A Convolutional Neural Network (CNN) classification to identify the presence of pores in powder bed fusion images

Additive manufacturing (AM) is a group of computer-controlled processes where three-dimensional objects are manufactured by depositing material layer by layer [1]. AM, also known as 3D printing, is the backbone of Industry 4.0. Automotive, defence, aerospace, healthcare, and general manufacturing are some of the prominent areas where additive manufacturing is replacing conventional manufacturing [1]. The main strengths of AM are the reduction in manufacturing costs and time, improved rapid prototyping, geometrical independence, rapid repair, and an ability to produce complex geometries using more sophisticated designs. Moreover, AM helps in reducing the weight of the object [2]. Lightweight production of metallic objects is significant, particularly in the aerospace industry, as the reduced weight contributes to reduced oil consumption and carbon dioxide (CO\(_{2}\)) emissions [3]. According to an estimate by Gebler et al. [4], by 2025, AM will reduce global manufacturing costs by 170-593 billion US dollars, with a 2.54-9.30 exajoules reduction in energy and 130.5-525.5 million tonnes (Mt) reduction of CO\(_{2}\) emissions.

Laser powder bed fusion (LPBF) is a manufacturing technique where complicated geometrical objects are produced by melting pre-defined regions, producing a solidification of the metal powder, layer after layer [5]. LPBF is the most recommended method for metal construction of objects [6]. Typical defects in LPBF include incomplete fusion of powder particles, porosity, powder contamination [7], cracks, surface deformation, irregularities in powder re-coating, and balling [8]. Among these defects, porosity is the most frequent and challenging to detect. Porosity compromises mechanical properties such as fatigue life [9]. It is particularly challenging due to its small size as it is difficult to observe with the naked eye. Several processing parameters, such as laser power, powder morphology, layer thickness, scan strategy, scan speed, gas flow, and hatch spacing, can either directly or indirectly contribute towards the creation of porosity [8]. Incomplete fusion holes, voids, and keyholes are the main porosity types. Incomplete fusion occurs because of the partial melting of the powder layer due to insufficient laser power [10]. The partial melting of the powder layer fails to merge with the layer below, causing porosity [11]. Keyhole pores are formed by a high-energy input that vaporises the powder and leaves gas bubbles in the solidified metal [12]. Voids, on the other hand, are caused by rapid cooling, which increases the residual stress of the melt pool [13].

The size and shape of the pores vary between different porosity types. Keyhole pores are round/spherical in shape and much smaller in size, typically bigger than 50µm [14‐17]. The gaps created due to a lack of fusion are irregular, narrow, and elongated in shape and usually more significant than 200µm in size [14, 16, 17]. The largest pore size observed by Zhang et al. [18] was 340 µm. Here, the pores were irregular in shape and could have resulted from a cluster of smaller pores. There were numerous factors that directly and indirectly influenced porosity. Increasing the energy density caused tiny, round pores (50-110µm), whereas decreasing it below the optimum value resulted in pores as large as 250µm [19]. Du Plessis [20] stated that at higher laser power, the keyhole pores increased in number and size (30µm up to 400µm), whereas keyhole pore formation decreased at higher scan speeds. Choo et al. [21] studied the effect of varying laser power from 50-150%. The maximum reported pore size was 340µm. Leung et al. [22] categorised porosity into two types: gas pores and pores near the oxide layers. Gas pores observed in the experiments were in the range of 250µm, whereas pores near the oxide layer merged with gas pores and grew as large as 50 to 500µm. Mireles et al. [23] had presented an image-based closed-loop control of the electron beam melting (EBM) process. The artificial spherical pores of 600 µm to 900µm were created in test cylinders and successfully detected from powder bed images. Mireles et al. [24] had designed spherical, triangular, cylindrical, and cubic shaped pores of 100µm to 2000µm in their test 3D specimens. The experiments revealed that re-melting the porous layer reduced porosity.

Real-time identification of porosity is complex and challenging. Porosity can be detected using post-build evaluation techniques. However, post-build analysis can be expensive, time-consuming, and laborious [25]. Several destructive (microscopic cross-sectional analysis) and non-destructive (Archimedes density measurement, gas pycnometry and XCT) methods are in practice to detect porosity [26]. Destructive methods like microscopic cross-sectional analysis slice the 3D object to identify defects. However, this results in extra cost, time, effort, and wasted material.

The broader acceptance of AM technologies, especially in aerospace and the medical domain, is hindered by the lack of in situ defect detection, as quality assurance is essential in these fields [27]. The main challenges to in situ monitoring are limited view of the build chamber, the poor spatial resolution of cameras, high temporal load, and the enormous amount of data collected [28]. Machine learning (ML) models are data driven and known for their efficient and effective handling of large data sets. However, applications of ML in LPBF are relatively new, and several issues are limiting the performance of ML solutions. The absence of publicly available data sets, the high cost of data capture, the installation of sensors, and data labelling are considerable challenges to solve for ML applications to be practical in in-process monitoring of the LPBF process [29]. Moreover, the scarcity of data, lack of experience in labelling data, lack of expertise in selecting good features, and the issue of over-fitting and under-fitting of the derived ML models are also hindering the applicability of ML solutions in LPBF [30].

This research aims to develop an ML model capable of identifying seeded defects in layer images from an LPBF build. Our study aims to resolve the issues hindering ML applications in AM, such as data capturing, data labelling, extracting valuable features from the data, class imbalance, and overfitting and underfitting of the ML models. A further aim is to demonstrate that effective hyper-parameter tuning could assist in producing an efficient model from scratch without the need for transfer learning. We constructed 3D metal test specimens with rich porosity defects to acquire data for experiments. We created seeded defects to simulate porosity artificially. The experiments covered a range of pore sizes to assess the ability of the LPBF to produce the pores and to identify any limits associated with the smallest detectable pore in the layer images captured during the build process. Three cylinders were designed containing seeded porosity. The porosity is inserted into the cylinders at various locations, with different shapes and dimensions. We constructed seeded defects as it is challenging to create porosity in a controlled manner, especially when building the objects with several other objects in the same printing job. It is also difficult to identify the pores on images due to their small sizes. Using seeded porosity accomplishes two main objectives: Artificial seeded porosity helps in data labelling as the exact location of the porosity in the powder bed images from the CAD file is known, reducing the need for experts in the labelling stage significantly and avoiding the need to use expensive, time-consuming destructive methods. We designed two labelling approaches for the image data set: The fundamental intuition here is that the seeded CAD information helped build a valid and reliable model, and the XCT showed how we could use non-destructive testing on the 3D samples to tune the model for non-seeded applications. The non-destructive XCT scans of the test cylinders were obtained and correlated with the in-process build images. A complete analysis was performed to accurately evaluate the ML model’s ability to detect different types and sizes of pores.

Convolutional neural networks combine advanced image processing techniques with deep neural networks. We trained a deep convolutional neural network (CNN) using the in-process images. A CNN approach was selected primarily due to their superior ability to extract features from the images without human interaction automatically. Manually extracting valuable features from images is extremely difficult [30] and can often fail to capture local, spatially related information. Automatic feature extraction makes CNNs extremely powerful and computationally more efficient when compared to the other standard ML models such as decision trees [31], random forests [32], and support vector machines (SVMs) [33]. In a comparison by Chouiekh and EL Haj [34], CNNs outperformed traditional ML models such as SVMs, random forests, and gradient boosting classifiers, both in terms of accuracy and training time. CNNs are scalable and capable of handling big data. Another significant advantage of CNNs over traditional ML classifiers is their ability to incorporate transfer learning, allowing complex models to be trained efficiently for a wide range of applications. Transfer learning is the process of keeping the weights and learning of a trained model and using it to solve new similar problems with little re-training [35, 36]. Going forward, the authors intend to use transfer learning on the current models to identify different defect types. The transfer learning will allow the learned weights of the current CNN model to be retained and transferred to new, possibly more complex, deep CNN models to identify a broader range of defects.

The remaining sections of the paper are organised as follows: Section 2 discusses the related work. Section 3 details the materials used in the experiments and the chosen model structure and explains the experimental methodology. Section 4 covers the results and discussion, followed by conclusions in Sect. 5.

Porosity detection from images captured during printing is a highly studied defect in AM. Capturing the porosity on images is challenging and requires a high-resolution camera and a well-lit build chamber. However, identifying the existence of porosity in LPBF parts is complex, and different researchers have followed different experimental approaches. Mireles et al. [24] designed pores of size in the range of 100 to 2000µm and of various shapes (sphere, cubes, circular, triangular, and prism) in their test specimens. HIPing and re-scanning were used to reduce the porosity. CT scanning performed 60% better than (Infra Red) IR cameras in recording porosity on images. Moreover, the IR camera used could not detect pores smaller than 600µm. Similarly, Mireles et al. [23] had created artificial seeded spherical pores of size 600µm to 900µm in test cylinders. Other researchers have created 3D metal objects with porosity induced by varying the laser power, scan speed, and hatch spacing [25, 37‐39]. ML algorithms are widely used for defect detection in LPBF with varying degrees of success. Many studies have used different ML algorithms to identify porosity from powder bed images. The most prominent of those studies are discussed here.

One of the significant works that developed an ML model using images from the build chamber was carried out by Aminzadeh and Kurfess [40]. They employed a Bayesian classifier to detect defects from images. The proposed framework achieved a precision of 89.5%. Gobert et al. [25] aimed to identify porosity types (gas pores and elongation voids) using images from the build chamber and XCT scan data, labelled by human experts. The layer-wise images were classified as nominal or flawed using a binary linear support vector machine (SVM) with high accuracy of 85%. A similar effort by Kwon et al. [37] identified porosity by using 13,200 images from the build chamber. The experiments were divided into seven groups based on different laser power ranges from 50W to 350W, keeping the rest of the printing parameters constant. The experiments revealed that laser power of less than 250W caused porosity. The trained neural network identified porosity images with less than a 1.1% failure rate. Zhang et al. [38] trained a CNN model on images of a 3D metal object printed with titanium powder using direct laser deposition. Five specimens of length 15mm each were printed with varying scan speeds of 1-4mm/s and laser power of 150-250W. Both destructive, cross-sectional analysis and non-destructive (XCT scan) were performed to identify the porosity location in the test objects. The model achieved an accuracy of 91.2% for micropores as small as 100µm [38].

In situ melt pool monitoring is another highly studied area. Many defects occur around the melt pool, such as keyhole porosity, spatter, balling, and under-melting. Various studies have focused on melt pool monitoring to investigate the causes and identification of defects. Yuan et al. [29] proposed a semi-supervised CNN for SLM process monitoring. One thousand and two hundred (1,200) individual tracks of 5mm were printed using 316L stainless steel powder. The data set consisted of frames of 250x250 pixels extracted from 1,200 LPBF videos, out of which 700 were labelled manually. The experiments showed that a semi-supervised approach achieved an accuracy of 93.8% compared to 92.2% accuracy of the supervised approach [29]. The experiments by Li et al. [39] aimed to identify porosity employing a semi-supervised ML approach to overcome the laborious, time-consuming, and sometimes unavailability of labelled data sets for supervised learning [39]. A study based on infrared images was carried out for zinc and its alloys. Data mining, statistical data analysis, and feature extraction techniques were employed to build a real-time in situ monitoring system. Cubes of size 5x5x5 mm were printed using zinc on an SLM prototype machine. The experiments revealed the effectiveness of plume stability monitoring for zinc [41]. In a study by Scime and Beuth [42], in situ monitoring of LPBF employed computer vision and unsupervised ML techniques to identify keyhole porosity and balling signatures from the melt pool. The test specimens were built using an EOS290 LPBF machine with Inconel 718 powder. The features detected in situ were related to the post-build analysis, enabling the classification of the in situ melt pool signatures as defects (keyhole porosity, balling). However, the authors concluded that defect detection based on melt pool signatures was not reliable and required further experimentation [42]. In addition to powder anomalies and melt pool monitoring, Scime et al. [43] presented comprehensive findings in a further study relating to microstructural defects such as porosity, spatter, and soot. They proposed a novel CNN model capable of real-time detection and identification of defects. The model was successfully tested on six different metal printers belonging to three different technologies: electron beam fusion, powder bed fusion, and binder jetting [43]. Instead of using optical camera images, Bartlett et al. [44] employed full-field infrared thermography to identify lack-of-fusion defects. Four cylinders of diameter 20mm and 6mm in height were printed using AlSi10Mg powder with a layer thickness of 50µm on an X-line 1000R SLM machine. The lack-of-fusion defects were detected with an 82% success rate. The pore sizes greater than 0.5mm were detected successfully. However, pores of size below 0.5mm were detected only 50% of the time [44].

Plumes and spatter signatures were used to train a deep belief network (DBN) on images captured by a near-infrared camera. The test specimens were printed on a custom design SLM machine with an integrated infrared camera. Five experimental scenarios were designed by varying scan speed 50-500mm/s and laser power 50-150W. The proposed DBN identified five different melted states with an 83.4% accuracy. Moreover, the proposed framework required fewer parameters, feature extraction, and signal processing [45].

An in situ, bi-stream, deep convolution neural network, which aimed to identify insufficient layer densification, trace discontinuity, and surface deformation defects in LPBF, successfully identified the process-induced errors with an accuracy of 99.4% [46]. Scime and Beuth [47] aimed to identify six anomalies in LPBF; recoater hopping, recoater streaking, debris, superelevation, part damage, and incomplete spreading. Previously, Bag of Words (BoW) and a CNN were used for multi-object detection from a single image. The BoW technique relies heavily on human input. The authors proposed a multi-scale CNN (MsCNN) based on reinforcement learning of the AlexNet CNN, which was trained on a colour image data set called ImageNet. The images were captured from 53 builds (of 3D printed objects) on an EOS M290 LPBF machine. The training data set was composed of 10,071 multi-scale patches, out of which 3,827 were defect-less, 1,896 recoater hopping, 527 recoater streakings, 666 superelevations, 1,297 disturbance, and 1,858 incomplete spreading patches manually labelled by human experts. The proposed MsCNN outperformed the BoW and CNN and was less affected by human bias [47]. In a similar effort to identify powder anomalies by the same authors, a computer vision-based solution was proposed that can be used as a real-time control model with some sufficient future improvements [48]. The image data set of 2,402 images, out of which 1,040 were fault-free images, 264 recoater hopping patches, 228 recoater streaking patches, 187 debris, 314 superelevation, 264 part failure, and 105 incomplete spreading patches, was captured from an EOS M290 LPBF machine using only the built-in camera and LED light. The proposed computer vision algorithm successfully identified failure modes and the exact location of flaws in the final product with microscopic accuracy. However, the use of deep learning algorithms with further improvements in the accuracy of the ML models will be required to use it as an in situ monitoring algorithm [48].

Identifying porosity from powder bed images is crucial for real-time identification and avoiding expensive post-processing methods. Porosity is a complicated defect to detect and avoid in LPBF. The reported studies that used powder bed images for porosity detection had used non-neural network ML models with accuracy generally less than 90%. The few neural network-based studies had better results than the non-neural network solutions, but their performances could be improved significantly by using various sophisticated deep learning techniques. CNNs are extremely powerful due to their ability to extract features automatically from the images. The cited studies covered a range of similar defects in LPBF like recoater problems, surface deformation, layer densification, and track discontinuity that were identified with excellent accuracy using neural network-based solutions. However, its incredibly small size makes porosity difficult to detect using CNNs with great precision. This paper aims to employ sophisticated deep learning techniques to enhance CNNs performance. The extensive and automated hyper-parameter tuning, balanced dataset, artificial data augmentation, and extensive evaluation of the CNN model using various criteria could drastically improve the model’s learning. Apart from that, data quality could be enhanced by inventing better and more accurate labelling techniques and improving the images’ sharpness using affine transformations. The proposed CNN model will become the pioneer in micro LPBF defects detection. The model’s learning will provide a solid base for identifying similar defects from powder bed images.

We constructed the specimens by utilising LPBF. A typical LPBF process is depicted in Fig. 1. The recoater system spreads a uniform dose of metal powder over the build platform. A laser (or lasers) melts specific regions on the newly spread metal powder layer. 3D objects are formed by successively melting 2D cross-sections of the whole 3D object. After each layer is melted, the powder bed is lowered by the thickness of the powder layer, and a new powder layer is spread. The oxygen level inside the chamber is kept to a minimum to avoid metal oxidation during the melting process by introducing inert gases such as argon or nitrogen.

A sample geometry was designed to incorporate seeded defects of different sizes to explore the minimum pore size created by the LPBF machine by validation with XCT of the built parts and to identify the smallest pore size that the layer imaging camera can capture. Spherical and cubic defect shapes are employed to see if this affect detection of the pores. The pores constructed in test cylinders varied in size from 20µm to 2mm, as shown in Fig. 2.

Three cylinders, each with a height of 30mm and a diameter of 12mm, were designed to contain seeded pores within them. Two cylinders, B1 and B2, had circular pores, whereas B3 had cubical pores. Mireles et al. [24] performed similar seeded porosity experiments. Spherical and cubic pores were selected here to study the application of deep neural networks in identifying defects from the images irrespective of the shape of the pores. The cylinders were produced using an SLM500HL machine (SLM Solutions, Germany) from A20X powder with 20-53µm size distribution. The cylinders were built on an aluminium substrate held at 150 \(^\circ\)C throughout the build process. The process parameters used were 360W laser power, 1500mm/s scan speed, 100µm scan spacing, and a layer thickness of 30µm. The stripe scan strategy was used.

The layer imaging camera used in this study is the LayerCam system in an SLM500HL machine. The imaging system comprises two Baumer TXG20 cameras, capturing 2600 \(\times\) 1,440-pixel images of the build area (500 \(\times\) 280 mm). The approximate pixel size of camera systems is 0.2mm/pixel. The cameras are not centred on the build area. The SLM500HL machine carries out image transformation operations before the images are saved for the user. The SLM500HL machine automates the capturing of layer images. Flash strips inside the machine are triggered when capturing the images. An image is captured after each powder layer is spread and after each layer is melted. Therefore, two images are captured per layer. The capturing of data on every layer of the build enables the examination of the specimens’ interior and exterior. The images are saved with a timestamp and layer number in the file name.

A comparison of the results obtained by following both labelling approaches, CAD-assisted and XCT-assisted labelling, is shown in Figs. 7 and 8. The model trained on images labelled using the CAD design achieved an accuracy of 90%. In contrast, the model trained on the XCT-assisted labelling approach achieved 97% accuracy in identifying porosity images from non-porosity images.

The experiments revealed that the circular and cubical shape pores appeared similarly on the powder bed images. We designed the pores in different forms to have a variety of pores’ shapes. However, only gas porosity appeared circular during printing, whereas lack of fusion and voids appeared irregular, elongated, and non-uniform. In our experiments, circular and cubical pores appeared irregular and distorted.

Class imbalance is one of the most common challenges in data-driven solutions. Class imbalance occurs when there is a big difference in the number of occurrences of one class compared to other classes within the data set. When this is the case, accuracy, used as an evaluation metric, can be misleading if the data are biased towards a particular class. A model trained on imbalanced data would result in a biased model. In our experiments, more than 88% of the images were without porosity in the XCT assisted labelling approach, whereas 83% of the images belonged to the non-porosity class in the CAD assisted labelling approach.

The CNN model was trained on both balanced and imbalanced data set to study the impact of balanced and imbalanced data on the model’s training. We used data augmentation methods to combat the class imbalance problem. The XCT-assisted labelled images were balanced by oversampling the minority class. The balanced XCT-assisted images had 2578 non-porosity (50.20%) and 2557 (49.79%) porosity images. The model’s accuracy on balanced XCT assisted images was 97%. However, it is always desirable to evaluate the model with different evaluation criteria instead of solely relying on accuracy. Apart from precision and recall, the F1-measure was employed, allowing a harmonic mean to weight the accuracy measure to account for the imbalance correctly. Further insight can also be obtained by considering the confusion matrix and additional metrics obtained from the confusion matrix. The precision of a binary classifier is given by:Precision is the measure of true positives among all the predicted positive cases. In other words, it specifies how many of all predicted positive cases are true. Precision can be an excellent metric for a manufacturer, where controlling false positives is critical. On the other hand, recall can better predict the model’s performance when false negatives have a high impact. The recall is calculated as:It is a measure of how many true positives are identified. In binary classification with significant class imbalance, a critical task is often to identify between the false-positive and false-negative rates. For instance, a false negative is more dangerous than a false positive in cancer diagnostics. Similarly, in our case, a false-negative is more critical. An image with porosity identified as non-porosity is more dangerous as it will build the object with porosity. This demonstrates why recall is a better evaluation criterion for the model’s performance. For non-porosity, the recall of the model trained on in-process CAD-labelled images was 97%, 99% for imbalanced XCT-labelled images, and 97% for balanced XCT-labelled images. However, there is a significant difference in the recall of the two approaches for images with porosity. The recall on CAD-labelled images was 54%, while the recall of imbalanced XCT-labelled images was 85%. The recall of the model significantly improved from 85% to 97% when trained on balanced XCT-labelled images. The models’ precision, recall, and F1-score on balanced XCT-labelled images were 97%, 97%, and 97%, respectively.

In comparison, the model’s precision, recall, and F1-score on imbalanced XCT-labelled images were 89%, 85%, and 87%, respectively. The balanced data set resulted in a better, more generalised, and unbiased training of the model, as the model outperformed in terms of precision, recall, and F1-score. The model performed well for the majority class, non-porosity images, and there was very little difference in terms of precision, recall, and F1-score as shown in Fig. 7.

The reason for a poor recall value for CAD-labelled images was due to the high false-negative rate. This is observed in Fig. 9. The false-negative rate of the CAD-assisted labelling approach was 45.65%. This means that out of all images with porosity, 45.64% was wrongly classified as non-porosity images by the model. This is a significantly high number of miss-classifications of a crucial class. At the same time, the rate of false negatives in the imbalanced XCT-assisted labelling approach was only 14.72% and 3.26% for balanced XCT-labelled images. This shows that only a fraction of porosity images was miss-classified by our model when trained on balanced XCT-labelled images. The better performance of XCT-assisted labelling is mainly due to the better and more correct labelling of the image set. The CAD-assisted labelling had many non-porosity images wrongly labelled as porosity images. The false-positive rates of the model were insignificant; in the CAD-labelling approach, the false-positive rate was 2.6%, 1.36% for imbalanced XCT-labelled and 2.84% for balanced XCT-labelled images. This means that both CAD and XCT labelling approaches worked well in predicting the non-porosity images, and only a tiny percentage of non-porosity images were wrongly predicted as porosity images. A significant difference between the false negative (14.72%) and the false positive rate (1.36%) was observed for imbalanced XCT-labelled data. However, the false positive (2.84%) and false negative (3.26%) rates for balanced XCT-labelled images were insignificant and resulted in better, unbiased model training.

As stated previously, we chose spherical and cubical shapes for the seeded pores in our test cylinders. Experiments have revealed that the spherical and cubic pores have a similar appearance in the layer images. The model achieved an excellent accuracy of 97% on the image data set labelled with the help of XCT. The model’s accuracy in predicting different-sized pores is shown in Fig. 10. The model attained an accuracy of 96.09% on the most considerable-sized pores; the 2mm.

Similarly, the model predicted the 1-mm-, 0.8-mm-, and 0.5-mm-sized pores with more than 80% accuracy. Using the current printing setup, the smallest visible pore on in-process images was 0.2mm. Our model was 66.67% accurate in predicting the 0.2-mm-sized pores. The model’s ability to correctly identify pores reduced with the pore sizes. Apart from accuracy, precision, recall, and F1-score, the model’s loss curves were also calculated to ensure good training and generalisation. The loss curve of the final model trained on balanced XCT-labelled images is shown in Fig. 11. The model is over-fitting as the train and test loss curves diverge after initial training epochs. That is why we employed \(l_2\) kernel regularisation with a regularisation factor of l = 0.01 and 50% dropout to avoid over-fitting the model. The final model’s loss curve with regularisation is shown in Fig. 12.

This study has established the efficacy of ML models in detecting LPBF defects from powder bed images. It investigated the application of a deep neural network model to predict the porosity from in-process images of LPBF. Besides data capture, labelling is the biggest challenge in developing accurate ML models. The study proposed two labelling approaches, CAD-assisted and XCT-assisted labelling of in-process images of the test cylinders designed with seeded pores of sizes ranging from 0.02mm to 2mm. Experiments revealed that the XCT information is a better benchmark for accurate labelling. The CAD-assisted labelling was unreliable as the pores designed into the CAD file might not be created in the final test specimen. The deep CNN model trained on the CAD-assisted labelled data set achieved an accuracy of 90%. However, the model failed to distinguish between porosity and non-porosity images due to the incorrect labelling of the images. In contrast, XCT-assisted labelling in-process images were reliable, accurate, effective, and produced better results. The deep CNN model distinguished the porosity images from non-porosity images with an accuracy of 97%. The model successfully detected pores as small as 0.2mm in size on in-process images. This is a big step towards porosity identification from in-process images. Moreover, we found that the balanced dataset resulted in a more generalised and unbiased learning/training than the imbalanced dataset. The experiments revealed that the balanced dataset significantly improved the model’s precision from 89% to 97% and the model’s recall from 85% to 97% compared to training on an imbalanced data set. The proposed model’s highly accurate predictability of porosity defects will help in post-processing cost reduction. The early in-process real-time porosity detection will enable AM machine operators to adjust the printing parameters to avoid defects.

In future experiments, a better camera setting will help capture higher definition images. This will, in turn, enable the capture of smaller-sized porosity. Defects such as porosity, balling, lack of fusion, and surface deformation will be created naturally in test objects by altering different process parameters such as laser power, scan speed, and scan strategy. More objects would be designed to encourage defect formation during the LPBF process to develop these findings further. Moreover, the current model’s learning will be tested on new defects by transferring its learning/weights to new deep learning models.