Metal Additive Manufacturing and Molten Pool Dynamic Characterization Monitoring: Advances in Machine Learning for Directed Energy Deposition

Abstract

Directed energy deposition (DED) has progressively emerged as a highly promising technology for the rapid, cost-effective, and high-performance fabrication of hard-to-process metal components with shorter production cycles. Recognized as one of the most widely utilized metal additive manufacturing (AM) techniques, DED has found extensive applications in critical industrial sectors such as aerospace and aviation. Despite its potential, challenges such as inconsistent part quality and low process repeatability continue to restrict its broader adoption. The core issue underlying these challenges is the complex, dynamic nature of the DED process, which involves the coupling of multiple physical fields. Within this context, the molten pool plays a pivotal role, serving as a key carrier that encapsulates abundant process characteristic information. The dynamic characteristics of the molten pool are intrinsically linked to the final part quality and the repeatability of the process. Consequently, integrating machine learning (ML) methodologies into the monitoring framework can offer robust data-driven support for enhancing both product quality and process consistency. This paper provides a comprehensive review of the research advancements and prospective trends in the dynamic monitoring and control of molten pool characteristics within DED processes underpinned by machine learning techniques. The review is structured around five key areas: an overview and fundamental principles of DED technology, methods for process information sensing during part monitoring, approaches for dynamically monitoring molten pool characteristics, the primary challenges currently faced in intelligent monitoring systems, and the potential future directions for further research and development. Through this detailed examination, the paper aims to shed light on the pivotal role of intelligent monitoring systems in advancing DED technology, ultimately contributing to more reliable and repeatable additive manufacturing processes.

1. Introduction

Additive manufacturing (AM) [1,2,3], as a cutting-edge technology in high-end manufacturing, involves the discretization of a digital model into layers, which are then sequentially accumulated to build the final part. This process offers numerous advantages, including shortened production cycles, high material utilization, digitization, automation, and the ability to cater to highly personalized and customized products. These advantages render AM a “transformative” technology that seamlessly integrates design, materials, and manufacturing processes [4,5]. AM has provided a significant opportunity for the transformation and upgrading of traditional manufacturing industries [6,7] and has gradually become a national strategy for leading manufacturing nations around the world. In the United States, the Department of Defense, along with five other government agencies, 85 enterprises, and 13 research universities, jointly established the National Additive Manufacturing Innovation Institute (NAMII). The core objective of this initiative is to secure leadership in the upcoming wave of technological advancements within the manufacturing sector and, based on this technology, launch the AM Forward program aimed at achieving the reshoring of high-end manufacturing. Similarly, countries such as the United Kingdom, Germany, France, and Japan have formulated and implemented national strategic policies and established research institutions dedicated to the advancement of AM. In China, AM has been prioritized as a key development area in both the “Made in China 2025” initiative and the “14th Five-Year Plan” for intelligent manufacturing.

High-end equipment is evolving towards larger sizes, greater complexity, higher reliability, and rapid design iteration. Metal components used in equipment manufacturing must be meticulously designed, precisely fabricated, and possess excellent performance, achieving the integration of functionality, performance, unity, and lightweight properties, essentially combining these features into one cohesive solution. Traditional manufacturing methods sometimes fall short of meeting these demanding requirements [8,9,10]. In this context, metal AM has emerged as a promising solution, capable of achieving lightweight, compact, and functional designs for complex metal components, as well as integrated material-structure net shaping. This technology is increasingly recognized as a potential game-changer for the rapid, low-cost, high-performance, and short-cycle manufacturing of difficult-to-process metal components, and it has already been widely applied in industries such as aerospace, aviation, and other industrial sectors [11,12,13,14,15]. Among the various AM technologies, metal AM stands out due to its significant volume in the materials field, emphasizing its importance and the many challenges associated with it. It has become a key research direction in advanced manufacturing technologies. Based on different process principles, metal AM technologies can be broadly classified into two categories: directed energy deposition (DED) and powder bed fusion (PBF). DED involves the use of focused thermal energy, such as lasers, plasma arcs, or electron beams, to melt and deposit metal powders or wires in an AM process [16]. When using a laser as the heat source, DED is known as laser-directed energy deposition (L-DED), which is one of the mainstream AM technologies.

Due to the complex non-equilibrium physical metallurgy and thermophysical processes resulting from the interaction between the laser, powder, and substrate during the AM process, materials often experience intricate behaviors that lead to internal quality issues such as cracks and porosity, as well as performance-related quality issues like deformation and surface roughness [17,18,19,20,21,22], as illustrated in Figure 1. Consequently, the stability and consistency of the DED and the resulting part quality have become significant challenges for the industry, presenting a major obstacle to large-scale production. Controlling the process parameters in DED is crucial for enhancing the stability and consistency of part quality. At present, the control of DED quality mainly relies on offline data and traditional empirical optimization methods, which result in high scrap rates and insufficient reliability of part quality. The conflict between “embracing” the transformative potential of additive manufacturing and “concern over” manufacturing quality issues has become a core bottleneck affecting the practical application of this technology. As a result, numerous research institutions worldwide have identified the monitoring of AM quality consistency and stability as a key technological development focus for the future. For instance, in October 2022, the U.S. Office of Science and Technology Policy released the National Strategy for Advanced Manufacturing, emphasizing the importance of optimizing the AM process and developing sensor technologies and other control measures to enhance process monitoring capabilities. In 2019, Germany published a guideline on industrial AM quality assurance processes, which, for the first time, clearly defined the consistency requirements for AM quality. Similarly, China has also formulated relevant policies, such as the “Additive Manufacturing Industry Development Action Plan”, which explicitly states the need to establish a testing and certification system for AM, carry out certification and evaluation analysis, and conduct core research on quality assurance technologies.

The direct fabrication of complex metal parts is the ultimate goal of metal AM technology proposed by the manufacturing industry [24]. Metal AM utilizes its inherent rapid heating and cooling characteristics to produce parts with higher strength and quality, but the parts are prone to defects such as porosity and cracks, as well as warpage and deformation during the forming process [25,26]. Therefore, online monitoring of metal AM characteristics is essential to investigate the relationship between part defects and process characteristic information. As an important carrier in the metal AM, the molten pool reflects the whole process of melting–fusion–solidification of the metal AM process, and the study of the molten pool feature information is of great significance for the online monitoring of metal AM [27,28,29,30].

Although DED technology has garnered widespread attention due to its advantages in manufacturing complex metal components, it still faces significant challenges in ensuring process quality consistency and repeatability. Existing studies indicate that the dynamic characteristics of the molten pool are critical factors affecting the forming quality. However, traditional monitoring methods struggle to capture and analyze molten pool characteristics in real time, limiting the industrial application potential of DED technology. To address these issues, machine learning (ML), as a data-driven analytical tool, offers new perspectives and solutions. With the rapid development of technologies such as big data, advanced sensing, and artificial intelligence (AI), intelligent monitoring technologies based on acoustic, optical, thermal, and magnetic information, combined with AI algorithms, have gradually been applied to AM [31,32,33,34]. Numerous scholars have provided detailed reviews on various aspects of metal AM, including defect types, process information sensing, quality assessment, and quality control [31,35,36,37]. However, current DED monitoring systems primarily focus on the “measurement” aspect, i.e., the collection of various process data. The effectiveness of quality assessment and control technologies remains inadequate, and their maturity is still insufficient. This limitation is largely due to the extreme complexity of the DED process, which leaves uncertainties regarding “when” and “how” to apply control measures. In 2023, ASTM International, in collaboration with NASA, the Marshall Space Flight Center, and various industry organizations, held a seminar on additive manufacturing monitoring technologies and published the “Strategic Guide for Additive Manufacturing In Situ Monitoring Technologies.” The guide highlighted the importance of in situ monitoring as an integral component of quality assessment and control. It emphasized that while many current quality assessment and control technologies are not yet mature, there is a significant gap in the study of the causal relationships between monitoring data and defect mechanisms. Therefore, by integrating advanced sensing technologies with AI methodologies, it is possible to enhance quality assessment and control techniques. The development of mature, intelligent monitoring systems for large-scale DED could serve as a powerful tool for optimizing production processes. Such systems are expected to play a crucial role in improving the stability and consistency of both the DED process and the quality of the produced components. This advancement will have significant implications for scaling up DED, providing a solid foundation for achieving higher efficiency, reliability, and quality in large-scale production.

This paper reviews the progress of research in intelligent monitoring and control of dynamic molten pool characteristics in DED using ML, as illustrated in Figure 2. It highlights the existing challenges in intelligent monitoring for large-scale production in DED AM and identifies potential directions and future trends. By addressing these aspects, the paper aims to explore feasible approaches to overcoming current challenges, thereby advancing the development and industrial application of intelligent DED monitoring systems. Unlike existing reviews, this paper is the first to systematically summarize the key techniques of machine learning in DED monitoring, especially in the dynamic characterization of the molten pool. Through multi-source information fusion, machine learning-based feature extraction and prediction, and intelligent control strategies with multivariate interactions, this paper provides new perspectives and methodological support for improving the consistency and stability of DED process quality.

This review paper aims to systematically summarize the research progress of ML in DED monitoring and explore its future development directions in conjunction with practical application cases. The main structure of this paper is organized as follows: Section 2 (DED Technology Profile) introduces the fundamental principles of DED technology and its key technical background, providing theoretical support for subsequent discussions; Section 3 (Process Information Sensing) focuses on the process information sensing methods in the DED process, including the acquisition and analysis of multi-source physical field signals; Section 4 (Molten Pool Dynamic Monitoring) discusses in detail the current status and challenges of dynamic monitoring of the molten pool, with particular emphasis on the application of machine learning in predicting and controlling molten pool characteristics; Section 5 (Challenges of Intelligent Monitoring) analyzes the main challenges faced by intelligent monitoring technologies in industrial applications; Section 6 (Future Outlook) looks ahead to future research directions, including the development of adaptive control systems and the establishment of standardized metrics; Section 7 (Conclusions) concludes the paper and presents key findings. Through this structural arrangement, the paper aims to systematically explore the research progress and future development directions of ML-driven intelligent monitoring technologies for DED processes.

2. DED Technology Profile

The concept of DED was first introduced in a patent, which described a method of using a laser to melt and deposit added metal powder for the repair of damaged components. Since then, DED technology has developed rapidly over the following years. Various research institutions, based on their understanding of DED and their own research focus, have given the technology different names.

Table 1. DED technical naming.

Naming Authority	Name	Abbreviated Name
Sandia National Laboratories, Albuquerque, NM, USA	Laser-engineered net shaping	LENS
Los Alamos National Laboratory, Los Alamos, NM, USA	Directed light fabrication	DLF
University of Michigan, Ann Arbor, MI, USA	Directed metal deposition	DMD
Stanford University, Stanford, CA, USA	Shape deposition manufacturing	SDM
University of Birmingham, Birmingham, UK	Direct laser fabrication	DLF
University of Liverpool, Liverpool, UK	Laser direct casting	LDC
Fraunhofer Institute, Aachen, Germany	Controlled metal build-up	CMB
Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland	Laser metal forming	LMF
Northwestern Polytechnical University, Xi’an, Shaanxi, China	Laser solid forming	LSF

lists the different names attributed to DED technology by various research institutions.

DED technology uses a laser beam as a high-temperature heat source to melt the surface of the substrate to produce a molten pool. Through powder/wire feeding equipment, it will be synchronized with metal powder/wire feeding into the molten pool. Then, the powder/wire will experience rapid melting and cooling, solidification, and the formation of a metallurgical bond with the substrate material. The laser deposition work head remains under the control of a computer in accordance with a pre-programmed path of movement, accumulating the parts layer by layer to achieve physical manufacturing [38,39]. Depending on the deposited material, the whole process is usually carried out in an inert gas atmosphere, such as argon and nitrogen.

A DED system is mainly composed of a laser deposition system, a motion control system, and a protection monitoring system. The laser deposition system is the core of the whole AM system, which is mainly composed of a laser, powder/wire feeding equipment, a cooler, a deposition head, and a molding platform and is used for material delivery and melting deposition to the molding platform. The motion control system is mainly composed of a CNC machine tool/mechanical arm, program controller, and CAM programming software and is used to realize the spatial positioning and movement of the deposition head/molding platform and to manufacture different shapes of components. The protection monitoring system is mainly composed of a safety enclosure, gas chamber, online monitoring equipment, and supporting software and is used to protect the processing safety and monitor the whole system. The protection and monitoring system is mainly composed of a safety shell, gas chamber, online monitoring equipment, and supporting software, which is used to protect processing safety and monitor the whole forming process to ensure processing accuracy. Figure 3 shows the schematic diagram of a typical L-DED system.

DED technology has the following characteristics [41,42]:

• It does not require molds, and the forming size is not limited, allowing for the direct forming of large components;
• It offers high flexibility in monitoring methods and can process complex parts without the need for support structures;
• It can be used for the direct repair of damaged parts and the manufacturing of multi-gradient parts;
• The processing cycle is short, material utilization is high, and post-processing is minimal;
• The as-formed parts exhibit excellent room-temperature mechanical properties, and the mechanical properties of heat-treated parts can reach the level of forged components.

3. Process Information Sensing

The use of advanced sensing techniques to detect cross-scale defects in the metal AM process is a fundamental prerequisite for realizing intelligent monitoring of the AM process. Since the metal AM process involves the coupling of multiple physical fields, including acoustic, optical, thermal, and magnetic fields, researchers, both domestically and internationally, have begun to collect diverse sensing information from different physical fields. Various methods have been employed to extract defect-related information and characterize these defects, leading to extensive research on this topic.

3.1. Optical Signals

Optical signal monitoring is one of the most straightforward and widely used methods in the process monitoring of metal AM. Generally, optical signal monitoring can be applied to molten pool and internal structure monitoring. Commonly used optical sensors include industrial cameras, high-speed cameras, photodiodes, and X-ray imaging systems.

Powder spreading images, part geometries, and surface roughness are crucial characteristics of the process of metal AM. Industrial cameras are widely used for monitoring these features due to their high resolution and low cost. For instance, Scime et al. [43] at the Oak Ridge National Laboratory utilized an off-axis industrial camera to capture powder bed images during the operation of the EOS M290 system. They monitored and analyzed anomalies occurring during the powder spreading stage and demonstrated the practicality of their proposed algorithm as an independent software package. Common sensors used in DED process monitoring include wired array cameras, high-speed cameras, and infrared cameras.

Table 2. Comparison of key performance indicators of commonly used sensors.

Sensor Type	Resolution	Frame Rate	Environmental Adaptability	Applicable Scenarios
Wired array camera	High	Moderate	High	Powder spreading, geometric monitoring
High-speed camera	Moderate	High	Moderate	Molten pool dynamic monitoring
Infrared camera	Moderate	Moderate	High	Temperature distribution monitoring

compares the key performance metrics of these sensors. The primary advantage of wired array cameras lies in their high resolution and real-time imaging capabilities, which are crucial for capturing powder spreading and molten pool dynamic characteristics. Furthermore, these cameras can be integrated with dedicated lighting systems to ensure consistent illumination conditions, thereby enhancing image quality and reducing environmental interference. In contrast, conventional industrial cameras may exhibit limitations in terms of dynamic range or sampling precision. Fischer et al. [44] from the Fraunhofer Institute for Laser Technology (ILT) installed an industrial line scan camera on the powder scraper and designed a corresponding lighting system. This setup allowed for line scan imaging during the powder spreading process. Line scan imaging provides high resolution, consistent lighting conditions and eliminates the need for additional sampling time. However, it is limited by the accuracy of motion and sampling precision, which may lead to image distortion. The team led by Shi at Huazhong University of Science and Technology [45] projected a series of sine wave stripe images onto the forming platform and used a phase-guided contour extraction method to accurately locate the contour center. This significantly reduced the impact of harsh manufacturing conditions and could potentially serve as an effective method for monitoring geometric defects in layer formation.

When a high-energy laser melts metal powder, a molten pool is formed. Since the formation of the molten pool occurs within an extremely short time scale, high-speed cameras are typically used to capture images of the molten pool morphology and its dynamic behavior during the process. Zhu’s team from the University of Chinese Academy of Sciences [46] used a coaxial high-speed camera to capture molten pool images and observed the oscillations of the molten pool in the plane perpendicular to the laser’s incident direction. These oscillations were used to characterize the molten pool’s size, shape variations, and movement characteristics (as shown in Figure 4). Rai’s team at Carnegie Mellon University [47] used industrial cameras to capture molten pool information during the additive manufacturing process. By combining ML techniques, they predicted the size and shape of the molten pool (the process is shown in Figure 5). Although their approach can support the monitoring of interlayer fusion behavior, it lacks information about the molten pool depth, and further research is needed to capture the formation and disappearance of the spatter. While high-speed cameras are capable of capturing molten pool details and dynamic changes at high frame rates, they generate large volumes of data during imaging, which increases the complexity of data transmission, storage, and processing.

Photodiodes can capture the light emitted by the molten pool, providing information about the light intensity and spatial distribution, which, in turn, allows for the extraction of molten pool size and shape information. Liu et al. [48] from Pennsylvania State University simulated the photon emission physics involved in molten pool image generation and used photodiodes to detect the simulated photons (as shown in Figure 6). They statistically analyzed the two-dimensional multimodal probability distribution of the molten pool and developed a multimodal distribution model to effectively describe the geometric variations of the molten pool. Photodiodes are highly sensitive and cost-effective; however, their spatial resolution is limited, making it difficult to simultaneously monitor both global and detailed information.

X-ray imaging, due to its non-contact nature, strong penetration, and high resolution, has become a popular technology in recent years for monitoring internal forming quality and micro-defects. In a study published by Ren et al. [49] at the University of Virginia in Science, high-speed synchrotron X-ray imaging and thermal imaging were simultaneously employed (the structure of the monitoring system is shown in Figure 7a). By utilizing multi-physical field simulations and ML methods, they identified two types of keyhole oscillations occurring during the fusion process and predicted the formation of porosity during the manufacturing process. In another study by Poudel et al. [50] published in Nature Communications, high-resolution X-ray computed tomography (CT) was used to analyze the morphological features and statistical distribution of volumetric defects in additively manufactured parts (as shown in Figure 7b). Additionally, they publicly released a dataset of the tomographic scans.

3.2. Acoustic Signals

There are numerous inherent relationships between the molten pool state, the internal quality of the formed part, and the acoustic signals during the metal AM process. For instance, the operation of the laser and the molten metal can generate acoustic signals or ultrasonic waves can be transmitted into the part via a transmitter to detect internal quality. Commonly used acoustic signal monitoring technologies include acoustic emission (AE) and laser ultrasound techniques, among others.

AE technology can capture the acoustic signals generated during the laser melting process. Drissi-Daoudi et al. [51] used ultrasonic microphones to capture the acoustic emission signals generated during the forming process, which were then classified for issues such as incomplete fusion, porosity, and conduction modes (the equipment is shown in Figure 8). However, this method is susceptible to environmental noise, and one of the key challenges in acoustic signal monitoring is how to isolate the acoustic signals generated from the interaction between the laser and the material from other sources.

Laser ultrasonic devices can be directed onto the surface of the sample, generating ultrasonic pulses that propagate into the sample and receiving the reflected signals to detect the internal quality of the part. Raffestin et al. [52] used a laser ultrasonic system to measure the propagation velocity of pulses, monitoring the evolution of the melt layer thickness and part height during the AM process (as shown in Figure 9a). They tracked defects and established a numerical model for analysis. A team led by Zhang at Wuhan University [53,54,55,56] conducted research on noise reduction, ultra-high-speed monitoring, and multi-feature fusion in laser ultrasonic monitoring of additive manufacturing, with the ultrasonic imaging of defects shown in Figure 9b. While laser ultrasound can detect small defects and subtle structural changes, it is affected by the material’s propagation characteristics, requiring parameter calibration and optimization for different materials.

3.3. Thermal Signals

During the metal powder melting process, a large amount of heat is generated, and the heat produced in the molten pool, along with its distribution in the formed part and powder, can significantly affect the quality of the part. Therefore, monitoring the thermal signals during the metal AM process is of great importance. Commonly used thermal sensors include infrared cameras and thermocouples, among others.

Infrared cameras can monitor and record the temperature distribution of parts and powder layers during the forming process in real time and track the temperature history. Williams et al. [57] integrated a wide-field infrared camera (structure shown in Figure 10) into a powder bed fusion system to measure the surface temperature of the powder layer, correlating it with the calibrated inter-layer cooling time. This is closely related to the density and lattice structure of the formed part. Liu et al. [58] used a long-wavelength infrared camera to track the temperature history after each powder layer was deposited, exploring the correlation between temperature history, powder layer thickness, and thermal conductivity, which was then used to predict the deformation of unsupported structure parts. However, the measurement accuracy of infrared cameras is also influenced by factors such as measurement distance and environmental temperature, and calibration must be conducted based on the actual conditions during use.

Thermocouples, based on the principle of the thermoelectric effect, measure temperature by utilizing the characteristic voltage difference that occurs due to the temperature variation between two different metal wires. They are commonly used to measure the temperature distribution in the forming cavity and substrate. Dunbar et al. [59] from Pennsylvania State University employed K-type thermocouples to measure the substrate temperature and analyzed the impact of laser scanning patterns on distortion accumulation during the multi-layer melting process based on the measured data. Nuñez et al. [60] from the Idaho National Laboratory designed various schemes to embed K-type thermocouples in the DED process, further improving the structural stability and tolerance of the components, reducing porosity, and effectively optimizing the system monitoring technology and its real-time capabilities. However, since thermocouples require contact with the measured object, the locations where they can be installed in metal AM equipment are limited, making it challenging to achieve complex measurements.

3.4. Other Signals

In addition to commonly used monitoring methods based on acoustic, optical, and thermal signals for AM processes, other classical detection techniques based on magnetic and vibration signals are also frequently employed to monitor the information during the forming process.

Eddy current testing (ECT) is a widely used magnetic signal-based method for defect detection. In metal AM, ECT is extensively applied to detect internal defects such as porosity, gas pores, incomplete melting, and inclusions. Guo et al. from the Zhang team at the Southern University of Science and Technology [61] employed ECT to detect sub-surface defects with diameters ranging from 0.4 to 1.6 mm in PBF manufacturing. They conducted a comprehensive study on the influence of factors such as excitation frequency, relative distance (distance between probe and sample), defect depth, and size on the defect detection performance of eddy current technology. The process and equipment structure are shown in Figure 11a and Figure 11b, respectively. Ehlers et al. [62] combined high spatial resolution giant magnetoresistance (GMR) arrays with a single-line excitation coil for defect detection, experimentally verifying the ability to detect defects smaller than 100 $\mathsf{\mu }\mathrm{m}$ in PBF-manufactured samples. However, ECT technology is highly dependent on the detection environment and installation position, and it is challenging to detect deep or small-sized defects.

Accelerometers are among the most commonly used sensors for measuring vibrations in industrial applications. In metal AM, accelerometers are often employed to monitor abnormal vibrations caused during the powder spreading process or the powder melting phase. Zhirnov et al. [63] developed a monitoring system consisting of a microphone and accelerometer (as shown in Figure 12), which can be mounted on the powder bed substrate to monitor defects such as spatter, spheroidization, and porosity that occur during the melting process. However, due to the complexity of the metal AM process, vibration detection is prone to interference from external factors.

In summary, it can be observed that scholars worldwide have conducted extensive research on multi-physics field information sensing in metal AM processes, achieving substantial results. These efforts have laid a solid foundation for the large-scale production of AM.

4. Molten Pool Dynamic Monitoring

In the metal AM process, the molten pool contains a wealth of temperature-related information. Monitoring the molten pool temperature provides valuable insights that enable real-time, online evaluation of the printing process. Moreover, by analyzing the distribution and variation patterns of the molten pool temperature, process parameters (such as laser power, powder feed rate, scanning speed, layer height, and so on) can be dynamically adjusted to ensure the forming quality of the workpiece and the reproducibility of the printing process.

Small variations in process conditions can lead to significant differences in the morphology of the molten pool, which, in turn, affects the forming quality of the workpiece. The stability of the molten pool morphology reflects the stability of the metal additive manufacturing process. Therefore, monitoring the morphology of the molten pool is particularly important. The molten pool morphology includes characteristics such as its length, width, area, height, depth, and intensity distribution. Existing monitoring systems typically utilize CCD cameras, near-infrared (NIR) cameras, CMOS cameras, pyrometers, photodiodes, or combinations of these devices. Additionally, the image data collected during the monitoring of the molten pool morphology have contributed to the development of image processing technologies.

4.1. ML Applications

ML is a form of AI technology that aids in identifying patterns in data and generating similar data without human intervention, allowing for decision-making or predictions. This paper systematically analyzes the potential application of ML techniques in the dynamic monitoring of DED molten pools, especially its advantages in feature extraction, predictive modeling, and intelligent control. For example, a deep learning method combining convolutional neural network (CNN) and autoencoder can realize efficient feature extraction and anomaly detection of melting pool images, and a regression model based on a support vector machine (SVM) can accurately predict the depth and width of the melting pool. In addition, this paper further proposes a multivariate interactive control framework to optimize the key process parameters using reinforcement learning techniques, which realize the closed-loop control between the dynamic characteristics of the molten pool and the forming quality. In general, research on identifying and analyzing molten pool image features using ML can be divided into three main categories: supervised learning, unsupervised learning, and reinforcement learning. Supervised learning enables us to learn from a set of labeled data in a training set, allowing the identification of unlabeled data from a test set, representing the inherent features of relevant samples, and revealing physical phenomena. Among these, classification and regression are typically two key functions for predicting the quality of finished parts in AM. Zhang et al. [64] utilized SVM and CNN to study the impact of these target features on build quality in the PBF process. The experiment showed that for constructed images (as shown in Figure 13a), the classification accuracy of CNN was higher than that of SVM. Based on this, a hybrid CNN method for PBF process monitoring was further proposed (as shown in Figure 13b). This method saves image processing steps and can automatically learn spatiotemporal representative features (e.g., molten pool) from raw images [65]. However, supervised learning also has certain drawbacks, namely, that only known categories can be effectively predicted within the model.

Unsupervised learning (such as clustering and principal component analysis, PCA) infers from unlabeled data, identifying hidden patterns or grouping similar data points in a given random dataset. It is widely applied in anomaly detection. Grasso et al. [66,67] studied defects in selective laser melting (SLM) and proposed an automatic defect detection method using a machine vision system in the visible light spectrum. This method combined PCA for statistical descriptor selection and k-means clustering for defect detection (as shown in Figure 14a). Yang et al. [68] proposed a new method called layer-wise neighboring-effect modeling (LNBEM), which uses scanning strategies to predict molten pool size. Specifically, they trained a fully connected neural network on the data with a new feature set to predict the molten pool area. In a later study, Fathizadan et al. [69] introduced a framework using a convolutional autoencoder neural network to process molten pool images. The results showed that the accuracy of feature extraction reached 95%, and they then proposed a control chart scheme based on Hotelling’s statistics to monitor process stability. Following this research trend, Kim et al. [70] developed a new deep learning-based method that directly estimates the molten pool coordinate position from AM molten pool images (as shown in Figure 14b). They used stacked convolutional autoencoders for image preprocessing (i.e., image transformation and spatter reduction) to denoise the raw molten pool monitoring images. In contrast to previous studies, Khanzadeh et al. [71,72] proposed a new method for porosity prediction based on molten pool top surface temperature distribution. To determine molten pool image features, they transformed the related data into a spherical coordinate system and used biharmonic surface interpolation to smooth and extract feature vectors. These vectors were then input into self-organizing maps (SOMs), and each image was processed to classify it as normal or porous. Additionally, to better study the process–structure–property (PSP) relationships in various AM processes, Ko et al. [73] proposed a new framework driven by machine learning and guided by physics (as shown in Figure 15). This framework involves three levels:

• Prediction of the PSP model and physical knowledge;
• Features of interest in PSP;
• Raw data from additive manufacturing.

This framework provides a systematic, physics-guided, data-driven approach to PSP in AM, combining physical knowledge with the generalizability of data-driven ML models.

In recent years, ML-driven monitoring techniques have shown great potential in industrial applications. For example, in the aerospace sector, ML has been used to monitor the dynamic characteristics of the molten pool in real time, resulting in a significant reduction in part defect rates. The practical application of these techniques provides valuable lessons for DED process optimization in industrial production. To validate the practical effectiveness of ML-driven monitoring techniques in the DED process, we take an aerospace manufacturer as an example, which utilizes a CNN-based monitoring system to analyze the molten pool dynamics of critical parts in real time. The system adjusts the laser power and scanning speed in real time by capturing the optical and thermal characteristics of the molten pool, which significantly improves molding quality and consistency. Experimental data showed that the system reduced part defect rates by 35% and increased productivity by 20% compared to traditional empirical optimization methods. In addition, the technology has been successfully applied in other domains (e.g., mold manufacturing and energy equipment repair), providing a reliable reference model for large-scale rollout. This case study demonstrates that ML-driven monitoring technology has significant potential for application in industrial DED processes. In the future, through further development and promotion of similar technologies, more efficient and reliable solutions can be provided to the manufacturing industry, thus promoting the large-scale industrialized application of AM technologies.

In the field of AM, deep learning (DL), a branch of ML, is gradually becoming an important tool for digital monitoring and intelligent control. By deeply mining and analyzing process data, artificial neural networks (ANNs) can efficiently identify complex patterns and trends, providing support for process optimization in AM. In the DED process, ANNs have shown significant potential in molten pool morphology monitoring, defect detection, and process parameter optimization. CNNs are typically used in DED monitoring for molten pool morphology and defect detection. Their convolutional layers efficiently extract spatial features from images, making them suitable for analyzing the width, depth, and morphological changes of the molten pool. Region-based CNNs (RCNNs) further enhance the detection capability of target regions but come with higher computational complexity, making them suitable for high-precision detection of small defects. In cases of insufficient or unevenly distributed data, generative adversarial networks (GANs) can significantly improve model training by generating high-quality synthetic data. Additionally, long short-term memory (LSTM) networks offer unique advantages in processing time-series data, such as molten pool temperature dynamics, supporting real-time feedback and control.

Table 3. Characteristics, advantages, disadvantages, and applicable scenarios of common ANN models in DED monitoring.

Model Type	Characteristics	Advantages	Disadvantages	Applicable Scenarios
CNN	Extracts spatial features from images	Suitable for large-scale image data and efficient	Limited ability to handle time-series data	Molten pool morphology monitoring and defect detection
RCNN	Region-based target detection model based on CNN	Precisely locates target areas	High computational complexity and slower speed	Small defect detection and molten pool feature prediction
GAN	Data generation and enhancement; adversarial training	Generates high-quality data and solves data imbalance issues	Unstable training and sensitive to parameters	Data augmentation and anomaly detection
LSTM	Processes time-series data	Suitable for dynamic processes and strong memory ability	Inefficient for high-dimensional data	Molten pool temperature dynamic monitoring and real-time feedback

summarizes the characteristics, advantages, disadvantages, and application scenarios of common ANN models (e.g., CNN, RCNN, and GAN) in DED monitoring. When selecting an ANN model, it is essential to weigh factors such as the specific application scenario and requirements. For instance, time-series models like LSTM are more suitable for real-time molten pool dynamic monitoring, while RCNN is better for detecting complex shapes or small defects with higher precision. Future research can focus on decision-making criteria for model selection, including data types, computational resource constraints, and industrial application demands.

4.2. Temperature Characteristics

In the field of temperature monitoring for metal AM, infrared cameras are one of the most widely used and effective sensors. The earliest use of infrared cameras for DED temperature monitoring was by Griffith et al. from Sandia National Laboratories, who conducted studies on laser-engineered net shape (LENS). The thermal images captured in their study only represented the temperature distribution of the molten pool and its surroundings in a simplistic manner. Compared to the early understanding of molten pool features by Griffith et al., recent studies have delved deeper, developing various methods and techniques. Jeon et al. [74] used coaxial infrared cameras to capture temperature images to measure the width and length of the molten pool at specific locations. They also measured the printing height and deposition trajectory profile at the same position using a laser linear scanner. The extracted features from the measurements were then input into an artificial neural network (NN) to develop an online molten pool depth estimation technique for the DED process. The overall accuracy was about 25.97%. High-precision infrared cameras with a wide temperature range offer high imaging quality, more accurate measurements, and broader applicability. However, their high cost has prompted researchers to attempt to design optical paths based on specific experimental conditions. Hao et al. [75] designed an optical path for high-speed temperature field measurement based on dual-wavelength temperature measurement technology (as shown in Figure 16a) and proposed a subpixel-accuracy dual-band image matching method. Compared to commercially available infrared camera products, their system had a measurement error as low as 1%, significantly reducing development costs while enabling real-time monitoring of molten pool temperature distribution during the DED process. Following the use of single infrared cameras for monitoring the AM process, multi-infrared camera systems have gradually emerged, providing more comprehensive and richer information. Herzog et al. [76] developed a novel system for monitoring DED-formed components (as shown in Figure 16b), consisting of three infrared cameras positioned along different optical axes. This system is capable of tracking the temperature, geometric shape, and vertical displacement of the molten pool throughout the metal printing process. By integrating data from the infrared cameras, an automated algorithm was developed to identify structural features and defect formation in the components. However, the data transmission speed of this system does not yet meet the requirements for real-time monitoring, which will need to be addressed through subsequent streamlining and downsampling.

In addition, pyrometers, CCD cameras, thermocouples, temperature sensors, and laser displacement sensors have also been applied to AM process monitoring. Among these sensors, with the deepening of molten pool mechanism research and the improvement of equipment accuracy, the use of pyrometers to monitor molten pool temperature has become widely adopted. Stockman et al. [77,78] utilized coaxially mounted dual-color pyrometers to collect temperature data during a single-layer dual-path printing process. They compared the porosity data obtained via X-ray computed tomography with the temperature-based data derived from thermal analysis to assess the effectiveness of anomaly detection techniques. Meanwhile, Bernhard et al. [78] used pyrometers to sample the molten pool temperature and proposed a method for determining anomalies in the parts based on temperature data from the AM process. Experimental results showed that this approach was able to identify internal defects such as porosity and voids in the component structure. Different from the above-mentioned temperature measurement methods, the thermal history of the molten pool during the metal AM process is also a crucial factor in process monitoring. Muvval et al. [79] used an infrared pyrometer to monitor the thermal history of the molten pool during the cladding process. By real-time monitoring of the molten pool’s thermal history and physical state during laser power modulation, they were able to optimize the microstructure of the component under the best processing parameters.

With the development of monitoring technologies and the updating of monitoring tools, molten pool temperature monitoring has become more precise. Leung et al. [80], in their research published in Nature Communications, demonstrated that high-speed synchrotron X-ray imaging technology can effectively uncover potential thermophysical phenomena during the molten track deposition process (as shown in Figure 17a). Their work revealed mechanisms such as Marangoni-driven flow, laser re-melting of porosity, and the dissolution and diffusion of pores, developing a mechanistic diagram to predict the evolution of molten pool characteristics and changes in the trajectory morphology. Zhang et al. [81,82] proposed a modified heat source model based on electromagnetic wave theory to study molten pool temperature. They calculated the heat transfer reduction coefficient of the beam power based on the modified model, which then allowed them to derive the temperature distribution and thermal history. Their work explored, from a mechanistic perspective, the interaction between the laser and powder particles. Khanzadeh et al. [72] employed an infrared thermography camera to monitor the temperature distribution on the top surface of the molten pool (as shown in Figure 17b). Using X-ray tomography (XCT) experimental methods, they determined the porosity of Ti-6Al-4V thin-walled samples. By combining the data from the infrared thermography with XCT measurements, they conducted a comprehensive analysis and compared it with the porosity location predicted from molten pool temperature data analysis. The results indicated that, after selecting an appropriate SOMs model, the proposed method could accurately predict the location of porosity.

In summary, temperature-based monitoring methods measure the temperature of the molten pool and its surrounding areas, as well as the temperature distribution between the layers of the workpiece. These measurements allow for the dynamic adjustment of important process parameters to maintain a constant temperature, thereby improving the quality of the part and ensuring the repeatability of the process. Currently, molten pool temperature dynamic monitoring mainly focuses on tracking temperature data during the manufacturing process. Sensors capture temperature data, which are then analyzed to monitor the quality of the part’s formation and detect manufacturing defects. However, there is a strong coupling relationship between molten pool temperature variations and visual features of the molten pool. Changes in temperature affect the geometric morphology of the molten pool, which, in turn, influences the formation quality of the part. Additionally, different process parameters and processing techniques also impact the temperature characteristics of the molten pool. At present, some research has been conducted on the dynamic monitoring of molten pool temperature. However, future work should not only continue to monitor the molten pool temperature but also further analyze the variation patterns of molten pool temperature characteristics and the influence of temperature gradient distribution on molten pool visual features. This approach will help better reduce the probability of defects during the manufacturing process and improve the quality of the parts being formed.

With the further development of metal AM technology and the advent of the big data era, AI technologies will be applied to molten pool temperature monitoring in metal AM processes. Traditional molten pool temperature monitoring methods struggle to meet the demands of big data analysis for molten pool temperature characteristics. Therefore, after acquiring molten pool temperature data through sensors, AI algorithms can be used to extract and analyze the molten pool temperature features, thereby constructing an intelligent molten pool temperature dynamic monitoring system. This system will enable the intelligent monitoring of molten pool temperature dynamics throughout the metal AM process. This approach will significantly advance the improvement and refinement of molten pool temperature monitoring, as well as promote the development of molten pool temperature monitoring technologies in metal AM. By leveraging AI for data analysis and decision-making, it will be possible to better understand and control the molten pool’s behavior, ensuring higher part quality, fewer defects, and more stable manufacturing processes.

4.3. Morphological Characteristics

The stability of molten pool depth is crucial for determining the forming quality in DED processes. In order to enhance the real-time prediction capability of molten pool depth, Tang et al. [83] proposed a method based on coaxial vision characteristics and molten pool convection state. Specifically, the method begins by analyzing the coaxial visual monitoring characteristics of the molten pool, dividing the image into convection, transition, and boundary regions (as shown in Figure 18a), representing the convection state based on these region features. Then, the relationship between process parameters, coaxial visual characteristics, and molten pool depth is established, enabling the prediction of molten pool depth under different convection states. For monitoring particle behavior during manufacturing processes, the calibration of the discrete element method (DEM) is crucial for simulating inter-particle interactions. Marco et al. [84] propose a calibration method for sticky particles that substantially improves the accuracy of the simulation by introducing a model for the adhesion forces between particles. This method provides a new tool for the study of particle flow properties in DED processes and helps to optimize powder transport and molten pool dynamics. This approach provides new tools for particle flow characterization studies in DED processes and helps optimize powder transport and molten pool dynamic characterization monitoring. Akbari et al. [85] developed an adaptive PI controller with hierarchical control gains (overall system structure shown in Figure 18b) to study the nonlinear response of the laser deposition process. This controller ensures that the molten pool width remains constant throughout the build process, thereby producing a more uniform and refined microstructure. In this system, molten pool width is used as a control input, while laser power is the control output, achieving real-time monitoring and closed-loop control of molten pool width. Ning et al. [86] proposed a real-time layer height prediction and parameter feedback framework to compensate for height inconsistencies during thin-wall component formation. This method enables the prediction of deposition height for the next layer, and experimental results demonstrate its practicality and applicability in real-time height consistency compensation. The surface flatness improvement can reach up to 22.8%. These advancements contribute to improving the precision of molten pool control and the overall quality of parts produced by DED, providing stronger support for high-quality, reliable manufacturing processes.

In addition to molten pool depth and width, the stability of deposition height is another key parameter that reflects the forming quality of DED processes. Donadello et al. [87] improved the triangulation method and proposed a new deposition layer height monitoring system. The system calculates the actual thickness of the deposition layer through triangulation and converts the measurement values into spatial coordinates to obtain a 3D spatial map of the deposited structure. By transmitting the measurement error signals to a feedback control system, process parameters are adjusted based on these error signals, which improves the forming quality. Subsequently, they further investigated the relationship between powder feed rate and deposition height stability during the DED process [88]. For this purpose, the researchers developed a new monitoring system that integrates layer height information obtained from coaxial optical triangulation with online measurement of the sample weight (as shown in Figure 19). This system enables more accurate monitoring and control of the deposition height, ensuring better process stability and quality. Additionally, Garmendia et al. [89] utilized a developed algorithm to precisely control the distance between the nozzle and the substrate, achieving dynamic adjustment of the layer height. This approach reduces the need for manual supervision and minimizes the occurrence of defective parts, thus ensuring better quality of the formed parts and improving the repeatability of the process. These advancements in deposition height monitoring and control contribute significantly to the overall quality and precision of DED manufacturing, offering more reliable and efficient means of achieving consistent layer deposition during the build process.

In order to obtain the morphology information of the molten pool for monitoring the metal AM process, the research has gone through a series of works, from the initial use of optical devices to monitor the morphology of the molten pool to real-time calculation and acquisition of the molten pool morphology information, to the accurate localization of the molten pool position information and the use of a low-coherence light source to increase the accuracy of the molten pool information, etc., and the monitoring system of the molten pool has been gradually developed and improved.

Compared with the traditional monitoring of pool geometry, such as length, width, and area, monitoring pool stability can provide more effective information. The information inside the pool with the same length, width, and area may be completely different, and monitoring a single pool geometry cannot capture the internal changes in the pool. Therefore, the traditional monitoring of the melt pool geometry can no longer meet the quality requirements. Obtaining more effective information on the dynamic characteristics of the molten pool through intelligent algorithms for the distribution of the molten pool intensity of feature extraction is not a better way. The distribution of the intensity of the molten pool can not only be extracted to the molten pool length, width, and area and other geometric features but it can also access to the quality of the formation of the molten pool of internal high-intensity has a great impact on the quality of the bubbles and droplets, as well as the molten pool internal high intensity. From the molten pool intensity distribution, we can not only extract the geometrical features such as length, width, and area of the melt pool but also obtain the information such as bubbles and droplets in the molten pool, which have a great influence on the forming quality, as well as the characteristics of the molten pool spattering, which reflect dynamic changes in the molten pool.

In addition, some of the above monitoring methods tend to only monitor one type of molten pool information, which makes the stability of the monitoring system poor, so the integration of a variety of molten pool information comparison is particularly important. All-round three-dimensional molten pool morphology monitoring will also increase the stability and reliability of the monitoring system.

4.4. Spatter and Plume Characteristics

The spatter and plume characteristics in the metal AM process, to some extent, determine the quality of the final product. In terms of obtaining raw data, high-speed imaging technology demonstrates significant advantages in real time, capturing the dynamic spatial and temporal changes in spatter and plumes [90,91].

Currently, several spatter and plume detection methods have been reported in the literature, including traditional computer vision, simple machine learning combined with image processing, and complex deep learning models. A typical example is the work by Tan et al., who studied the characteristics of spatter in PBF and used an image segmentation-based NN to extract the spatter [92,93] (as shown in Figure 20a). The results showed that the proposed method could extract 80.48% of the spatter within 70 ms. Criales et al. [94] investigated the percentage of spatter in each frame of images captured by high-speed cameras and found that the spatter percentage was over 20%, reaching up to 70% in some cases. In addition, Repossini et al. [95] developed a logistic regression model to analyze spatter characteristics and classify different energy density conditions corresponding to various quality states (as shown in Figure 20b). By treating spatter as a process feature-driving factor, they observed a significant improvement in the ability to detect unmelted and over-melted conditions. However, due to limitations in image capture methods and image processing algorithms, only three spatter characteristics were studied. In a later study, Baumgartl et al. [96] proposed a CNN architecture with depth-separable convolutions to train thermal imaging data (as shown in Figure 20c). The study found that DL outperformed traditional computer vision techniques in spatter detection, but the limitation of this method was that it could only detect spatter and layering. DL models such as YOLO (you only look once) are widely used to detect and classify the dynamic properties of the molten pool in the AM process. Chebil et al. [97] developed a DL architecture (YOLOv4) to address the specificity of spatter detection, establishing a fast and automatic control method to assess the stability of the PBF process, using spatter as a process feature. The methodology provides a reference for real-time monitoring of AM processes. However, these studies still suffer from insufficient optimization of specific process parameters.

5. Challenges of Intelligent Monitoring

The stable consistency of the DED process and forming quality has been a challenging problem for its scale-oriented production, as shown in Figure 21. With the rapid development of big data, advanced sensing, AI, and other technologies, intelligent monitoring technology based on acoustic–optical–thermal–magnetic information and ML algorithms has been gradually applied to DED, but for large-scale production, its monitoring means are still limited, and the maturity of quality judgment and regulation technology is not enough. Therefore, combining advanced sensing technology and AI methods to enhance the intelligent monitoring technology and develop the intelligent monitoring system for the scale production of DED encounters new challenges.

Clarifying the key mechanisms of defects is challenging. Most existing studies focus on a single type of defect, establishing its mapping relationship with process information. However, quality issues in AM are often the result of multiple defects interacting with each other, and different types of defects may also influence each other. Therefore, a comprehensive evaluation of various defects in the manufactured parts is needed to obtain more reasonable assessment results. Additionally, most studies so far have focused on internal defects, with relatively fewer studies on defects like cracks. Further research is still needed to explore the generation mechanisms, monitoring methods, and control mechanisms of crack defects.

The fusion of multi-source heterogeneous information is challenging. Although most studies have used different information sources for defect monitoring and quality assessment, they are mostly based on a single type of signal. In reality, the formation of defects in the metal AM process is highly complex, and a single information source may not fully characterize the defect. The AM process contains rich multi-scale heterogeneous information, and it is essential to develop new monitoring techniques, effectively integrate multi-source heterogeneous information, and enhance the accuracy and robustness of quality assessment and control. In the medical field, one of the major challenges in multi-source data fusion is the problem of aligning heterogeneous data, such as aligning different modalities (e.g., CT images with MRI data) for joint analysis. By introducing a deep learning-based cross-modality alignment algorithm, researchers have successfully improved the accuracy of data fusion. Similarly, in the field of autonomous driving, time synchronization and noise filtering of sensor (e.g., LiDAR vs. camera) data are major difficulties. Techniques based on timestamp correction and Bayesian filtering are used to solve these problems. These methods provide useful references for the fusion of acoustic, optical, thermal, and magnetic signals in the DED process. For example, time-synchronization techniques can be used to ensure the consistency of multi-physics field signals, while deep learning-based alignment methods can optimize the accuracy of cross-modal feature extraction.

The interpretability of quality assessment is challenging. While many studies have established mappings between process information and various defects in metal AM based on DL models, the major drawback of these intelligent quality assessment models based on DL is their lack of interpretability. This limitation is particularly prominent in risk-sensitive fields like metal AM, often leading to inaccurate quality assessments. Therefore, it is necessary to develop explainable intelligent models that incorporate physical knowledge, aiming not only to identify defects but also to reveal the causal relationships between process information and various defects.

The interaction and regulation of multiple parameters is challenging. Existing quality control methods mostly focus on single parameters and single-objective designs without considering the comprehensive impact of multi-parameter interactions on the quality of components in DED. In reality, the DED process involves multiple interrelated process parameters and is susceptible to various disruptive factors, making the optimization of these parameters a significant challenge. Additionally, existing simplified state identification models and single-input, single-output control systems cannot meet the complexity of DED, leading to potential issues where improving one aspect may sacrifice another, ultimately failing to ensure the stability and consistency of the final part quality. Therefore, there is a need to develop multi-parameter interaction-based multi-objective process parameter optimization and control methods, particularly for effective quality control within the limited time and layers of deposition.

6. Future Outlook

In multi-source data fusion, the DED process involves the interweaving of acoustic, optical, thermal, and magnetic signals, and it is difficult to accurately characterize the process defects with a single piece of information, so the fusion of multi-source data containing rich information can maximize the process characteristics and improve the accuracy of quality assessment. To this end, the following work can be carried out:

• Establishing an effective database and utilizing data-layer fusion methods to complement the advantages of multi-source information.
• Learning and fusion of multisource data representations to extract rich defect features from the data, achieving feature-level fusion.
• Multisource data alignment and collaborative learning to align the corresponding relationships between different data sources and defect representations, transferring knowledge learned from information-rich sources to information-scarce sources, and enabling mutual assistance in the learning of each information source.
• Based on data fusion experiences in other fields, the keys to multi-source data fusion in DED processes are ensuring temporal and spatial alignment of sensor data to reduce error accumulation, developing downscaling and dimensional feature extraction methods for high-dimensional data to optimize computational efficiency, and using physical knowledge to guide data fusion algorithms and enhance model interpretability. These solutions provide new technological paths to realize efficient DED process monitoring.

Interpretable quality intelligent judgment methods must overcome the “black box” attribute of traditional deep learning-driven methods, which is of great significance in enhancing the reliability of quality judgment and revealing the causal relationship between process information and performance quality assessment. It is proposed to carry out research in the following aspects:

• Causal representation learning: Achieving quality assessment while uncovering the intrinsic causal relationships between process information and defects.
• Integration of physical knowledge in quality assessment: Embedding physical principles into the assessment models to ensure that the models align with certain physical laws.
• Post hoc interpretability and visualization methods: Analyzing the mapping relationships between assessment results and information sources, facilitating post hoc explainability and visualization of the decision-making process.

Multivariable interaction-based quality control: In the process of controlling the relationship between process information, macro/micro defects, and performance quality, the control approach shifts from a single-parameter control to a multidimensional parameter control. This utilizes the strong interdependence between various control parameters to achieve high-precision online quality control. The following research work is recommended:

• Offline optimization of multiple process parameters driven by multiple performance indicators.
• Providing an operational window for process parameters for online control.
• Development of multi-input, multi-output adaptive control methods to eliminate the “trade-off” problem inherent in single-input, single-output control systems.
• Machine learning-driven online control, integrating the advantages of reinforcement learning in environmental interaction.

Establishment of forming process monitoring standards: For large-scale production of DED-manufactured parts, it is essential to establish reasonable and practical standards and industry norms for forming process monitoring. This may involve the following research areas:

• Development of online monitoring indicators and quality evaluation system: Establishing standards for DED online monitoring indicators and the construction of a quality evaluation system, leading to the formation of DED manufacturing process monitoring norms and quality grading standards.
• Establishment of control boundaries based on quality control mechanisms: Defining the boundaries for quality control based on process control mechanisms and developing constraints and standards for DED process parameter optimization and quality control.

Development and deployment of an online monitoring system: The construction of a DED monitoring platform that can quickly and effectively sense multi-source information such as sound, light, and heat during the manufacturing process. The platform should be capable of conducting real-time quality assessment, enabling effective control of process parameters such as powder spreading and laser power. Additionally, the development of a modular, combined AM monitoring hardware and software system is crucial to ensuring quality consistency for large-scale DED manufacturing.

With the development of ML and intelligent monitoring technologies, future research can focus on the application of adaptive control systems in DED. These systems can dynamically adjust process parameters (such as laser power, scanning speed, etc.) in real time to accommodate changes in the environment and variations in workpiece characteristics. Specifically, research can concentrate on the following areas:

• Feedback control-based adaptive parameter adjustment algorithms for real-time optimization of molten pool dynamic characteristics.
• Integration of reinforcement learning with traditional PID control to enhance the system’s response speed and stability.
• Development of ML-based adaptive control systems that can automatically learn and adjust control strategies under different process conditions.

With the industrialization of DED technology, the standardization of monitoring systems and their indicators will be crucial for improving production efficiency and product consistency. Future research should focus on defining and verifying the following standardized metrics:

• Standardized measurement methods for key physical parameters such as molten pool depth, width, and temperature distribution.
• Unified standards for part quality assessment, such as surface roughness, internal defect rate, and interlayer fusion quality.
• Real-time monitoring indicators based on multi-source data fusion, ensuring process consistency and part traceability.

By establishing these standardized indicators, a quantifiable quality assurance framework can be provided for the industrial application of DED technology. By strengthening the research on adaptive control systems and formulating standardized indicators for DED monitoring, future studies will offer more practical and systematic solutions for improving the stability, precision, and productivity of DED processes. It is hoped that these research directions will provide a more concrete pathway for academic research and industrial applications in this field.

7. Conclusions

The molten pool in DED is a complex physical field, and the changes in its temperature characteristics are influenced by multi-source information factors. There is still significant room for development in the dynamic monitoring of molten pool temperature. Furthermore, after years of development, online monitoring technology for molten pool morphology in DED has made significant progress. It is now capable of real-time acquisition of morphological information such as length, width, and area of the molten pool, which is applied to process feedback control. However, the monitoring of molten pool morphology still has many areas that need improvement. Compared to traditional parameters such as length, width, and area, the molten pool intensity distribution contains more valuable and effective information, and the extraction of molten pool intensity distribution data is expected to be a key trend in future development.

Currently, most studies obtain relatively simple molten pool information, with issues related to reliability, robustness, and the inability to process molten pool data in real time. However, multi-source heterogeneous information fusion and ML algorithms offer potential solutions to these problems. In future research, a multi-sensor information fusion monitoring platform can be developed, with real-time molten pool data processing models built based on ML algorithms. This approach will enable the extraction and recognition of characteristic information from the DED process, providing theoretical and methodological support for improving the quality stability and process repeatability of DED parts. Dynamic monitoring technology in DED processes has been continuously developing, and in the future, intelligent monitoring based on ML and multi-source heterogeneous information fusion will be an important direction for dynamic monitoring.

The main contributions of this paper include (1) proposing an intelligent monitoring framework for DED based on the fusion of heterogeneous information from multiple sources, (2) developing a physical knowledge-guided explanatory machine learning model for revealing the causal relationship between process information and defects, and (3) exploring an intelligent optimization and control method for multivariate interaction. These innovations provide theoretical and methodological support for improving the consistency and stability of DED process quality and lay the foundation for large-scale production applications.