Page 78 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

groups or downsample overrepresented groups. In AI-driven to collaboratively train a shared model while keeping
AM, statistical data augmentation methods such as synthetic their local data private. In the context of AI-driven AM,
124
minority oversampling technique, statistical shape federated learning allows the involvement of different
96
analysis, bootstrapping, and stratified sampling were sites with varying machines, materials, and sensor
97
98
99
implemented for AM quality prediction and defect detection configurations in the training process of a global predictive
modeling tasks. Recently, ML-based data augmentation model while preserving data privacy. This can potentially
104
algorithms such as GANs, autoencoders, and diffusion mitigate data scarcity as it encourages more manufacturers
101
100
models have been applied to generate synthetic AM defect to contribute their data without data safety and security
102
images, addressing the imbalance where defective examples concerns. For instance, Shi and Kontar proposed a
104
are typically outnumbered by defect-free ones. Although personalized federated learning model that integrates
these data augmentation methods showed promising model domain adaptation to address data heterogeneity across
performance improvement in AI-driven AM, synthetic local Internet of Things devices. Applied to distributed
data must be validated using reality and diversity metrics to desktop AM machines, this method improves printing
avoid inducing more biases. Nonetheless, synthetic data speed prediction by allowing devices to share knowledge
122
evaluation has not been widely utilized in this domain. while retaining personalized models, enhancing efficiency
105
without sharing raw data. Lei et al. presented a federated
5.6. Self-supervised and semi-supervised learning learning framework for predicting section-wise heat
Self-supervised learning and semi-supervised learning emission in LPBF. It employs a customized model for
leverage unsupervised learning to learn generalized each client and integrates federated learning techniques
representations and reduce the reliance on labeled data, like FedAvg, FedProx, and FedAvgM to aggregate model
thus mitigating data scarcity in AI-driven AM. Self- weights, ensuring privacy while achieving performance
123
supervised learning enables models to learn useful patterns comparable to individually trained models. Despite its
from unlabeled data by creating their own supervision potential to mitigate data scarcity, federated learning
signals (e.g., pseudo labels and pretext tasks), reducing remains underexplored in AI-driven AM, with fewer than
the need for large annotated datasets. For example, Kim ten published studies, representing only a small fraction of
123
et al. employed self-supervised representation learning relevant research.
96
with deep generative models to improve anomaly detection
in AM, addressing the challenge of data imbalance 5.8. Physics-informed learning
between normal and defective products. By transforming ML models informed through process and material
time-series data into images, applying StyleGAN for data physics, 106,125 domain understanding, 114 context
augmentation, and optimizing a boosting-based classifier, awareness, and engineering knowledge 113,127 can
126
the approach enhances anomaly detection performance, lead to improved performance as compared to regular
demonstrating its effectiveness in CNC milling and wire algorithms when applied to AM concerns. As a result,
arc AM. Semi-supervised learning, on the other hand, recent applications saw a significant increase in physics-
utilizes a small amount of labeled data along with a larger informed ML approaches. Several studies have focused
set of unlabeled data, allowing the model to generalize on optimizing process parameters and mitigating defects
123
better with a limited amount of manual annotation. through physics-informed ML. Process and material
128
Zheng et al. proposed a semi-supervised autoencoder physics has been used to model and predict defects such
103
framework for DED real-time quality monitoring as porosity and surface roughness. Another significant
106
that extracts melt pool features through unsupervised aspect among these applications was their focus on
training and leverages a small amount of labeled data for modeling complex thermal and material behavior during
classification, which achieved high prediction accuracy the manufacturing process. AM, in particular metal
108
while significantly lowering data labeling costs. In reality, AM, involves deposition phenomenon characterized by
the representations learned through pseudo labels, pretext material melting, solidification, and remelting, prompting
tasks, and unsupervised learning might contain both efforts to integrate physics into data-driven pipelines to
relevant and irrelevant information to the prediction task. avoid computationally expensive numerical simulations
Extra annotations and finetuning might be necessary to while enhancing the fidelity of these empirical solutions.
achieve desirable performance. As a result, these studies integrate physics to predict aspects
of the deposition process such as melt pool behavior,
5.7. Federated learning
temperature distribution, and the effect of thermal stress
Federated learning is a decentralized ML approach that on the final product, to name a few. Another motivation
110
enables multiple AM systems or manufacturing facilities behind the incorporation of physical knowledge is to make

Volume 1 Issue 1 (2025) 11 doi: 10.36922/ESAM025040004