Page 76 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

Table 2. An overview of the advanced approaches in AI‑driven AM
Advanced approach Description Target data challenge Example
publications
Data integration Combining datasets from multiple data sources Small data size and high heterogeneity 49,73,88
Knowledge transfer Transferring knowledge between data sources. Small data size and high heterogeneity 89-92
Feature engineering Reducing dimensionality and extracting representative features High dimensionality and heterogeneity 49,55
Adaptive sampling Iteratively sampling additional data to improve data Small data size and low data quality 93,94
representativeness
Data augmentation Oversampling underrepresented groups and/or downsampling Small data size and low data quality 95-102
overrepresented groups
Self-supervised learning Learning representations from unlabeled datasets based on pseudo Small (labeled) data size 97
labels or pretext tasks
Semi-supervised learning Learning representations from both labeled and unlabeled datasets Small (labeled) data size 103
Federated learning A decentralized learning approach that trains a model using Small data size and high heterogeneity 104,105
multiple data sources while ensuring data privacy
Physics-informed ML Embedding material physics, domain understanding, context Small data size 106-114
awareness, and engineering knowledge into ML
Note: The example publications are not exhaustive.
Abbreviation: ML: Machine learning.

integrating knowledge across sources, scales, and processes.
Table 2 provides an overview of the advanced approaches
discussed in this section and indicates the data challenges
that each approach tackles.
5.1. Data integration
Data integration in AI-driven AM combines datasets from
multiple sources across different materials, geometries,
process parameters, sensors, modalities, and prediction
tasks. Although data integration increases the total
amount of information, special data handling and model
architecture are required to bridge format discrepancies
Figure 8. Absolute and relative composition of dominant ML methods
and algorithms across a decade of AM applications (2015 – 2025). and information gaps between different datasets. For
49
The methods grouped under “Others” include Boltzmann machines, instance, Chen et al. proposed multi-modal data fusion to
quadratic regression, sparse representation, locally linear embedding, integrate data of different formats and definitions acquired
hidden Markov models, latent Dirichlet allocation, neural operator from a thermal camera, a CCD camera, and a microphone
networks, mixture models, quadratic discriminant analysis, diffusion for in situ DED defect detection. The authors synchronized
models, and RL techniques (e.g., Q-learning, dual deep Q network, classic
deep Q network, deep Q network, deep deterministic policy gradient, and the multi-modal data and extracted physics-informed
multi-armed bandit). features from each modality to align the input data
Abbreviations: AM: Additive manufacturing; ML: Machine learning; type. Scime and Beuth integrated multi-scale powder
73
RL: Reinforcement learning. bed camera images to train a defect detection model by
preprocessing them into the same shape and concatenating
transfer, feature engineering, adaptive sampling, and them at the input layer. Olleak and Xi combined low-
88
data augmentation methods to improve the information fidelity simulation data of the LPBF process with limited
richness of AM datasets. Model-centric approaches such experiment data to reduce the amount of experiment data
as self-supervised learning, semi-supervised learning, required for training a melt pool prediction model.
federated learning, and physics-informed ML have also
been proposed to leverage advanced ML architectures 5.2. Knowledge transfer
while embedding AM domain knowledge. Overall, Knowledge transfer in AI-driven AM has advanced
advanced data handling and ML methods in AM are significantly to address challenges such as data scarcity
evolving to handle diverse data types and modalities, and high data collection costs. Initially, knowledge transfer

Volume 1 Issue 1 (2025) 9 doi: 10.36922/ESAM025040004