Page 80 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

Figure 9. Methods that mitigate and expand the gap between data and models in AI-driven AM
Abbreviations: AEs: Autoencoders; AI: Artificial intelligence; AM: Additive manufacturing; CNNs: Convolutional neural networks; GANs: Generative
adversarial networks; LLMs: Large language models; ML: Machine learning; RCNNs: Recurrent convolutional neural networks; RNNs: recurrent neural
networks; VLMs: Vision-language models.

transferring knowledge from similar datasets or reducing complexity, there is a pressing need to develop additional
noise and bias in the original dataset. Feature engineering methods tailored to specific data types and applications
methods were introduced to extract salient signals from in AI-driven AM. Furthermore, there is significant
AM processes, starting with image data in 2016 and later potential to incorporate AM domain knowledge into these
extending to time-series and point cloud data in 2018. Data mitigation approaches, enhancing their adaptability and
augmentation has been applied to mitigate representation effectiveness (e.g., physics-informed knowledge transfer
bias, beginning with image data in 2019 and acoustic and data augmentation). 131
signals in 2020. Active learning, introduced in 2021, has
primarily been used for adaptively sampling image data 6.2. Large AM datasets
but has yet to be extended to other data types. Knowledge The research and applications of ML in AM can benefit
transfer methods, categorized by their use cases in from the creation of large-scale and standardized
Figure 9, have evolved significantly. Since 2017, researchers datasets across the process chain. For instance, advanced
have transferred knowledge between AM geometries, monitoring techniques in AM have focused intensely
materials, and machines. Starting in 2018, this approach on capturing common process signatures during layer-
expanded to include transfers across AM technologies and by-layer fabrication. Common examples include time-
from general-purpose datasets to AM processes. In 2021, series representations of process emissions (e.g., point
knowledge transfer between different parameter ranges temperatures, light intensities, and vibrations ) and
132
gained attention, and recently, cross-modality transfers two-dimensional (2D) graphic representations of process
have been implemented to remove noise and extract phenomena (e.g., deposition profiles, material feedstock,
representative features. Semi-supervised learning and self- and melt pool zone 133,134 ) among other signatures to model
supervised learning, incorporating unsupervised learning process concerns of defects and anomalies. Collectively
techniques, have been proposed to leverage unlabeled processing these datasets into a standardized and diverse
examples in AI-driven AM in 2019 and 2020, respectively. representation can lead to expedited research on new
Federated learning has been introduced to this domain materials, designs, and process variations. These large
since 2022 to mitigate data scarcity while preserving data datasets can be a source of domain knowledge for newly
privacy. Multi-fidelity learning further addresses this gap developed ML architectures by leveraging knowledge
89
by leveraging cost-efficient, low-fidelity simulation data transfer techniques. Moreover, these datasets can support
to reduce reliance on expensive experimental data. Given domain-specific benchmarking to evaluate the potential
the considerable gap between data availability and model and contributions of advanced learning approaches (e.g.,

Volume 1 Issue 1 (2025) 13 doi: 10.36922/ESAM025040004