Page 77 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

was applied to leverage large, publicly available general- reduce the dimensionality and heterogeneity of the input
purpose datasets for enhancing models trained on small- feature space to remove noisy, redundant, and irrelevant
scale datasets, enabling improved defect detection accuracy signals from the raw features. Reducing the dimensionality
and faster convergence. 115,116 Researchers then extended of the dataset alleviates data sparsity, thereby reducing the
120
knowledge transfer to cross-material applications, such as risk of overfitting. Notably, feature engineering methods
transferring learned features from stainless steel to bronze must be evaluated to avoid information loss, which instead
or titanium alloys. 89,90 Cross-process transfer further compromises the model performance.
allowed adaptation across different AM techniques, such as
from LPBF to DED. Recently, cross-modality knowledge 5.4. Adaptive sampling
117
transfer between DED visual and audio monitoring Adaptive sampling techniques are employed in active
data has been conducted to transfer knowledge between and sequential learning to acquire additional data
heterogeneous data types and formats. Despite the using advanced methods, aiming to improve model
92
performance improvements in the literature, knowledge performance and/or address dataset representation
transfer might induce negative transfer that compromises bias. Active learning dynamically acquires new training
54
model performance if knowledge is not transferrable examples based on their potential to improve prediction
between two datasets. In this domain, most researchers performance with minimal data acquisition effort. For
118
54
decide to perform knowledge transfer based on domain example, van Houtum and Vlasea proposed adaptive
94
expertise without conducting transferability analysis. weighted uncertainty sampling, an active learning method
Safdar et al. introduced transferability analysis to this that balances uncertainty with random sampling to
91
domain and decomposed AI-driven AM into AM and AI improve model performance while reducing annotation
components. They developed similarity metrics tailored to requirements. Applied to AM quality prediction and
each knowledge component to evaluate the feasibility of benchmark datasets, this method outperforms state-of-
knowledge transfer between datasets. These metrics assess the-art strategies, reducing annotations by 20 – 70% in
whether the datasets exhibit sufficient similarity to justify most cases. Raihan et al. proposed a surprise-guided
93
and guide knowledge transfer, enabling more systematic sequential learning framework, integrating adaptive
and effective adaptations of models across different AM sampling that efficiently explores and exploits the feature
processes or conditions. space inspired by Bayesian optimization and a conditional
GAN to model the melt pool morphological characteristics.
5.3. Feature engineering Additional examples acquired using adaptive sampling can
Feature engineering is the process of selecting and significantly increase the information-richness of small-
transforming input features to improve the performance of scale datasets, thereby effectively mitigating representation
an ML model. In AI-driven AM, it encompasses domain bias and data scarcity. However, its application is limited
63
knowledge-based methods, feature selection, and feature to scenarios where additional data acquisition is feasible.
119
learning. Domain knowledge-based methods derive
features informed by physical phenomena or sensing 5.5. Data augmentation
technologies, such as thermodynamics, signal processing, Data augmentation has been widely implemented to mitigate
and computer vision. Feature selection algorithms rank representation bias and data scarcity when additional data
raw features based on redundancy and input-output acquisition from physical experiments is unavailable. 54,55
dependencies, while feature learning transforms raw Data augmentation methods generate additional data for
inputs into representative features using statistical or the underrepresented groups by synthesizing new examples
ML methods. In a DED multi-modal monitoring system or partially altering existing examples, thereby improving
established by Chen et al., the authors utilized domain data diversity. According to literature, data augmentation
54
knowledge to extract temperature field features like peak techniques can be categorized into domain knowledge-based,
and mean temperatures, visual features like melt pool width statistical, and ML-based methods. Domain knowledge-
and length, and acoustic features like spectral variance and based methods manually synthesize underrepresented
skewness. Spearman’s correlation coefficients among examples in the dataset using domain knowledge. For
49
the input features and the label were computed to select example, Becker et al. generated in-process AM acoustic
121
the most important input features. Feature learning signals by applying signal-processing techniques such as time-
techniques, including linear discriminant analysis and stretching and amplitude-shifting. Wong et al. synthesized
95
principal component analysis, were performed to generate AM X-ray computed tomography images using computer
representative features for dimensionality reduction and vision techniques such as zooming, rotation, and flipping.
dataset visualization. Feature engineering methods usually Statistical methods statistically oversample underrepresented

Volume 1 Issue 1 (2025) 10 doi: 10.36922/ESAM025040004