Off-axis high-speed camera-based real-time monitoring and simulation study for laser powder bed fusion of 316L stainless steel

In order to develop smart laser powder bed fusion (LPBF) devices that autonomously identify a defect and remove it during the process for first-time-right and zero-defect parts, it is important to develop reliable on-machine defect measuring capabilities. As defects in LPBF parts often occur below the layer that is being processed, capturing the information of a printing layer may not give information about physical phenomena that are occurring below this layer. Therefore, to predict volumetric features such as porosity only by looking at the layer being processed, the correlation between process signatures identified in-process and defects measured via post-process inspection methods (for example X-ray computed tomography) needs to be conducted. Hence, in situ monitoring and post-process metrology form a basis to better understand the fundamental physics involved in an LPBF process and ultimately to determine its stability. By utilizing high-speed imaging, various process signatures are produced during single-track formation of 316L stainless steel with various combinations of laser power and scan speed. In this study, we evaluate whether these signatures can be used to detect the onset of potential defects. To identify process signatures, image segmentation and feature detection are applied to the monitoring data along the line scans. The process signatures determined in the current study are mainly related to the features like the process zone length-to-width ratio, process zone area, process zone mean intensity, spatter speed and number of spatters. It is shown that the scan speed has a significant impact on the process stability and spatter formation during single-track fusion. Simulations with similar processing conditions were also performed to predict melt pool geometric features. Post-process characterization techniques such as X-ray computed tomography and 2.5-D surface topography measurement were carried out for a quality check of the line track. An attempt was made to correlate physics-based features with process-related defects and a correlation between the number of keyhole porosities, and the number of spatters was observed for the line tracks.


Introduction
In laser powder bed fusion (LPBF), melt pool flow dynamics are driven by surface tension gradients resulting from strong temperature gradients beneath the laser in the melt pool region [1]. This complex hydrodynamic flow sometimes leads to melt pool instability and generates features that can be observed by sensors in real time. Melt pool, vapor plume, and spatters are the key features observed in the LPBF process when using a camera-based monitoring system [2,3]. These features can provide significant information for understanding the LPBF process. The laser-metal powder interaction leads to formation of a pool of molten metal known as the melt pool. It is well known that the build part properties are significantly influenced by the melt pool geometry and dynamics [4]. Real-time monitoring of the melt pool in LPBF has enabled the elucidation of these process phenomena to detect any unpredicted faults and control the process parameters [5,6]. The vapor plume is another feature that sometimes can be observed in the monitoring data. The build chamber also encloses metal vapor plumes and a vapor jet ejected from the melt pool due to laser melting. Though the amount of metal gas is relatively small, still the concentration above the melt pool can be significant and its concentration varies depending on processing parameters. When the metal plume and vapor jet absorb the laser input, the metal gas could go into an ionization stage and result in a plasma [7]. The laser absorption into this plasma further reduces the laser power arriving at the powder and affects the laser absorption for the powder in the powder bed. A shielding gas such as argon is used to move the metal plume and vapor jet away from the melt pool, but the extraction can never be 100%.
Spatter refers to the molten metal droplets ejected from the melt pool during the melting of metal powder [8] (Fig. 1). The presence of spatters during the process may be a signature of an unstable process, and the number of spatters may be an indicator of the occurrence of certain types of defects [9][10][11]; therefore, spatter features are extracted from the monitoring data. Different spattering mechanisms lead to different types of spatter morphology. As mentioned by Young et al. [12], two main mechanisms are responsible for spatter formation: recoil-pressure driven spatter and entrainment-driven spatter. When the temperature of the melt pool exceeds the boiling temperature, a strong metal evaporation happens. Strong metal evaporation causes a high recoil pressure that exerts downward pressure on the melt pool, resulting in a rapid fluid convection there and sometimes causing the ejection of molten metal from the melt pool. On the other hand, entrainment of powder particles occurs by a local pressure difference of ambient gas flow due to the Bernoulli effect [13,14]. Spatter usually occurs when high laser power, high layer thickness, or low laser scanning speed are used. Spatter particles sometimes undergo an oxidation reaction during in-flight cooling, which may therefore form an inclusion in the build parts [15,16]. Spatter may also land on a build part and cool down faster than the build part, thus forming a different microstructure than the build part. These droplets increase the layer thickness and surface roughness of the top layer [17,18] while also enhancing process-induced porosity [19,20]. Collisions between spatter particles can further affect the particle dynamics [21]. The spatter direction and spatter velocity also provide useful information and are considered as key physics-based image features in defect detection. According to Ji and Han [22], the spatter direction may be a useful feature for identifying pore formation. Processing parameters, material density, and the direction and amplitude of inert gas flow across the powder bed all affect the spatter ejection trajectory. The spatter velocity is inversely proportional to the material density, e.g., dense materials such as steel have spatters moving at lower velocity as compared to spatters of relatively less dense materials such as titanium.
The characterization of process-induced spatter has been the focus of many studies. Bayle and Doubenskaia measured the spatter ejection for stainless steel 904L by an infra-red camera and pyrometer [24]. These spatters were observed to be ejected in the laser scanning direction (i.e., in the direction of higher temperature gradient) and their velocity was in the range of 0.5-5 m/s. Qiu et al. [25] studied the effect of scanning speed and layer thickness on the melt flow behavior. High-speed imaging was conducted at 10,000 frames per second using a high-speed camera. The velocity of spatter particles was found to increase with increasing scanning speed. Lane et al. [26] designed an off-axial setup with a 45° viewport and used a camera for measuring lower temperature phenomena. Although a measurement of smaller particles was not possible due to the optical resolution limit, some spatter particles of 200 μm diameter could be observed being ejected from the melt pool. Matthews et al. [27] observed the depletion of powder particles (i.e., denudation) around the solidified track using ultra-high-speed imaging at 500 k frames per second with an optical resolution of about 5 µm. Ly et al. [28] carried out high-speed imaging of melt pool dynamics similar to Matthews et al. [27] for 316L stainless steel tracks produced on top of a metal substrate and compared the experimental results with simulations. The authors concluded that 15% of the spatters are recoil pressure-induced droplets with velocities of 3-8 m/s. Gunenthiram et al. [29] performed single-track experiments similar to Matthews et al. [27], but with a powder bed that is moved with a high-speed (x, y) table below the laser. Repossini et al. [30] recorded spatter formations for different melting conditions using an off-axial high-speed camera at 1000 frames per second. Repossini et al. [30] observed a higher number of spatters in the over-melting conditions as compared to the normal melting conditions. Zhao et al. [31] used high-speed X-ray imaging and diffraction to image the melt pool dynamics of Ti-6Al-4 V at 50 k frames per second. The keyhole porosity formation dynamic could be seen with high spatial and temporal resolutions.
However, while many studies are already done regarding real-time LPBF observation, only a few studies are performed that correlate the resulting signature features to real defect formation [3]. This paper discusses a feature-based real-time optical monitoring system and the effect of processing parameters on both the optical features and defect formation. For that purpose, an off-axis high-speed camerabased monitoring setup was used to measure melt pool, vapor plume, and spatter features. Spatter features along the line track were automatically identified from post-process 2.5-D surface topography measurements. The melt pool geometry was predicted from a finite element model and later compared with the experimental results. Also, a correlation analysis was performed between the optical features extracted from the monitoring data and the quality features observed in the X-CT reconstruction of the line tracks. The structure of the paper is as follows. First, the experimental details are given in Section 2. Next, an overview of the experimental results is discussed in Section 3, followed by the conclusions in Section 4.

LPBF process
To provide insights into defect and feature formation, an experiment was performed by printing single tracks of 316L stainless steel (316L) on top of an LPBF 316L substrate (Fig. 2). The substrate of 9 mm × 9 mm × 7.5 mm was printed on a 3D Systems ProX DMP320 machine with nominal processing parameters (laser power: 215 W, laser speed: 900 mm/s). The LPBF machine's laser source was an Ytterbium fiber with wavelength of 1064 nm, and the laser spot size was 75 µm. The process was carried out using argon as a shielding gas. The powder particle size distribution of the 316L stainless steel powder from 3D Systems ranges between 20 and 50 µm. In total, 14 line scans were printed on top of the substrate with seven different process parameter sets (Table 1). A layer thickness of 30 µm and a hatch spacing of 100 µm were used for both the substrate and the line tracks.
Seven variations in the energy input were accomplished by either changing the laser power or the scanning speed   (Fig. 3). These are summarized in Table 1. The linear energy density (LED) was calculated for all settings by dividing the laser power (P) by the scanning speed (v). Line scans were printed twice in opposite directions for every single combination of processing parameters. This step was done to check the reproducibility and to investigate the effect of the gas flow direction. The length of the line track was 8.8 mm, and the distance between the line tracks was 0.5 mm. Prior to printing the line scans, the substrate layer below the line scans was remelted without recoating to reduce the surface roughness and to remove pores (if any) from the layers below. As a result, all the defects that form underneath the line tracks can be considered as originating from the line tracks themselves. The test object was cut by wire-electrical discharge machining (wire-EDM) from the baseplate after processing.

In-process monitoring and image processing
Monitoring data were recorded by using an off-axis highspeed camera (Mikrotron Eosens 3CL) placed with an inclination angle of about 25° with respect to the normal of the build plate. The advantage of using an off-axis highspeed camera-based system is being able to monitor the spatter signature and not requiring any modifications in the machine or the optical path of the laser beam. The video stream is acquired at 20,000 frames per second, with each frame having 120 × 120 pixels. The camera had a 30-µs exposure time and a 50-mm lens with an aperture of f/16. The aperture and shutter time were set manually so that spatters are clearly visible in the video stream. The resulting pixel size corresponds to approximately 100 µm × 100 µm over a field of view of 12 mm × 12 mm.
A short-wavelength pass filter with a cut-off value of 975 nm was added before the camera to protect it from the laser source. The monitoring system captured radiations in the visible to near infrared wavelength range between 350 and 975 nm. The LPBF machine was equipped with the Materialise Control Platform (MCP), and controller data was also collected at 100 kHz. This controller data consisted of the (x,y)-position of the laser on the build area, the laser on/off signal, the laser power, the laser speed, and timestamps. The high-speed camera was connected via Camera Link to NI PXI. Data was stored to an NI SSD drive with a high-speed data transfer rate of around 1.5 GB/s (Fig. 4).
Physics-based image features were identified from the monitoring data after binarization of the image frames. A static threshold of 40 intensity units is applied to the 8-bit video frames (whose values range between 0 and 255). The motivation for using this static threshold is because the resulting image has a low level of noise and the appearance of the environment in the field of view remains consistent over time. The melt pool, vapor plume, and spatter regions have a higher intensity than the background, and therefore, no sophisticated method is used for segmentation. After binarization, the image is segmented automatically per frame into background, a combined melt pool and vapor plume which is labeled "process zone" in this work, and spatter regions by using a connected component analysis algorithm (Fig. 5). From the segmented images, several established optical features were extracted: process zone length-to-width ratio, process zone area, process zone mean intensity, and number of spatters [2]. Additionally, spatter speed and spatter direction were also computed. A detailed discussion of individual features is presented in Section 3. To identify the process zone and spatters on each frame, connected component labeling is used [32] to uniquely label each high-intensity region within the field of view on the off-axis camera. The process zone is taken to be the largest of these regions while the remaining regions are labeled as spatter. The trajectory for each spatter is then calculated by matching its position in neighboring frames using the Euclidean distance: a spatter in one frame is matched to the closest spatter in the subsequent frame. Once a spatter cools down or disappears from the field of view of the camera, the tracking of the spatter is completed, and the spatter's velocity and direction are computed. The velocity of a spatter is derived from its total distance traveled (in meters) divided by the number of frames in which the spatter appears. To measure the direction of a spatter, a linear vector, v Sp , is fit to the spatter's full trajectory, and the angle between the spatter and the laser scanning direction is computed as follows: where v MP (y), v MP (x) represents the coordinates of the melt pool trajectory vector during the frames where the spatter is within the field of view of the camera, and v Sp (y), v Sp (x) represents the coordinates of the trajectory vector of the spatter. The area of the spatter is calculated as the maximum pixel count from all the frames where the spatter appears, capturing the moment where the spatter appears the closest to the camera with the highest luminous intensity. The spatter number is calculated per frame as the total number of the labeled spatters obtained from the connected component labeling. Also, the mean number of spatters per line track is calculated. All other process zone and spatter features were calculated as described in [2].

X-CT
The test object was scanned by X-ray computed tomography (X-CT) using a Nikon XT H 225 ST machine at KU Leuven that has a maximum tube potential of 225 kV. The instrumental parameters were set at a tube voltage of 220 kV, 45 µA, and an exposure time of 4000 ms. A total of 3600 radiographs were taken. The chosen target material for the X-ray generation was tungsten. The source-to-detector distance was 1048 mm and the source-to-object distance was 52 mm, leading to a measurement resolution (voxel size) of 10 µm × 10 µm × 10 µm. A tin filter of 0.5 mm thickness was used in front of the X-ray tube window to reduce the beam hardening effect. The X-CT volume reconstruction was performed using the Feldkamp-Davis-Kress (FDK) algorithm [33].
From the X-ray computed tomography images along the line tracks described in Section 2.1, several surface and subsurface process-related defects were found. Based on these results, process parameters were either labeled as belonging to a "defect" or "no-defect" region. Line tracks with laser parameters P = 215 W and v = 900 mm/s (ID#1) were marked as having the optimum processing parameters as no defects were observed along the track ( Fig. 6 (a)). A series of keyhole porosities were formed below the two line tracks with laser parameters P = 215 W and v = 225 mm/s (ID#4), and these were therefore marked as keyhole porosity regions. As the resolution of X-CT reconstruction resulted in a 10-µm voxel size, it was possible to measure keyhole pore sizes from 10 µm and above. With laser parameters P = 54 W and v = 225 mm/s (ID#7), incomplete melting of the two line tracks was observed and, therefore, these tracks were labeled as tracks having lack-of-fusion defects. Discontinuous line tracks were observed for the two tracks with laser parameters P = 500 W and v = 2093 mm/s (ID#6) due to Plateau-Rayleigh instability. While the energy inputs for both the optimum and the Plateau-Rayleigh instability conditions were the same, defects form at higher power and speed driven by surface tension in the melt pool region resulting in discontinuous melt tracks. It can be seen in Fig. 6 that lackof-fusion and Plateau-Rayleigh instability result in surface defects, whereas a keyhole porosity is a sub-surface defect that occurs several layers below the line track.

2.5-D surface topography measurements
After performing the X-CT scan, the surface topography of the top surface of the test object was measured by a 2.5-D confocal optical microscope (S neox -Sensofar) in order to identify spatter on the line tracks. A stitched area of 8.33 mm × 7.26 mm was measured with a pixel size of 0.65 μm using an objective giving 20 × magnification and having a numerical aperture (NA) of 0.45. After measurement, the topography was pre-processed by removing spikelike artifacts and subtracting the least-squares plane. Spatter particles were identified along the line track using crosscorrelation with a typical 70 × 70 pixel spatter feature kernel.

Finite element modeling of the melt pool
A finite element model was developed to simulate the production of a single track of 316L stainless steel using Addi-tiveLab software [34]. This commercially available software is particularly tailored to the modeling, simulation, and visualization of metal additive manufacturing processes. In this software, the moving laser beam was modeled as a rotary Gaussian heat source. The mesh size varies from 8 to 100 µm. Conductive heat losses were considered to the side, the bottom, and the top of the simulation domain. In addition, the radiative heat losses were considered at the top of the simulation domain. The influence of process parameters on the melt pool dimensions was evaluated. To do so, melt pool dimensions for different process parameters were predicted and validated experimentally. The calculation time for this problem is approximately 1-10 min on an 8-core machine. Therefore, it can provide a rough estimation of the melt pool dimensions in a fast way, avoiding the need of longer calculation times. The used thermophysical properties of 316L stainless steel are given in Table 2.

Metallography and optical microscopy
Once the X-CT and 2.5-D surface topography measurements are performed, the test object is prepared for metallography. The object was resin mounted, grinded, polished, and etched to measure the melt pool width and depth. These width and depth measurements can be used among others to validate the simulated results of the line tracks. Electrolytic etching was performed with 10% oxalic acid to etch the 316L sample cross-section.

Results and discussion
Physics-based image features related to process-related defects are studied and discussed in this section. As the resolution of the off-axis camera does not allow to individually detect the melt pool and the vapor plume, both are combined in an area that is labeled the "process zone." The first part of this section discusses the process zone features extracted from off-axis camera monitoring for the line track experiment, specifically process zone area, process zone length-to-width ratio, and process zone mean intensity. The second section presents the spatter features segmented from raw images, specifically the number of spatters, spatter speed, and spatter direction. The third part shows the results of automatic spatter identification from 2.5-D surface topography measurements. The fourth subsection presents melt pool geometry features such as depth and width extracted from both finite element modeling and optical microscopy of the line tracks cross-section. In the final part, a correlation between signature features extracted from real-time monitoring data and X-ray computed tomography for the line track is discussed.

Melt pool and vapor plume feature analysis from monitoring data
Several process zone (combined melt pool and vapor plume) features are extracted, measured in the XY plane, and discussed in the current section for the line tracks. The process zone dimensions such as area and length-to-width ratio are of major importance in LPBF since these dimensions have a large influence on the process stability. The process zone area and length-to-width ratio depend on the processing parameters being used (laser power, scanning velocity) and on the thermal diffusivity of the material being processed.
The melt pool will be more elongated (i.e., a larger lengthto-width ratio) for the material with smaller thermal diffusivity, whereas it will lead to a more equiaxed melt pool (i.e., a length-to-width ratio closer to 1) for the material with a high thermal diffusivity. On the other hand, the influence of the processing parameters on the process zone area and length-to-width ratio is rather complex. Laser power and scanning speed are not interchangeable, i.e., the same energy density with different laser power and scanning speed will lead to different results. To understand the effect of combinations of laser power and scanning speed on process zone area and length-to-width ratio, line tracks with varying laser power and scanning speed were produced as discussed above (Table 1). Figure 7 shows the mean value of process zone area along the different line tracks on the process window that covers the seven combinations of laser power and scanning speed conditions that were studied. It was observed that the process zone area was similar for optimum as well as keyhole porosity conditions, whereas it varied for Plateau-Rayleigh instability and lack-of-fusion conditions. A significant influence of laser power on the process zone area was observed. It is known that the laser power influences the temperature reached in the molten part and the powder bed since it determines the total amount of energy added per second by the laser source to the powder bed. At higher laser powers, process zone temperatures will be larger as compared to process zone temperatures when using low laser power; therefore, a higher process zone area was observed at higher laser power as compared to the process zone area at lower laser power. It was observed from the monitoring data that conditions that form Plateau-Rayleigh instability result in the highest process zone area, whereas the lowest process zone area was observed for the lack-of-fusion conditions. In contrast to the process zone area, quite different results for the process zone length-to-width ratio are obtained for a combination of several laser power and scanning speed conditions. Here, a significant influence of the scanning speed on the process zone length-to-width ratio was observed. The length-to-width ratio increases with increasing scanning velocity, specifically for the Plateau-Rayleigh instability conditions. The possible reason for the resulting higher length-to-width ratio at the higher scanning speed condition is that the amount of radiative heat loss is lower since there is simply not enough time to radiate a significant amount of energy. It is also possible that a higher laser speed will heat up a larger area over the 30 µs that the camera is recording the image. At low scanning speeds, a large part of the added energy is lost through heat radiation, which results in smaller melt zone dimensions. In this study, a length-to-width ratio of 2.5 was observed for Plateau-Rayleigh instability condition, whereas a length-to-width ratio of 1.5 was measured for optimum, keyhole porosity and lack-of-fusion conditions (Fig. 8).
In general, any object radiates heat. By measuring this radiation, one can estimate the temperature of an object. Fig. 7 Interpolated contour plot indicating the mean value of the process zone area along the line tracks. The interpolation was carried out using a distance method. The range of the main process zone area and the corresponding color are shown at the right side of the contour plot The melt pool emits thermal radiation at the wavelengths that correspond to the melting temperature (Planck's law). One way to estimate the temperature of the melt pool is by capturing the intensity of the process zone [6,36]. The grey value of a pixel in an image from a CMOS camera is a function of the emitted light intensity. Here, a linear relation is taken between the process zone intensity and the grey value (on an 8-bit grayscale). The mean value of intensity within the process zone is calculated by taking an average of process zone mean intensity for all the frames along a line track of 8.8 mm. Figure 9 presents the average value of the process zone mean intensity for the line tracks in the process window.
For the optimum process parameters, heat transfer happens mainly via heat conduction; however, heat transfer via heat radiation also takes place. For keyhole mode melting, convective heat transfer (also known as thermo-capillary convection) is the most common form of heat transport. As mentioned earlier, for the higher scanning speed such as the Plateau-Rayleigh instability condition, the heat transport via heat radiation is lower. As can be seen from Fig. 9, the process zone mean intensity value for optimum condition is in the range of 91-96, and for Plateau-Rayleigh instability and keyhole porosity region, the mean intensity of the process zone is in the range of 86-91. Clearly, the grey value decreases with decreasing laser power such as for lack-of-fusion defects, since with lower laser power, the temperature of the process zone is quite low. The influence of laser power and scanning speed on the process zone mean intensity seems complex and further research is required to characterize it.

Spatter feature analysis from monitoring data
In this section, the effect of laser power and scanning speed on spatter features is discussed, specifically the number of spatters, spatter speed, and spatter direction. It is known that molten metal is ejected from the melt pool once the dynamic pressure, due to the recoil pressure and the Marangoni force, is higher than the pressure produced by the surface tension [28,37]. In this work, spatter features extracted from raw monitoring data are of high intensity (similar to the process zone) and are therefore called "hot spatter." Moreover, powder spatters are ejected near the process zone, and due to their low intensities, these "cold spatter" cannot be recorded in our real-time monitoring data. The spatters move in three dimensions, but in this work, we can only observe spatter particle motion and trajectory in the projection on the XY plane parallel to the sensor of the camera. Some studies showed the feasibility of spatter tracking using a high-speed stereo vision and 3D particle tracking velocimetry [38,39].
From Fig. 10, it can be seen that scanning speed has a significant influence on the spatter generation, and it will later be correlated with quality of the part in Section 3.5. For the keyhole porosity condition, the mean value of the number of spatters was approximately 1 for all the frames along a line track of 8.8 mm, whereas for all the other conditions, almost no hot spatter was observed along the line track. The spatter that forms during the keyhole mode melting is because of the recoil pressure and the Marangoni force as mentioned earlier.
In addition to the number of spatters, spatter speed and orientation might also provide useful information of the process stability. On the macroscopic scale, the spatter ejection trajectory is possibly dominated by the vapor plume and the Fig. 9 Interpolated contour plot showing the mean value of the process zone mean intensity along the line tracks. The range of the mean process zone intensity and the corresponding color are shown at the right side of the contour plot gas flow [13,28,40,41]. Most of the droplet spatters are of 100-400 µm size as calculated from the monitoring data. This is much larger than the size of the powder particles. The speed of spatters was calculated by considering the spatter distance between two consecutive frames. The calculated spatter speed ranged from ∼ 0.4 to ∼ 7.8 m/s in the viewing plane. The average spatter speed along the line track was lowest with 0.2 m/s for the lack-of-fusion condition and highest with value of 3.3 m/s for the Plateau-Rayleigh instability condition. For the optimum and keyhole porosity conditions, the spatter speed is almost similar with mean value of 1.4 m/s and 1.6 m/s per 8.8-mm line tracks, respectively. Figure 11 presents the polar histogram of the spatter orientation for different laser power and scanning speed conditions. Here, 0° represents the melt pool and laser scan direction, and 90° represents the gas flow direction. The radius in the polar histogram is the normalized amount of spatter (in pixels) along each frame of 120 × 120 pixels. For the optimum condition, spatters are oriented to both gas flow direction and backward to melt pool scan direction, whereas for both Plateau-Rayleigh instability and lack-of-fusion conditions, spatters were observed moving mainly in the opposite direction of the moving melt pool. A uniform distribution in all directions was seen for the keyhole condition, similar to what is found in the literature [28]. In [28], spatter seems to rise almost vertically in the case of keyhole mode melting due to the use of a side-view camera, whereas in this work, spatter seems to be uniformly distributed in all directions in the case of keyhole mode melting due to the use of a top-view (but still off-axis) camera. Due to the vapor plume emission being constrained in the vertical direction by the keyhole walls, Ly et al. [28] observed the spatters rising almost vertically in the keyhole mode melting condition.
Conversely, for the low energy input, the vapor plume emits mainly in the backward direction, which causes droplet spatters to orient in the similar direction. Meanwhile, Zhao et al. [42] discovered a new mechanism for droplet spatter in keyhole conditions based on the X-ray synchrotron imaging technique. A tongue-like protrusion on the front keyhole wall undergoes a bulk explosion, which causes spattering to occur. Consequently, the momentum released by the explosion of the protrusion determines the spatter speed, and spatters are seen being ejected along both the front and the rear keyhole rims. This is in agreement with the results obtained in the current study about spatter orientation in keyhole mode melting.

Spatter identification from 2.5-D surface topography measurements
Confocal microscopy was performed for post-process spatter analysis where several spatter features were extracted from monitoring data. Several studies have already been performed to identify the spatter features from surface topography measurements [43,44]. In this work, spatters were automatically identified along the line track via crosscorrelation, where the kernel is a 70 × 70 pixels spatter feature ( Fig. 12 (a)). The cross-correlation map shows spatters being identified along the line track ( Fig. 12 (c)). A spatter map was extracted after applying a thresholding to the crosscorrelation map (Fig. 12 (d)). The threshold value selected is 0.015. To confirm the number of spatters observed from monitoring data, a spatter map is obtained from the surface topography measurements. A comparison of spatter maps for keyhole mode melting and optimum conditions is made (Fig. 12 (d) and (e) respectively), which confirms that more spatter is identified along the line track for keyhole condition than for the optimum condition. It was assumed in this work that spatters segmented along the line track in keyhole mode ( Fig. 12 (d)) are due to melt pool instabilities. Another possible mechanism of spatter formation is powder particles sucked into the wake due to the metal vapor jet. However, spatter formation in this way results in cold spatter particles, which could not be recorded in the current off-axial setup. As shown in Fig. 10, a high number of spatter particles were recorded during the keyhole mode. For this reason, it was assumed that spatters observed along a line track are hot spatter and not cold powder particles. To confirm this, high dynamic range monitoring data for keyhole mode may be needed to detect cold spatter as well.

Finite element modeling of melt pool
The relationship between the melt pool dimensions and the laser parameters (scanning velocity and laser powder) were predicted using a finite element model and validated experimentally. Conservation of mass, momentum, and energy are considered in order to model the melt pool and heat-affected zone. The model solves the heat equation considering conduction. 3D finite element modeling of LPBF for conduction and keyhole modes has already been carried out as reported in the literature [45][46][47]. In this work, the FEM simulation results seem to be in good agreement with the experimental data for 316L stainless steel considering the experimental reproducibility and the numerical approximations (Table 3). Figure 13 indicates the dimensions of the melt pool obtained experimentally and the temperature distribution obtained from simulations. With P = 215 W and v = 600 mm/s, the melt pool cross-section has a semi-circular shape, similar to the melt pool cross-section in conduction mode. By reducing the scanning speed further to v = 225 mm/s, the melt pool depth increases due to high recoil pressure and enters into keyhole mode. As can be seen from Fig. 13, in a keyhole mode, the keyhole upper part is wide, whereas the bottom region has a V-shape and is narrow.

Correlation between real-time monitoring and X-ray computed tomography
In order to validate the signatures of process-related defects, it is crucial to correlate the part quality with the real-time signatures from the monitoring. Both keyhole porosity and droplet spatter formation are stochastic in nature. However, correlating porosity and spatter was possible using the signal showing spatter orientation with respect to the process zone for a P = 215 W and v = 900 mm/s (optimum); b P = 215 W and v = 225 mm/s (keyhole porosity); c P = 500 W and v = 2093 mm/s (Plateau-Rayleigh instability); d P = 54 W and v = 225 mm/s (lack-offusion). The radius in the polar histogram is the normalized amount of spatter (in pixels) along each frame of 120 × 120 pixels. In the histogram, 0° represents the melt pool and laser scan direction, and 90° represents the gas flow direction   1  215  900  137  137  149  105  2  215  600  152  161  190  167  3  215  450  154  166  265  218  4  215  225  204  215  443  336  5  500  523  178  131  556  605  6  500  2093  149  114  204  240  7 54 900 0 0 0 0 recorded from real-time monitoring and X-ray computed tomography. By correlating the 2D monitoring information of the process zone and spatter with 3D information of the surface and the subsurface defect formation from X-CT, it could be possible to predict for future experiments the susceptibility for defect formation just from the monitoring data. The correlation analysis between in situ process signatures and post-process part characterization has already been studied by several authors [2,3,6,[48][49][50]. Among all the physics-based features discussed in previous sections, a strong correlation was observed between the number of keyhole porosities below the line track and the number of spatters recorded from the off-axis monitoring system (Figs. 14 and 15). Correlating monitoring signature features to defects seen in the X-CT is difficult for fully printed parts; therefore, single tracks were studied in this work. The spatter features' formation depends mainly on material properties and the processing parameters. As seen in Section 3.2, both laser power and scanning speed have an impact on the amount of spatter formation. An increase in energy input results in keyhole formation and in more spatter event formation (Fig. 15). A large number of spatter formations indicate a high recoil pressure in the melt pool. Compared to the keyhole condition, less spatter formation is observed for conduction mode melting (Fig. 14), and therefore, spatter formation can be related to the quality of LPBF part.

Conclusions
In this work, several physics-based features were extracted from off-axis high-speed camera-based monitoring data, and process-related defects in LPBF were studied using X-ray computed tomography.
• The process zone (melt pool and vapor plume) and spatter features were extracted from images and evaluated for line tracks with various laser power and scan speed conditions. A significant influence of laser power on the process zone area, and of scanning speed on the process zone length-to-width ratio, was observed. The highest process zone area and length-to-width ratio were observed for tracks having Plateau-Rayleigh instability, a defect that forms under high laser power and scan speed conditions. The influence of laser power and scanning speed on the process zone mean intensity is not clear, and further research needs to be done. • In addition, several spatter-based features were calculated after image segmentation. The droplet spatter speed ranged from ∼ 0.4 to ∼ 7.8 m/s for 316L stainless steel. A more uniform spatter orientation was observed for keyhole mode melting conditions. • To confirm the spatter variations seen in the monitoring data, post-process surface topography measurements were performed along the line track, and spatters were segmented by cross-correlation template matching. The spatter number segmented from surface topography measurements agrees qualitatively with the spatters segmented from the monitoring data. • Additionally, we experimentally validated a finite element model that predicted the melt pool dimensions for several laser parameters. • An attempt was made to correlate physics-based features with process-related defects. A strong correlation was observed between the number of keyhole porosities from X-CT measurements and the number of spatters from real-time monitoring data.
Once these process signatures are confirmed for more materials and process conditions, it is possible to apply realtime closed loop feedback control to control defects based on the features extracted from real-time monitoring. In the future, physics-based features and quality measures will be studied for different part geometries, materials and machine sets.