Page 33 - IJAMD-1-1

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

material into the desired shape. When a laser beam serves Machine learning (ML) is a subset of artificial
1,2
as the heat source for melting the powder, the technique is intelligence (AI) and can learn from existing data, analyze
termed laser powder bed fusion (L-PBF) or selective laser data relationships, and autonomously generate predictions.
melting (SLM). The L-PBF production process involves As a result, the integration of ML into L-PBF offers the
3
the construction of a model using computer-aided design advantage of bypassing the need for extensive experiments,
(CAD). The data model is then sliced into two dimensions thereby saving both time and costs. 15,16 ML leverages
and imported into the L-PBF equipment. Next, the past L-PBF experimental data to train models capable of
4-6
equipment’s process parameters are configured. During predicting defects that may occur in the manufactured
the fabrication process, a high-power laser selectively parts during the printing process. This capability of
melts the metal powder layer in a specific area. After ML enables real-time problem resolution or prevention
each layer is melted, the feeding system seamlessly stacks during actual production. Simultaneously, the model can
17
a new layer. The L-PBF process is shown schematically establish correlations between input process parameters
7,8
in Figure 1. and product performance, allowing for the prediction of
The advantage of L-PBF lies in its ability to rapidly specific product quality attributes. For example, Scime
manufacture complex-shaped parts or products by et al. utilized an unsupervised ML algorithm for anomaly
utilizing laser melting technology. This method proves detection and classification of defects such as debris,
versatile, accommodating a wide range of materials and incomplete spreading, and recoater hopping based on
18

contributing to the reduction of the production cycle. in situ monitoring images. In addition, Sanchez et al.
Despite its widespread use in various sectors, including predicted the minimum creep rate of alloy 718 samples
automotive, medical, and aerospace industries, L-PBF in L-PBF using an ML model with input parameters such
still encounters certain challenges. The complexity of the as build orientation, scan strategy, and number of lasers,
19
L-PBF process, coupled with the diverse materials it can achieving an accurate error of 1.40% for the creep rate.
handle, has given rise to challenges concerning low process Considering the aforementioned issues and
repeatability and unspecified machine process parameters. conditions, this paper outlines different ML methods
These issues hinder the establishment of standardization of applicable to quality improvement in L-PBF. The second
best practices, resulting in a paucity of guaranteed quality section presents a detailed categorization of ML and
for L-PBF manufactured parts. 10,11 hyperparameter optimization methods. In the third
The traditional optimization method in the design section, the specific definition and classification of quality
of experiments (DoE) process is directed at enhancing improvement are introduced, and current approaches and
the quality of manufactured parts through a substantial methods of parameter optimization and in situ monitoring
number of experimental iterations. However, it has limited are described in detail. In the former, an overview is
capacity to capture non-linearities in variables and is both presented based on several critical characteristics of
costly and time-consuming. 12,13 Furthermore, sophisticated manufactured parts, including melt pool, porosity, and
techniques such as computational fluid dynamics (CFD) hardness. The latter is presented according to different
encounter challenges in making accurate predictions. ML tasks grounded in a broad series of quality measures,
14
Therefore, identifying the factors influencing disparate including segmentation, regression, and classification. The
part performance and determining the optimal set of concluding section will summarize the key findings and
process parameters becomes a challenging endeavor. offer future recommendations on the application of ML for
quality improvement in L-PBF.

2. Machine learning
ML is a subset of AI with the aim of training machines
to execute tasks based on patterns learned from existing
knowledge acquired through data, observation, and
interaction with the world. The fundamental stages of ML
involve data collection and preprocessing, model training
and testing, and model evaluation. ML can be classified into
four categories based on the learning approach: supervised
learning, unsupervised learning, semi-supervised learning,
and reinforcement learning (RL), as illustrated in Figure 2.
20
Supervised learning involves employing labeled data to
Figure 1. Schematic diagram of a process for laser-powder bed fusion. 9 train the model, while unsupervised learning operates

Volume 1 Issue 1 (2024) 27 https://doi.org/10.36922/ijamd.2301