Page 44 - IJAMD-1-1

International Journal of AI
for Material and Design ML for quality improvement in L-PBF

the defect-free regimen, facilitating the segregation of tasks. This could be achieved by employing more complex
anomalies such as balling, keyhole pores, and lack of fusion ML methods for complex classification tasks. By harnessing
pores during the L-PBF process. Okaro et al. introduced the advantages offered by in situ monitoring, such as real-
a Gaussian mixture model (GMM) with labeled data, time monitoring, there exists the potential to improve the
applying it to distinguish between “acceptable” and “faulty” quality of manufactured parts.
L-PBF builds using photodiode data. 69
(b) Quality prediction 4. Discussion
Besides defect detection, certain research works have The attainment of process repeatability and stability is
focused on classifying the quality of printed parts into essential for L-PBF to fabricate high-quality parts that
low, medium, and high levels based on predefined criteria. satisfy the strict requirements of critical applications
Mohammadi et al. utilized K-means clustering and an such as aerospace and biomaterials. The complex physics
NN method on acoustic signals to predict the quality of interactions and the multitude of variables inherent in
fabricated parts across three levels. In another study, Li L-PBF underscore the efficiency of ML methods as viable
70
et al. adopted a two-step approach. Initially, they used a solutions for improving part quality during parameter
71
GAN data generation model to generate minority defect optimization and in situ monitoring stages. Despite the
samples. Subsequently, a DL-based quality classification significant progress in quality improvement achieved
model was established to classify the augmented balanced through current ML-based applications in L-PBF, several
datasets. Wasmer et al. implemented a spectral convolutional obstacles and challenges persist. 74
NN on acoustic emission to classify the quality of 316L The substantial data requirements pose a significant
stainless steel into three levels. Furthermore, Wasmer et al. challenge for ML methods applied in L-PBF for quality
72
conducted a research study exploring the applicability of RL improvement. Supervised learning methods, which
in the L-PBF process. The study utilized an RL method require extensive labeled data, are commonly used in
73
trained in a supervised manner to connect acoustic signals the reviewed studies for ML applications in quality
with three different quality levels of printed parts evaluated improvement, specifically in parameter optimization
by porosity. The results of the study are promising. and in situ monitoring. This reliance on labeled data

3.2.4. Summary of in situ monitoring necessitates tedious and laborious labeling work, especially
in in situ monitoring, where a vast amount of data is
In Section 3.2, we introduce and analyze research papers collected during the L-PBF process, thereby rendering the
focusing on ML methods applied in in situ monitoring training and testing of the ML methods challenging. In
75
scenarios of L-PBF. Given the widespread adoption of ML other words, the active learning capability of ML, where
methods in predicting a broad series of quality measures the system can interact and autonomously label new data,
and considering that ML methods are the key focus of has not been completely achieved in L-PBF. To ease this
76
this work, representative research works mentioned in predicament, some studies have explored the application
this section are categorized into specific, significant ML of unsupervised and semi-supervised algorithms for
tasks, namely segmentation, regression, and classification. quality improvement in L-PBF, as detailed in the reviewed
During this classification of current works employing ML literature. 18,51,55,65,69 However, it is important to note that
methods in in situ monitoring of L-PBF, several noteworthy the current unsupervised and semi-supervised algorithms
observations emerge. The majority of studies have centered applied in quality improvement are relatively basic and
around defect classification tasks, especially those aimed only achieve high accuracy in straightforward tasks.
at classifying fabricated part quality into three quality
classes: low, medium, and high quality, assessed through In addition, in line with typical scenarios where
measures such as porosity. 54,56,57,71-73 In addition, most tasks the normal regimen significantly outweighs the defect
undertaken are relatively simple, typically involving two- or regimen, research cases often yield imbalanced datasets,
three-class classifications. This phenomenon highlights consequently impacting the optimal performance of ML
a limitation in the impact of applying ML methods. methods in quality improvement. To address this issue, a
Comparatively, the regression task is the least explored subset of current works utilizes generative algorithms, such
among current works, especially in relation to the fitting as GAN, to generate a more balanced dataset. For example,
68,71
of mechanical properties such as yield strength, tensile in the reviewed works, GAN is utilized to generate
strength, and elongation. This underexplored aspect minority defect samples or to learn the representative of a
16
of the regression task implies that there is considerable specific regimen.
potential for further exploration, allowing for proposing The consideration of model generalization is rarely
work that addresses both regression and segmentation addressed in the reviewed works. While current works have

Volume 1 Issue 1 (2024) 38 https://doi.org/10.36922/ijamd.2301