Page 69 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

challenges related to quality control, material performance, neural networks, function as “black boxes,” making it
process repeatability, and scalability. Machine learning difficult to interpret their decisions. Lack of explainability
(ML) technology, which has undergone a remarkable hinders adoption in the industry where understanding the
11
transformation over the past two decades, has shown rationale behind predictions is critical. AM process is very
great potential to address many of these challenges dynamic requiring potential ML applications to adapt in
through intelligent and data-driven approaches. Driven real time, which is challenging for many static complex ML
by exponential growth in computational capabilities, models. With improved data quality, efficient and domain-
unprecedented access to vast datasets, and cutting-edge constrained ML approaches can be more effective in AM.
algorithmic innovations, ML models potentially excel Thus, data quality, quantity, and ML model complexity are
in analyzing complex datasets, making them suitable related matters that deserve more research efforts.
in AM environments where large amounts of factors/ This perspective paper explores the fundamental ML
features are interconnected or coupled. For example, ML themes of data and models in the context of academic
models can predict mechanical properties and optimize research and industrial adaptation. It highlights key data-
process parameters, such as laser power, print speed, and related challenges, including data scarcity, poor data quality,
layer thickness, to achieve consistent part quality. They and high dimensionality, along with ongoing research
6,7
can also enhance defect detection by analyzing in situ efforts to address these issues. The paper provides insights
monitoring data, such as thermal images or acoustic into the development and evolution of ML techniques
signals, to identify anomalies during the printing process over the past decade, emphasizing the increasing model
in real time. Moreover, ML has the potential to accelerate complexity and the emerging learning trends designed to
8
the design process in AM. Generative design approaches overcome current limitations. Finally, the paper outlines
enabled by ML can support structure/shape/topology and proposes key research directions that could accelerate
optimization to balance strength, weight, and material the adoption of ML within the AM industry.
usage. These designs often surpass traditional engineering
9
approaches, unlocking new possibilities in industries such 2. Status of ML in AM
as aerospace, healthcare, and automotive. In addition, ML The status quo of ML applications in the AM field can
enables material discovery and development by identifying be summarized through existing reviews and surveys.
optimal material compositions and predicting their Figure 1 categorizes the scope of these reviews into five
behavior under various conditions, reducing the need for
extensive physical experimentation. major types: Core ML themes, AM process chain, AM
applications, AM software (SW), and AM hardware (HW).
In the past decade, research in applying ML in AM Literature reviews concerning core ML themes focus on
has undergone exponential growth. A large number of elements of ML pipelines including input data, learning
12
publications have come out reporting investigations on algorithms, 13,14 learning techniques, and advanced
15
individual ML models for specific prediction targets in learning approaches. AM process chain (design,
5,16
17
AM. However, the challenges posed by data quantity and process, and post-process phases ) has been the focus of
10
7
10
quality issues were rarely discussed or investigated. Data ML applications aimed at addressing challenges to product
quality and quantity are one of the most important factors design, process repeatability, and manufacturing quality.
18
that determine the performance, reliability, and scalability As a result, several existing reviews have summarized these
of ML-driven solutions. Issues such as noisy and incomplete ML applications in standardized processing technologies
data, data diversity, data standardization, and data labeling (e.g., material extrusion [MEX], 19,20 laser powder bed
are some of the pressing obstacles to overcome. fusion [LPBF], directed energy deposition [DED] ). As
22
21
As the new ML models become more capable, their technology matures, AM applications are expanding in
complexity also increases significantly. Complex ML different domains as evidenced by some of the reviews on
23,24
25
26
models, especially deep learning (DL) architectures, manufacturing, construction, and health, to highlight
are prone to overfitting when trained on limited or the most significant fields. ML has also advanced AM SW
27
28
biased datasets. Moreover, a lack of process information, solutions through progress in monitoring, control, and
29
engineering guidance, and domain awareness during the enhanced digitization. Research on AM HW elements
30
27
training can further degrade the performance leading to (systems and sensors ) to support the integration of ML
models with limited generalization and scalability. Training with AM through systems integration, and sensor fusion
certain large ML models requires high computational represents another theme of these reviews.
resources which is difficult for the AM industry to have According to the Scopus database, more than 130 review
access to. Many advanced ML models, such as deep articles (with search keys: “Additive manufacturing” OR “3d
Volume 1 Issue 1 (2025) 2 doi: 10.36922/ESAM025040004