Page 76 - IJAMD-2-2

International Journal of AI for
Materials and Design Prediction of AM defect based on DL

essential approaches to improving the comprehensive training; however, large amounts of data in engineering are
performance of LPBF parts. 2 often not available. For example, experimental studies on
Defect modeling in an LPBF process is important. LPBF often use the design of experiments to reduce the
Modeling involves simulating and predicting the formation number of experiments to save costs and time. Therefore,
of defects and guiding the process of optimization. It helps much of the experimental data of LPBF is for specific
achieve better control over LPBF parameters, including the tasks; it is not prepared for ML/DL and the ML/DL model
laser power, the scanning speed, etc., to reduce defects and creation. The work in this paper is part of the author’s effort
improve the final part quality. Due to inadequate melting to explore a DL model on a small experimental dataset
and bonding occurring between adjacent layers, LOF is a with unbalanced data.
serious defect that compromises the overall strength and The main objective of this research paper is to
cohesion of the printed structure of LPBF. 3 establish DL models on a small dataset (with unbalanced
Sequentially learned random forest with enhanced data) and predict the LOF defect of LPBF based on the
sampling was presented for robust and efficient defect established DL models. The remainder of this paper
classification in an LPBF process. A machine learning is organized as follows: the second section introduces
1
(ML) framework that combines a fuzzy logic scheme, a LPBF; the third section introduces data (the dataset used
self-organizing map, and a tailored U-Net architecture in this paper), data pre-processing techniques (min-
was presented to improve the defect prediction capability max normalization and z-score standardization), and
of an LPBF process. The framework and methodology evaluation metrics for DL; the fourth section presents
were employed to predict general defects, for example, four DL methods utilized in this paper, including the
keyholes and LOF, by analyzing in situ optical tomography Elman neural network, the Jordan neural network,
data. Furthermore, a quality assurance professional the DNN with weights initialized by the deep belief
was permitted to use the expert knowledge through network (DBN), and the regular DNN based on the four
customizable fuzzy rules. 4 algorithms (rprop+, rprop−, smallest absolute gradient
[sag], and smallest learning rate [slr]); the fifth section
A deep learning (DL)-based approach to defect gives results and discussion; and the sixth section
detection was proposed that uses various convolutional presents the conclusion and future research.
neural networks and transfer learning techniques to
automatically segment and detect the melt pools of LPBF 2. LPBF
and porosity from microstructure images. The research Among the most studied laser-based AM process for
demonstrated the ability to detect and segment melt pools metals and alloys is LPBF or selective laser melting. It is
and porosity accurately, even with a limited set of training suggested that the term “LPBF” should be used according
data. It also paves the way for effective quality monitoring to ASTM standards. LPBF utilizes a high-power laser beam
and quantitative evaluation of defects in LPBF. 5
to melt the pre-defined contours selectively in subsequent
LPBF has been a widely utilized AM method for metals. layers of powder. The molten metal pool solidifies rapidly
In situ, sensing and monitoring have been an effective by cooling. The underlying build platform is lowered,
approach to detecting LPBF defects. A real-time defect followed by another layer of powder deposition. This
detection system with feedback control was proposed cycle is repeated successively till a 3D solid object is
based on the integration of 3D point cloud data processing constructed. The process is finished inside a chamber full
with DL. This research demonstrated that there would be of atmospheric gas (argon, nitrogen) to avoid oxidation.
8
great potential for 3D point cloud-based DL in improving Figure 1 illustrates the LPBF process. A nickel-based alloy
defect detection and quality control in LPBF. 6 is used in this research.
There are more hidden layers in a DL network than Although there is great potential for LPBF, surface and
in a traditional artificial neural network (ANN). DL can subsurface defects such as LOF and internal porosities
automatically learn and discover relevant features from affect the application of internal LPBF, especially where
data or examples; therefore, it can detect patterns and fatigue life is a major concern. The optimization of
trends and make predictions. Generally, much data is the fatigue performance of LPBF parts is substantially
needed for training a DL model. A DL model can improve dependent on the manufacturing process parameters and
its performance if more data is used for model training. post-processing of LPBF. It is significant to eradicate the
Overfitting and data availability are two major challenges in formation of surface and subsurface critical defects (e.g.,
DL. Generally, it is necessary to collect a lot of quality data LOF) by optimizing the melting parameters to obtain
7
(such as balanced data, not biased data, etc.) for DL model materials with good fatigue performance. Figure 2 shows
9
Volume 2 Issue 2 (2025) 70 doi: 10.36922/IJAMD025060005