Page 36 - IJAMD-2-2

International Journal of AI for
Materials and Design ML-driven optimization in additive manufacturing

into supervised learning, unsupervised learning, semi- detecting anomalies. Techniques such as k-means clustering
supervised learning, and reinforcement learning, each and principal component analysis have been employed to
offering distinct advantages depending on the nature analyze sensor data and monitor inconsistencies in layer
of the available data and the specific manufacturing formation and material distribution. 45
objectives. A visual summary of the major ML categories, Semi-supervised learning, which leverages a
representative models, and their applications across various combination of labeled and unlabeled data, has proven
AM processes and materials is provided in Figure 2, particularly advantageous in AM, where acquiring large
highlighting how different learning paradigms contribute labeled datasets for defect classification can be time-
to process, property, and design optimization.
intensive and costly. Methods such as self-training and
Supervised learning relies on labeled datasets to establish generative adversarial networks (GANs) have been
relationships between input parameters and output quality explored to augment datasets, improve defect detection
metrics, enabling predictive modeling and automated accuracy, and enhance model robustness. Reinforcement
31
defect classification. Commonly used algorithms such learning, which enables an agent to learn optimal strategies
as SVM, random forests (RF), artificial neural networks through trial-and-error interactions with the environment,
(ANN), and convolutional neural networks (CNN) have has been increasingly applied in AM for real-time process
been applied in AM for tasks such as predicting mechanical control and adaptive optimization. Deep Q-networks and
properties, optimizing printing parameters, and identifying proximal policy optimization algorithms have been utilized
defects through image-based analysis. 31,44 Unsupervised to dynamically adjust laser power in PBF and extrusion
learning, which identifies patterns in unlabeled data, is rates in FDM, leading to improved process stability and
particularly useful for clustering process variations and reduced defect formation. 16

Figure 2. Summary of machine learning (ML) and additive manufacturing (AM) techniques commonly used for process optimization. ML approaches and
representative models are shown on the left, while AM technologies, materials, and composites are displayed on the right. ML-driven optimization tasks
in AM include process optimization, property prediction, and design optimization.

Volume 2 Issue 2 (2025) 30 doi: 10.36922/IJAMD025130010