*3.3. Balling*

Another common defect in the L-PBF process is called "balling", which leads to disruption in interlayer bonding, poor surface quality, and porosity in the final part [36]. It occurs mainly in high scan speeds, which leads to an elongated melt pool that breaks into small beads (10 μm) to reduce the surface tension due to Rayleigh instability. Additionally, partially melted powder due to low input energy leads to limited melt formation and produces coarser sized beads (500 μm). Oxide layer or presence of moisture on the powder surface can also disrupt the melt pool surface tension, resulting in balling. These impurities reverse the capillary motion at the melt pool surface, and positive surface tension gradient forces the melt pool to flow inwards to the center, which develops unfavorable wetting conditions and leads to spheroidization. Remelting/double exposure or deoxidizing agents can significantly reduce the balling phenomena [37,38]. Li et al. [39] presented the SEM images of the balling phenomena for stainless steel w.r.t di fferent scanning speeds in the L-PBF process (as shown in Figure 3d). Other authors, for example [38,40], reported the balling phenomenon in L-PBF for di fferent materials.

## *3.4. Cracks*

The residual stresses are the stresses that remain in the part after fabrication; these stresses are significantly detrimental to the mechanical strength of the final part. Higher thermal gradient due to rapid heating and cooling cycles develops residual stresses in the part and results in delamination, part distortion, and warping [41,42]. The two mechanisms responsible for the evolution of residual stresses are temperature gradient mechanism (TGM) and rapid solidification mechanism. During TGM, the pre-solidified layer underneath the melted layers is heated up very rapidly upon laser exposure, which tries to expand but is restricted by the cold solidified material. This restriction in thermal expansion develops compressive residual stresses at the regional heat a ffected zone (see Figure 4a) [43]. During the solidification mechanism, the melted zone cools very rapidly, and shrinkage is again restricted by the underlying layer, developing tensile residual stresses in the upper layer (Figure 5a). Therefore, the cyclic heating and cooling phases during printing results in larger tensile residual stresses in the top surface layers and residual compression stress at the bottom of the printed part (Figures 4b and 5b) [43].

**Figure 4.** (**a**) Temperature gradient mechanism during laser heating, (**b**) representation of the evolution of thermal stress and strains in the irradiated zone [43].

**Figure 5.** (**a**) Evolution of stresses during solidification phase, (**b**) representation of the evolution of residual stresses and strain in the irradiated zone [43]. Reused under Creative Commons Attribution License.

At a given time and temperature, when the tensile residual stresses cross the ultimate tensile strength of the material, it leads to cracking in the part [44]. As reported [45], the crack initiation and propagation start on the as-built surface as it adhered to the partially melted material and leads to interlayer delamination, also called interlayer cracking. Materials like nickel-based alloys, copper, and stainless steel are highly susceptible to cracking due to their low thermal conductivity [46–48]. The pre-heating of the build plate is often applied to reduce the e ffect of the residual stresses in the final part.

#### *3.5. Part Distortions*

Another critical aspect of residual stresses on the overall quality of the part health is "part distortion". If the thermally induced local residual stresses exceed the yield stress of the material, it can lead to part distortion in the middle of the process. Part distortion can also damage the recoater permanently. Kruth et al. [49] reported the warping in the SLMed parts due to high thermal fluctuations experienced by the material in a short interval of time. The relaxation in the residual stresses after the removal of the finished part from the base plate and support structures, leads to macroscopic distortion in the final part (as shown in Figure 3e). Mugwagwa et al. [50] studied the influence of process parameters such as laser power, scanning speed, scanning strategies on the evolution of residual stresses, and part distortion experimentally and with finite element modeling. Sometimes, the warping of the parts can lead to failure of the bolts of the build plate as well. Mishurova et al. [51] studied the influence of support structures on part distortion using the X-ray di ffraction method. The stress redistribution takes place for the samples with support structures after removal from the base plate, whereas for parts without support structures, only simple stress relaxation observed. They have concluded that the support structures decrease the amount of residual stresses in the part.

Other defects or anomalies such as surface defects, material contamination, and geometrical inaccuracy are not discussed in this article. However, interested readers can refer to references [52–55].

#### **4. Origin of Defects**

L-PBF process involves a large number of critical parameters that can influence the quality of the final part and leads to the above mention defects if not optimized. Sharratt et al. [26] categorized the influencing parameters into three categories, which are (i) process defects, (ii) equipment-induced defects, (iii) design defects. Equipment defects are basically due to aberration in the equipment of the machine such as f-theta lens aberration, improper calibration, wear of recoating blade, etc. Nowadays, due to significant technological advancements, most of the commercial machines are well-calibrated and standardized to minimize such defects. Design defects are due to improper optimization of the design as per AM standards such as lack of support structures, overhang angle, and bad orientation. These defects can be minimized by using commercial software such as AdditiveLab, Simufact, Virfac, etc., which helps to find the best support structure placement based on the thermomechanical simulations.

A lot of time and money is consumed to find best-fit process parameters such as laser power, scan speed, hatch distance, layer thickness, etc., to reduce the third type of defects called "process defects". All the parameters mentioned above can be combined in the empirical formula of 'Energy density (*E*)' given by Equation (1), where *P*, *v*, *h*, and *t* represent laser power, scan speed, hatch distance, and layer thickness, respectively.

$$E = \frac{P}{v \times h \times t} \tag{1}$$

Each paramete in Equation (1) has its e ffect on the final quality of the part. For example, at lower scan speed and high laser power, the energy density is higher, and its attributes to porosity defects are due to entrapment of gas. Thijs et al. [27] stated that high energy input in Ti6Al4V powders leads to large melt pool size, which causes denudation. Thus, not enough material is available to fuse the adjacent scan vector, which leads to bigger pores in the part. The higher residual stresses are also reported for low scan speed and higher laser power [46].

Similarly, at low laser power and higher scan speed, a lack of fusion defects is more pronounced due to a lack of energy input to fully melt the powder layer [29,35,56]. Sun et al. [57] showed the relationship between the part density and input energy density for Al0.5CoCrFeNi High Entropy Alloys (Figure 6). It can be observed that for lower energy density, the highest porosity is reported. Therefore, it is crucial to optimize the process parameters to attain the maximum dense parts.

**Figure 6.** Relative density as a function of energy density [57]. Reused under Creative Commons Attribution License.

It is important to note that the scanning strategy also makes an impact on the overall quality of the part. Although it has not been considered in the energy density equation, the scanning strategy directly influences the heat balance of the part. For example, Aboulkhair et al. [30] showed the relation between the scan speed and different scan strategies, as shown in Figure 7. The different strategies used are: "2X" is similar to "X" but each layer is scanned twice, "X&Y 2HS" means that each layer was scanned twice having each scan pass perpendicular to the one before and with different hatch spacings for each scan, "pre-sinter" means to first scan the layer with half the power followed by a second scan with full power, "overlap" strategy indicates where each layer was scanned twice with the second scan melting the overlap between each two adjacent melt pools, and "alternating" strategy indicates scanning each layer in a direction rotated by 90◦ to its precursor. As can be concluded from Figure 7, the pre-sinter scanning strategy and scanning speed of 500 mm/s report a high density of AlSi10Mg parts.

#### **5. Melt Pool Signatures**

Laser–material interaction in L-PBF is a complex thermophysical phenomenon occurring in a short interval of time. The shape and size of the melt pool are significantly influenced by the process parameters and geometrical factors as well. Therefore, melt pool signatures provide essential information about the quality of the part. A different range of electromagnetic spectrum emanates from the melt pool, which serves as the input variable for in situ sensing devices. The types of electromagnetic radiations observed from the melt pool are shown in Figure 8. As we know, scattering, absorption, and diffraction of the laser light are the common phenomena that occur. Hence, the backscattering of the laser light is a common phenomenon that occurs in L-PBF. Therefore, it is vital to eliminate the backscattering of the laser light in the in situ sensing module by installing correct mirrors. Thermal emissions from the melt pool occur at the wavelengths that correspond to the melt temperature. High laser energy can lead to the evaporation of melt and ionization of the gas, which results in the formation of a plasma plume above the melt pool. The electromagnetic radiation from the plasma can also add noise to the captured signatures [58–61]. Sometimes, the plasma also hinders the laser beam, which results in low input energy to the powder bed. Thermal radiation, which ranges from 900 to 2300 nm, is captured to monitor the process and links it to the temperature field of the melt pool. Plasma radiations, which are in the range of visible wavelengths (400–600 nm), are eliminated [58,62]. These melt pool signatures are captured by using in situ sensing sensors, which are discussed in the next section.

Scanning Strategy

**Figure 7.** Effect of scanning strategies and scan speed on the porosity % of AlSi10Mg parts [30]. Reused under Creative Commons Attribution License.

**Figure 8.** Illustration of types of emitting radiation from melt pool [58].

#### **6. In Situ Sensing Devices**

As discussed in the last section, capturing the melt pool signatures is essential to monitor the process. Therefore, in this section, we review the working principle of the most used devices or sensors for in situ sensing of melt pool signatures in the L-PBF process. Generally, two entities related to the melting pool are measured, i.e., melt pool morphology (shape and size) and temperature. For temperature measurements, there are two types of sensors available, i.e., photon detectors, also called non-contact sensors, and thermal detectors, referred to as contact detectors. As can be seen in Figure 9, the thermal detector response is constant with the wavelength (radiations emanating from the melt pool), whereas the photon detector response was greatly influenced by the wavelength of the light. The linking of the photon detectors output to the melt pool temperature will be discussed in Section 7. Besides monitoring the melt pool signatures, other process signatures such as powder bed spreading and part geometry are also studied using different sensor techniques, such as ultrasonic, line profiling, optical coherence tomography, etc., as reported in references [63–65].

**Figure 9.** Wavelength response of detectors.

#### *6.1. Non-Contact Temperature Measurement Devices*

Most of the studies involve the monitoring of the temperature field or electromagnetic radiations from the melt pool both in a co-axial/on-axis and off-axis manner. The melt pool signatures capture the most invaluable information such as the temperature profile of the melt pool and the morphology of the melt pool that directly influences the quality the part [66–68]. Electromagnetic emissions are associated with the melt pool and plasma plume [27]. Therefore, to capture electromagnetic signatures, mostly non-contact measurement sensors are used. Broadly, these sensors can be classified as spatially integrated, i.e., photodiodes and pyrometers; spatially resolved, i.e., cameras; spectrally resolved, i.e., spectrometers [62,69]. Regardless of the sensors, the essential requirement for AM applications is the ability of the sensor to capture the rapid melt pool dynamics. Below, we discuss the working mechanism of the two most common photodevices used in the L-PBF process.

## 6.1.1. Photodiodes

Spatially integrated, single-channel detectors such as photodiodes are semiconductor devices that convert light (radiation) into electric current. The basic principle of the photodiode or photodetector is that when photons of sufficient energy (radiation) hit the active area of the device, it generates an electron-hole pair, which changes the effective resistance or conductance of the detector. A bias voltage is applied to collect the charge carriers, and the signal (photocurrent) is measured across a load resistor RL, as shown in Figure 10. The photocurrent is directly proportional to the intensity of the photons hitting the active area of the device. The intensity of photons is directly linked to the melt pool temperature. Photodiodes have the advantages and disadvantages of reducing the signal to a single data point, i.e., photovoltage corresponding to the amount of light hitting the photodiode. Reducing the signal to a single point relates to data shrinkage, which makes it easier to post-process. Low cost, high sensitivity, and high sampling rates (typically ~50 kHz) make them suitable for in situ sensing of the L-PBF process. The detection range of these devices is limited to over a range of the spectrum; for example, germanium (Ge) and InGaAs-based photodiodes are used for visible to infrared light detection. In contrast, silicon-based photodiodes are used for the UV and visible spectrum [70–73].

**Figure 10.** Schematic of working principle of PIN photodiode [70].

#### 6.1.2. Digital Cameras

The digital cameras, also called spatially resolved sensors, are an array of photodiodes where each photodiode is labeled as a pixel. The different types of photodiodes with their detection wavelength range are given in Figure 11. Each pixel detects the photon coming from the melt pool and converts it into photocurrent. Afterward, signal processing is done to extract a picture which corresponds to the temperature fields of the melt pool. Based on the detection range of the devices, it can be classified either as a regular digital camera that has detection range in the visible spectrum or a thermal camera, which has detection range in the infrared spectrum. Mostly, for monitoring melt pool signatures, thermal cameras are preferred with a detection range over 700 nm. Based on the signal processing method, the digital cameras can be classified into two categories, i.e., charge-coupled device (CCD) and complementary metal-oxide-semiconductor (CMOS). In a CCD camera, all the pixel information is processed by the single circuit, whereas, in a CMOS camera, each pixel has its processing circuit. Therefore, the CMOS cameras have better signal processing speed, but the capture area is limited compared to a CCD camera. The main disadvantage of using cameras for in situ sensing of the L-PBF process is the speed of the process, which requires a high frame rate camera that results in an enormous size of data that are difficult to post-process [74–76].

**Figure 11.** Wavelength sensitivity of common detectors.

#### *6.2. Contact Temperature Measurement Devices*

Another category of the devices or sensors used for monitoring the temperature fields in the L-PBF process is direct contact-based measurement devices such as thermocouples. The thermocouple is an electric device that produces a temperature-dependent voltage due to thermoelectric effect. The working principle of the thermocouple is based on the Peltier effect. The thermocouple consists of two electric conductors joined together to form two junctions. If junctions are attached to the surface at different temperatures, an electromotive force (EMF) is generated due to the temperature gradient between the two junctions of the circuit; then, total EMF can be calculated based on the temperature of the junctions and the properties of the electric conductors used. A point of measurement is connected to one of the junctions in the circuit, also called a hot junction. Another body of known temperature is connected to the other junction, also called a cold junction or a reference junction. A voltmeter measures voltage or current generated due to the temperature gradient between the two junctions. This voltage or current is then converted into the temperature [79,80]. Thermocouples can be used to monitor the overall heat flow of the part while printing. However, one limitation of the thermocouple is that it requires physical contact to measure the temperature, which is difficult in the L-PBF process and often requires extra preparation such as drilling holes or slots. Some of the researchers have monitored the L-PBF process using thermocouples, which can be found in the reference article [71].

#### **7. Temperature Correlation**

One of the key advantages of using photodiodes and digital cameras is the possibility to monitor the melt pool temperature fields or process signatures without any physical contact with the target surface. It provides the freedom to monitor the processes like AM, in which it is challenging to access the target surface. Another advantage of these sensors is the possibility to detect high temperatures such as 2000 ◦C. Although these sensors provide grea<sup>t</sup> flexibility and easy detectability, calibration of these sensors properly is a vital issue. The main principle of the sensors is to generate a photocurrent or photovoltage based on the emitted radiation from the melt pool. Therefore, it requires much understanding about the process and complex thermophysical phenomena occurring in that short interval of time. Usually, the melt pool is compared with the black body phenomenon to compute the actual temperature. However, in reality, the melt pool does not behave as an ideal black body rather than a real body. In addition, the emissivity of the material concerning di fferent temperatures is not known. Therefore, a better understanding of black body phenomena is inevitable. Here, we present an insight into black body radiation and their correlation with the melt pool physics in L-PBF.

All bodies above absolute zero (0 K) emit thermal radiation due to the thermal vibrations of atomic particles. The quantity of the power emitted is expressed by the radiative intensity. The total energy is distributed across the electromagnetic spectrum. The emission intensity and shape of the distribution is dependent on the temperature. Conversely, the temperature of the body can be calculated by the quantity and spectral distribution of its emitted energy across the spectrum.

Firstly, a black body is a physical body that absorbs all the incident electromagnetic radiation and emits electromagnetic radiation called black body radiation. The radiation emitted by the black body is dependent on the temperature alone, not on body shape and composition. For an ideal black body at a specific temperature (T), spectral radiance (*B*) is given by Planck's law (Equation (2)), where h is Planck's constant; *c* is the speed of light; *KB* is the Boltzmann constant.

$$B = \frac{2hc^2}{\lambda^5} \left( \exp\left(\frac{hc}{k\_B \lambda \, T}\right) - 1\right)^{-1} \tag{2}$$

Figure 12 shows the spectral radiance of ideal black body radiation at di fferent temperatures. It can be seen that the radiant intensity increases with increasing temperature at all wavelengths and for all temperatures. For any given temperature, the radiant intensity reaches a maximum at a specific wavelength, and there are two wavelengths where the radiant intensity is equal for intensities less than the maximum or peak intensity. The wavelength at maximum radiant intensity decreases with an increase in temperature (Wien's displacement law) [70].

**Figure 12.** Spectral radiance of an ideal black body at di fferent temperatures [70].

Nevertheless, real surfaces do not emit radiations like an idealized black body. The fraction of the emitted radiation from the real body is relative to a black body, which is known as spectral emissivity. However, spectral emissivity is temperature and wavelength-dependent. Therefore, it is challenging to obtain spectral emissivity of every material for each temperature and wavelength. Therefore, the direct application of Planck's law for temperature correlation cannot be used. Instead, to overcome this challenge, there are three possibilities which are as follows:


Mostly, the ratio method is used to calculate the actual temperature. In the ratio method, the two intensity measurements (*I*1 and *I*2 ) at two di fferent wavelengths λ1 and λ2 are calculated. The ratio of intensities *I*1 and *I*2 are taken to calculate the temperature as given by Equation (3).

$$\frac{I\_1}{I\_2} = \frac{A\_1 \varepsilon\_1 \lambda\_2^5}{A\_2 \varepsilon\_2 \lambda\_1^5} \exp\left(\frac{hc}{k\_B T} \left(\frac{1}{\lambda\_2} - \frac{1}{\lambda\_1}\right)\right) \tag{3}$$

The di fference between the two wavelengths (λ1 and λ2) chosen is very small so that it is assumed that the emissivities (1 and 2) are equal. Once the constants *A*1 and *A*2 are known, the calibration curve can be calculated, which relates the intensity ratio to the temperature of the object irrespective of the emissivity. Therefore, the choice of the two wavelengths is very critical. Because too close wavelengths can lead to low signal and noise ratio from the sensors, i.e., the measurement will be dominated by the noise. If too distinct wavelengths are chosen, the assumption of equal emissivities will no longer be valid. Most of the engineering materials have a melting point above 1000 K and vaporization temperature around 3000 K. Therefore, most of the monitoring devices have detection ranges in the visible-near infrared region. Typically, for calibration of the sensors and to find the values of the constants *A*1 and *A*2, the tungsten halogen light source with known spectral radiance is used. The halogen lamp is placed inside the build chamber, and the spectral radiance measured by the sensors is then compared with the radiance obtained by the calibrated spectrometer. Then, the *A*1 and *A*2 values are computed [70].

Moylan et al. [77] also used a multispectral method to calculate the imaged temperature from captured infrared images using a hyperspectral camera in the L-PBF process. The imaged temperature can be calculated by using this method. It shall be noted that the imaged temperature is di fferent from the true temperature, as the true temperature is highly dependent on the emissivity of the material, which is subjected to the temperature.

#### **8. Overview of ML Techniques**

In the L-PBF process, the quality of the final part and the reliability of the whole process are highly dependent on various process parameters, as discussed in Sections 3 and 4. The above-discussed process parameters incorporate inconsistency in the part quality, which is di fficult to monitor. The inconsistency during the process can be e ffectively controlled and monitored by using in situ monitoring systems. However, there are numerous problems associated with monitoring systems, which hinders the full exploitation of these systems. Firstly, the size of the data, real-time processing, and storage of the data are significant challenges. For example, consider a laser printing a part with geometrical tolerances of ±100 μm with a scanning speed of 100 mm/s. The sensors will have a data collection rate of ~10 kHz, i.e., one data point for every 10 μm to achieve a spatial resolution of 10 μm. Most of the researchers have reported a sampling rate of 50 kHz and used more than two sensing devices, such as a digital camera, photodiode, or combination of both. Therefore, if a single data value is stored in binary fashion, it occupies 32 bits. With a 50 kHz data sampling rate and with five sensing devices, the data will be stored at 1 MB/s. Therefore, if the machine runs for three days, it will generate at least 233 GB of data. The post-processability of this dataset and storage is difficult [78]. Secondly, linking the obtained in situ sensing data to the quality of the product is a significant challenge. In recent years, the research community is working on integrating ML methods into AM.

Machine learning (ML) is a subset of Artificial Intelligence (AI) that can learn and improve in an automated fashion from experience without being explicitly programmed. The process of learning starts with data observation and data pattern recognition to allow algorithms to learn automatically and adjust actions accordingly. The main aim of the ML is to learn and improve from data without human intervention. The ML algorithms are used in AM for design and process optimization, in situ monitoring, inspection and validation, and cost estimation. The data used for training the ML algorithms decide the effectiveness of the ML approach. In other words, the ML algorithms are as good as the training dataset [79].

ML algorithm can be broadly classified into three categories: supervised learning, unsupervised learning, and reinforcement learning. In supervised machine learning, the algorithm is trained with a labeled dataset. The labeled dataset is the dataset in which the output is known for the input variables. Therefore, the model is trained based on inferring the functional relationship between input and output variable. Supervised learning can be used for regression and classification problems. For instance, classifying the part quality as "good" and "bad" is a classification problem, whereas predicting the porosity level or mechanical properties of the part is a regression problem. In unsupervised learning, the output labels are not known for the input training dataset. The algorithm separates the training dataset into different clusters based on the relationship among input data. Unsupervised learning is useful in clustering problems such as detecting anomalous conditions. The third ML category, called reinforcement learning (RL), is learning by interacting with an environment. The RL algorithm learns from the consequences of its actions, and it selects its actions based on its experience and new choices, which is trial and error learning. The RL algorithm receives a numerical reward which encodes the success of an action's outcome. The RL algorithm seeks to learn to select actions that maximize the accumulated reward over time. Applications of the RL algorithm are game theory and self-driving cars. The list of various ML algorithms is shown in Figure 13 [80].

There is another type ML algorithm called "semi-supervised learning", which combines unsupervised and supervised learning. Semi-supervised learning combines the small or limited amount of the labeled dataset with a large amount of the unlabeled dataset during training the algorithm [81]. This type of learning is beneficial in AM applications, where obtaining the labeled dataset is a very challenging task. The decision for using a type of ML approach is based on the available dataset, and specific output preferred. The use of the abovementioned ML approaches for monitoring the L-PBF process is discussed in the next sections.

**Figure 13.** Schematic of machine learning algorithms.

#### **9. Problem with ML Techniques**

Although there are tremendous possibilities of the ML approach in the AM industry, there are a few limitations or problems associated with the AM methods that can have a deterministic effect on the ML outcomes.


4. Machine learning training problems: Overfitting and underfitting are the most common problems that can occur while training machine learning models. Overfitting occurs when the model tries to fit every data point precisely in the training dataset. It makes the model vulnerable to noise and outliers. Usually, overfitting is most likely with non-parametric and non-linear models that are more flexible while learning a target function. For example, the Decision tree (DT) is a nonparametric algorithm that is very flexible and subjected to overfitting training data. Overfitting problem can be eliminated by running a tree after it has learned, to remove some of the detail it has picked up. Underfitting occurs when the model is unable to extract a reasonable relationship among training data points. An underfit model will perform poorly on training data and will not be able to generalize to new data. Drop out and regularization techniques can be used to overcome underfitting and overfitting problems [82,83].

In the next sections, we discuss the class-wise (supervised, unsupervised, semi-supervised) ML approaches used for in situ monitoring of the L-PBF process.

#### **10. In Situ Data Processing Using ML**

In this section, post-processing of the in situ monitoring data is presented as per the used ML approach classes. This section mostly focuses on in situ defects related to melt pool signatures and the powder layer spread. The use of the ML approach in other domains such as geometric control, cost estimation, process optimization, etc., is not covered in this review.

#### *10.1. Supervised Learning Approach*

Supervised learning is one of the most commonly used ML approaches for detecting the defect in the L-PBF process. The first step in supervised learning is to prepare a labeled training dataset. Labeling the in situ data is the most challenging and error-prone task, which can affect the overall accuracy of the supervised ML model. Therefore, labeling the in situ data, such as IR images, photodiode signals, pyrometer signals, can be done via making a statistical comparison between in situ data and CT images, which captures the ground truth labels. The same can be used for labeling the di fferent clusters obtained by unsupervised learning. Gobert et al. and Petrich et al. [84,85] used a 36.3-megapixel digital single-lens reflex (DSLR) CCD camera (Nikon D800E) mounted inside the build chamber of EOS M280 (L-PBF system) to detect discontinuity defects such as overheating, pores and unmolten powder. The ground truth bales were extracted by comparing the post-CT scan data with captured in situ images. The labeled dataset was then used to train supervised ML approaches such as support vector machine (SVM) and neural network (NN). Imani et al. [86] used ML techniques like SVM, K-NN, and NN to find the process parameters such as hatch distance, laser power, and scanning speed that likely to produce more porous parts. The link between process parameters and the location, size, and frequency of the pores was studied. Aminzadeh et al. [87] used an 8.8 megapixel USB Digital Camera with high focus lenses to capture the post-printed images of every layer in L-PBF. They used Bayesian inference to detect the porosity and quality of the final part. The training dataset was labeled as "defective" and "non-defective" based on visual inspection. The frequency-domain features were extracted, and a trained Bayesian classifier was used to predict the quality of the part. Zhang et al. [88] used a high-speed camera to capture process images such as melt pool, plume, and spatters. These images were used to train SVM classifiers and CNN to detect the process anomalies. The CNN model showed a higher accuracy of 92.7% compared to the SVM classifier, which showed an accuracy of 90.1%.

Baumgartl et al. [89] used a convolutional neural network (CNN) to detect defects from thermographic images. CNN showed an average accuracy of 96.80%. The thermographic images were captured using an infrared camera, which was placed above the build chamber with an angle of 60o to the build plate. Delamination and spatter defects were successfully identified and cross-validated with the CT scan. However, the model does not identify other defects, such as balling, lack of fusion, and keyhole.

Scime et al. [90] used a one megapixel Photron FASTCAM Mini AX200 high-speed camera in the visible range to capture melt pool images. The machine learning technique, called Scale Invariant Feature Transforms (SIFT), was used to extract features from the captured melt pool signatures. The bag-of-words (or Keypoints) were used to obtain a scale-agnostic description of melt pool morphology. The melt pool morphology represents the shape of the melt pool, spatter, and vapor plume. With the ML approach, the author classified the individual melt pool morphologies into four categories, such as desirable, balling, under-melting, and keyhole porosities. A supervised classification ML technique called support vector machine (SVM) was used to train and test the model. The percentage of the melt pool morphologies linked to process parameters such as laser power and scan speed was obtained.

In another study, Scime et al. [91] used a computer vision algorithm to predict the percentage of types of defects present in the part based on the powder bed images. Scime et al. extracted the regions of the images and classify such as anomaly-free, recoater hopping, recoater streaking, debris, superelevation, part failure, and incomplete spreading. These categories were used as the fingerprints for the computer vision algorithm. A percentage of the particular defect in the respective layer was classified. The approach did not use melt pool signatures and instead used powder bed spreading as their main signature. Repossini et al. [92] used a high-speed camera to capture spatter images and used them as an indicator for process stability. They have developed a supervised learning logistic regression model to classify di fferent energy density conditions to various quality states using spatters as a descriptor.

Some researchers also studied the acoustic signals to detect the defects in the L-PBF process using a supervised learning ML approach. For example, Ye et al. [93] used deep belief networks (DBN) to analyze the acoustic signals to detect defects such as balling, keyhole, and cracking. Shevchik et al. [94] used Bragg grating acoustic sensors to capture the acoustic signals during melting, solidification, and spattering processes. The time-domain captured signal was transformed into the frequency domain using the wavelet transformation technique. The spectral CNN was used to identify the particular defect signals. The presented sCNN model showed an accuracy of 83–89%.

#### *10.2. Unsupervised Learning Approach*

The use of unsupervised machine learning is limited in the field of the L-PBF process. However, unsupervised can be useful when obtaining the labeled training dataset is not possible. Another advantage of using unsupervised learning compared to supervised learning is that it does not require human interaction to the label training dataset. Grasso et al. [10] demonstrated the use of the T-mode Principal Component Analysis (PCA) for image data obtained by an OlympusTM I-speed 3 camera mounted outside the build chamber at an angle of 40◦ w.r.t build plate to define a spatial statistical descriptor and to detect local over-heating phenomena along the scan path by analyzing the intensity profile of each pixel. These local hotspots correspond to the regions where slow cooling occurred due to variation in heat flux; it can be observed at sharp corners and overhang structures. K-means clustering learning was applied to the spatially distributed PCA descriptors to detect a defect in an automated fashion. Although an iterative updating of the k-means clustering based on the new dataset allowed to detect defects in complex shapes, the approach is dependent on the comparisons between data collected at di fferent layers or on the signal from successful builds of the same geometry, which limits its applicability to other situations. Grasso et al. [95] used an o ff-axis IR camera to monitor the plume of zinc material to design a data-driven and automated approach for process monitoring. The region of interest (ROI), which is the plume, was extracted from the IR images. A multivariate control-charting method was proposed to monitor the ROI features, i.e., mean intensity and area of the plume, to detect the deviation from the stable plume patterns. However, in this work, the author did not take into consideration the morphology of the melt pool.

#### *10.3. Semi-Supervised Learning*

Okaro et al. [96] proposed another perspective of the ML approach for data treatment and predicted the quality of the parts based on their mechanical property as a descriptor. Key features were extracted from the photodiode signals, and a semi-supervised classification algorithm called "Gaussian Mixture Model-Expectation Maximization (GMM-EM)" was applied to classify the samples as "faulty" and "acceptable" based on the ultimate tensile strength of the tensile bars. This approach showed the possibility for automatic certification of the L-PBF parts based on their mechanical properties. However, the method of Okaro et al. does not illustrate any possibility to detect the defects in the parts, which is the cause of inferior mechanical properties.

Yuan et al. [97] used a high-speed 1 kHz video camera, with its focal point moving along with the laser spot, to acquire in situ video data to monitor scan track width and continuity. Image processing algorithms were used to label a small amount of data using structured light measurements. The labeled in situ videos were used to train a semi-supervised CNN model based on the temporal ensemble method, with a small amount of labeled data and a large amount of unlabeled data. Yuan et al. showed that the semi-supervised CNN approach performs better than the supervised CNN approach. They also studied the average measuring time for 100 videos, which was 1.4 ms and concluded that the trained CNN approach is adequately fast to monitor in real-time. The summary of the discussed literature is presented in Table 1.


**Table 1.** List of the literature on the processing of L-PBF in situ data using machine learning approaches.

The final aim of in situ monitoring is to improve the overall reliability of the process via a feedback control loop. This will significantly reduce the material waste and time needed to qualify the part for quality assurance. Yao et al. [98] developed a closed-loop feedback control system for real-time monitoring of the L-PBF process. Firstly, the defect level of each printed layer was calculated based on irregular and non-homogeneous patterns using multifractal analysis. Secondly, the composite index was decided for each porosity level based on Hotelling T<sup>2</sup> statistics from the multifractal spectrum. Finally, the stochastic dynamics of the layer to layer defect conditions were modeled as a Markov process to determine an optimal control policy.

#### **11. Commercial Monitoring Systems**

Due to the increasing demand for L-PBF systems in manufacturing industries, the commercial suppliers of L-PBF systems are now equipped with the inbuilt in situ sensing modules. The monitoring system layout, detection range, and types of sensors di fferentiate di fferent commercial supplier monitoring systems. There is no common consensus on the detection range of the melt pool emissions among di fferent suppliers. The reason to choose a specific detection range is protected under the confidentiality clause by the supplier. Therefore, each module has its disadvantages and advantages. For example, the EOSTATE Optical Tomography (OT) module detects the light emitted from the met pool for every layer. The EOSTATE Exposure analysis tool provided by EOS GmbH uses a statistical algorithm to detect so-called "hotspots" and "cold spots", which are the regions of high and low intensity compared to rest of the layer.

Similarly, Melt Pool Monitoring (MPM) provided by SLM solutions captures the melt pool emissions in the near-infrared region by using two photodiodes with di fferent spectral ranges. Then, captured thermal emissions can be studied in the software module supplied by SLM solutions in a layer by layer fashion. The Layer Control System (LCS), also supplied by SLM solutions, uses the visible range camera to monitor the quality of the powder bed spread. After the powder bed spreading, an automatic algorithm processes the captured image in real-time and automatically activates the recoating operation in case of inhomogeneous powder spread. Renishaw supplied the InfiAM module composed of MeltView, which captures the in situ data in a wide spectral range and LaserView to monitor the part build and laser power, respectively. It provides the captured data in 2D and 3D view as the build progresses. Table 2 summarizes the in situ sensing equipment with their sensing process signature capabilities. It shall be noted that the author's aim is not to comment on the capabilities of the available commercial systems as the authors have not personally used all these systems.


**Table 2.** List of commercial in situ sensing modules.

## **12. Conclusions**

From the literature available, it is evident that the in situ sensing of the L-PBF process is still in the infancy stage. Much of the research conducted so far focuses on developing in situ sensing systems to understand the process and less concentrated on detecting defects in real-time. The ultimate goal of in situ sensing in the L-PBF process is to develop a closed feedback control system to improve and control

the quality assurance in real-time. Understanding the process signatures and their link to the potential defect in the final part is the first step towards this. Commercial systems equipped with monitoring systems work on a black box phenomenon, and the critical information obtained during the process is protected by the manufacturer. The large size of the data generated for a particular job is another limitation of these systems. Thus, there is a huge demand for other advanced machine learning models for treating and extracting crucial information from the available captured data. To date, only a few researchers have used a machine learning approach to counter this problem.

The complexity of the L-PBF process is a significant challenge to understand all the phenomena occurring due to controllable and uncontrollable parameters. Therefore, monitoring each stage of the process, such as powder health, powder spreadability, build chamber inert gas flow, oxygen content, melt pool stability, etc., are closely interlinked. Therefore, linking the di fferent process states such as powder bed spreading, layer heatmap, and melt pool stability can enhance the quality control in real-time. Real-time monitoring will also limit the use of other quality checks, such as non-destructive testing, which costs a lot of time and money. The improved reliability and repeatability of the L-PBF process will allow risk-averse industries to adopt the process entirely.

**Author Contributions:** Conceptualization and methodology, P.Y., E.L., O.R., C.A., E.L.G.; investigation, E.L., O.R., C.A.; writing—original draft preparation, P.Y.; writing—review and editing, P.Y., E.L., O.R., C.R.; visualization, P.Y.; supervision, E.L., O.R., C.R., E.L.G.; All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was conducted as part of the "ENABLE" project funded by the European Union's Marie Sklodowska-Curie Actions (MSCA) Innovative Training Networks (ITN) H2020-MSCA-ITN-2017 under the gran<sup>t</sup> agreemen<sup>t</sup> Number 764979.

**Acknowledgments:** The authors would like to thank the T-ADD team at SIRRIS for providing valuable suggestions and seminars related to the L-PBF process.

**Conflicts of Interest:** The authors declare no conflict of interest.
