Page 41 - IJAMD-1-1

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

performance. Researchers must comprehensively weigh enhancing process failure detection. However, traditional
other factors to evaluate the performance of a predictive image segmentation algorithms, including threshold-
model. based and edge-based methods, rely solely on the principle
of gray change and fall short of achieving precise object
3.1.4. Summary of parameter optimization localization and recognition. In addition, these algorithms
Section 3.1 comprehensively discusses the impact of are susceptible to disturbances caused by changes in the
L-PBF technology on the quality improvement of metal experimental environment or the presence of noise. ML
3D-fabricated products, with a particular focus on three algorithms, especially deep learning methods, are proven
critical quality indicators: melt pool characteristics, efficient in handling complex segmentation scenes and can
porosity, and hardness. also be applied in the L-PBF process. For example, Fang
It is imperative to note that while enhancing quality, et al. proposed a novel method using a lightweight U-Net-
the energy consumption and environmental impact of based convolutional NN (CNN) with a high-speed camera
50
the L-PBF process should also be considered. This aspect to capture melt pool signatures (melt pool contour).
represents a relatively under-researched direction in the This approach achieved the highest mean intersection
field. The sustainable application of L-PBF technology over union (MIoU) (0.9806) when compared to two
necessitates a careful balance between optimizing product conventional segmentation algorithms (the threshold
quality and minimizing its ecological footprint. Future segmentation method and the active contour method).
research in this area should aim to develop more energy- In addition, it obtained a low processing time, averaging
efficient L-PBF processes and investigate the environmental 37 ms. In another study, Taherkhani et al. proposed a self-
implications of different process parameters and materials organizing map (SOM)-based segmentation method for
used in metal 3D printing. detecting voids and pores during the L-PBF process using
51
a photodiode light intensity dataset. The SOM model
3.2. In situ process monitoring clustered the photodiode signal into different clusters, and
Even when the process parameters fall within a favorable each cluster was then mapped with printed part geometry
window, the dynamic laser-material interaction, including to complete the segmentation task.
multi-physics and multi-scale processes, has the potential 3.2.2. Regression
to cause defects that affect the mechanical properties
of fabricated parts. Consequently, in situ monitoring In in situ monitoring of L-PBF, the regression task is
methods employed during the L-PBF process are essential often performed to predict certain quality measurements
to detect critical events through the monitoring of signals, values like porosity or mechanical properties using
such as optical and acoustic signals. This marks the first monitoring signals or process features as input. Compared
step toward achieving real-time control systems capable to classification, regression seeks to predict a continuous
of identifying and correcting defects within a suitable variable and is thus more precise and informative. For
response time, thereby improving the repeatability and illustration, the main steps in the work of Huang et al. are
quality of L-PBF parts. Given the abundance of data outlined. The dataset encompasses production parameters
generated during monitoring and the intricate relationship measured during the process, and the mechanical properties
between inspected signatures and part quality, ML of fabricated parts tested are first formed. Subsequently,
methods demonstrate advantages for application in L-PBF ML methods are utilized to predict the yield strength,
in situ monitoring. Previous research endeavors have tensile strength, and elongation of 316L stainless steel, with
successfully applied ML methods to address various in situ downstream production parameters serving as input.
52
monitoring tasks, including segmentation, regression, and Different ML methods are applied for a comprehensive
classification of quality and defects based on data collected comparison, including LR, SVM, DT, and KNN, among
in situ by different sensors. The ensuing discussion delves others. Notably, tree models performed the best, obtaining
into specific tasks and noteworthy works in this domain. an F1 score of more than 0.9. In the study conducted by
Feng et al., tree-based ML models are used to predict local
3.2.1. Segmentation porosity utilizing multi-layer in situ optical tomography
The application of image segmentation algorithms in in situ monitoring images. The main steps and data acquisition
53
monitoring during the L-PBF processes plays a critical role stages are shown in Figure 11. In a separate work, Paulson
in accurate melt pool measurement or defect detection. et al. present four statistical ML models, encompassing
The analysis of images captured during the printing process LR, random forest classification (RFC), gradient boosting
allows for the precise extraction of regions of interest, such classification (GBC), and Gaussian process classification
as the melt pool and potential defects, at the pixel level, (GPC), to correlate temperature histories with sub-surface

Volume 1 Issue 1 (2024) 35 https://doi.org/10.36922/ijamd.2301