Page 48 - IJAMD-2-2

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

approach enables real-time detection of surface quality provided with process parameters, this model offers a
degradation in PBF and more accurate layer-by-layer visual prediction of the expected surface roughness,
surface finish monitoring. Moreover, employing high- showing high alignment with experimental observations
speed imaging data to generate process maps and variability and underscoring its utility for improving mechanical
maps for various alloys can further improve the reliability properties.
and quality stability of metal 3D printing. Notably, the In addition, ultrasound-assisted DED (UADED) has
use of vision transformers enables real-time analysis been explored as a method to improve melt pool stability
of dynamic melt pool changes, significantly enhancing and reduce process defects. An unsupervised learning
process control efficiency (Figure 6A). In DED processes, model has been applied for in situ monitoring of melt pool
143
a vision-based real-time powder stream defect detection dynamics, spatters, and plume formation in UADED. The
152
technique has also been introduced. 144,145 By analyzing melt model effectively classifies process variations by extracting
pool images acquired through a coaxial camera to classify and reconstructing features from high-speed imaging data,
normal versus abnormal states and identifying defective enabling real-time detection of anomalies and enhancing
nozzles during deposition, researchers have demonstrated the overall stability of the deposition process.
improvements in the precision of in-process quality
monitoring. Ultimately, such real-time monitoring and 3.2.3. Design optimization
ML-based analytical methods are critical for automating To meet the dual demands of complex geometries and requisite
and advancing metal 3D printing processes, and they form performance in metal components, effective approaches
indispensable components for establishing stable mass- must be adopted at the design stage, in addition to process
production systems. optimization. Traditional simulation-based optimizations

3.2.2. Property optimization can require prohibitively large amounts of iterative
computation. However, ML, particularly reinforcement
Optimizing the mechanical, thermal, and functional learning, allows for the rapid and accurate exploration of
properties in metal AM is paramount for simultaneously additively manufactured shapes and structures. Combining
ensuring the performance of the final product and the reinforcement learning with multi-objective optimization
stability of the manufacturing process. In particular, algorithms can effectively address diverse objectives, such
combining microstructure analyses (e.g., electron as topology optimization, weight reduction, and structural
backscatter diffraction, X-ray CT) with deep learning is an integrity. For instance, a study integrating support vector
active area of research aimed at quantifying correlations regression with non-dominated sorting genetic algorithm-II
among material properties. 146,147 (NSGA-II) demonstrated the precise prediction of a
For PBF processes, ML techniques have been used to component’s final height and width, which was then
153
improve repeatable quality by analyzing how part location leveraged to refine process control strategies. By enhancing
and chamber pressure variations impact mechanical geometric accuracy and optimizing the interplay between
properties, and by optimizing process control strategies design and manufacturing processes, metal 3D printing can
accordingly. Furthermore, by employing Bayesian be extended to a far broader range of applications.
148
optimization and GP regression, researchers have 3.3. Ceramic
explored expanded process windows to optimize the
mechanical properties and density of Ti-6Al-4V alloys. Ceramic AM has gained significant attention due to its
149
This approach has demonstrated the effective mitigation ability to fabricate complex geometries with high thermal,
of balling instabilities in newly discovered high-density chemical, and mechanical stability. Unlike traditional
process regions. ceramic manufacturing methods that require extensive
post-processing, such as sintering and machining, AM
In DED processes, a CNN-based real-time monitoring techniques such as DIW, SLA, and powder-based extrusion
and melt pool segmentation approach has been introduced offer a more flexible and efficient approach to producing
for process stability and optimization. A YOLO-based ceramic components. However, challenges such as viscosity
150
model was employed to predict geometric features of the control, printability optimization, defect detection, and
melt pool, such as area, height, and width, and to analyze microstructural consistency remain critical obstacles in
the correlations between process parameters and bead achieving high-quality ceramic prints. To address these
geometry, thereby facilitating real-time process control issues, ML has been integrated into various stages of the
(Figure 6B). An artificial intelligence (AI)-based method ceramic AM process, including material formulation,
using conditional GAN was also proposed to generate process parameter optimization, defect detection, and
virtual surface morphologies of Ti-6Al-4V parts. When quality assurance.
151

Volume 2 Issue 2 (2025) 42 doi: 10.36922/IJAMD025130010