasets. These approaches struggle to accommodate continuously evolving material characteristics and manufacturing techniques, limiting the development of appropriate prediction systems. To address these limitations, PINN approaches have been investigated, which integrate physical laws into NN learning, enabling accurate prediction of manufacturing outcomes with minimal datasets.

Research has demonstrated that appropriate predictions are feasible using NNs trained solely on physical laws without requiring labeled training data. Rahul Sharma developed a PINN model capable of predicting three-dimensional temperature profiles at specific locations in Laser Metal Deposition (LMD) processes by applying NNs and physical laws without utilizing labeled training data [33]. Since no external data is provided, the PINN in this study integrates the loss of governing PDEs, initial conditions, and boundary conditions into the loss function. The developed PINN architecture comprises four hidden layers, each consisting of thirty-two neurons. The output of each layer is derived by applying the swish activation function, as illustrated in Fig. 10 below, which depicts the structure of the PINN model used. To evaluate the prediction accuracy of the PINN model, the results were compared with those derived from Finite Element Analysis (FEA). Fig. 11 presents a comparison between the results obtained using the PINN model and the FEA results. The PINN model predictions demonstrated high accuracy when compared to FEA results, with a relative error of 2.1% at the top boundary and 3.7% within the domain near the bottom. This framework reduces the need for iterative experiments and demonstrates the advantages of combining physics-based analysis with NN learning, enabling real-time quality control and automatic optimization in AM processes.

Beyond this, numerous studies have been conducted to improve accuracy with limited data by applying PINN models. Min Xia developed research incorporating thermal transfer and adiabatic relationships as physical laws into a PINN model to predict temperature distribution and weld pool dimension in metal AM processes [50]. The prediction accuracy was validated, and the possibility of an advanced model reflecting fluid flow physics in the melt area was presented [50]. Pouyan Sajadi conducted research applying the concept of online learning to overcome the limitations of conventional PINNs that are trained only in the offline stage [51]. Through this approach, real-time temperature prediction during metal AM processes was achieved, and an adaptive technology was developed that can dynamically enhance prediction accuracy by continuously learning new data during the manufacturing process [51].

To design an enhanced PINN model, additional training techniques can be applied during the NN learning process. Qingyun Zhu developed a PINN model, which is enhanced through a transfer learning method, for predicting melt pool profiles in the SLM process [34]. The SLM process involves numerous complex physical phenomena. Currently, low-fidelity data struggles to accurately capture these complex phenomena, while high-fidelity data containing intricate physical laws makes it challenging to apply the developed PINN models. To address these limitations, Transfer Learning Enhanced PINN (TLE-PINN) model was investigated. The TLE-PINN model incorporates heat transfer laws and boundary conditions into the loss function and learns subtle correlations between data through transfer learning. Fig. 12 illustrates the architecture and operational workflow of the TLE-PINN framework in this study. In this study, training require high-fidelity Finite Volume Method (FVM) data trained with a transfer learning framework. Although inference is computationally efficient, the model’s heavy reliance on precomputed simulation data substantially constrains its integration into fully online process control systems. The TLE-PINN model provides excellent convergence speed and imposes strong physical constraints on the data. Superior performance was demonstrated compared to conventional PINN approaches, and the potential for industrial applications across various fields was validated [34]

Pouyan Sajadi researched the prediction of temperature fields in metal AM processes by combining PINN and Convolutional Long Short-Term Memory (ConvLSTM) architectures [52]. The Physics-Informed Convolutional Long Short-Term Memory (PI-ConvLSTM) framework, which integrates PINN and ConvLSTM, predicts 2D temperature fields based on previous thermal images obtained in real-time during the manufacturing process and input parameters. This research addresses a high-dimensional task that generates 2D outputs from 2D inputs. CNNs offer advantages in managing high-dimensional data through efficient parameter sharing, while LSTM architectures excel at capturing temporal trends and dependencies in data sequences. To leverage both advantages, a ConvLSTM architecture that combines CNN and LSTM was employed. The study demonstrated prediction errors within 3% for temperature fields in thin wall structures and prediction errors below 1% for cylinder and cubic structures, exhibiting high accuracy. The manufacturing process can be optimized based on temperature fields predicted by the PI-ConvLSTM framework.

This framework utilizes only 2,760 experimental images data and is proper for real-time digital twin implementations. Also, it successfully enabled temperature field predictions across various structures, deposition modes, and input parameter configurations [52].

Furthermore, numerous studies have been conducted to enhance PINN performance by incorporating novel techniques into NN architecture. H. Safari addressed the complex heat transfer problems in LPBF processes through an operator learning approach combining Deep Operator Network (DeepONet) and PINN [53]. Notably, the method enables accurate prediction of three-dimensional temperature distribution during melting and solidification in both single-track and multi-track scenarios and allows for rapid analysis of the effects of scanning paths and laser parameters [53].

Table 4 summarizes the key characteristics of major NNs that have been investigated for process optimization.