Page 38 - IJAMD-2-2

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

reliance on traditional empirical trial-and-error methods. of ML models, numerous studies have attempted to
54
These ML-driven strategies significantly enhance process monitor and control defects that emerge during FDM
control, reduce production time and material waste, and printing. In parallel, experimental approaches based on
improve overall manufacturing reliability. physical sensing techniques have been used to measure
process-induced deformation. For example, Kantaros
3. ML-based 3D printing optimizations and Karalekas utilized fiber Bragg grating (FBG) sensors
3.1. Polymers to measure residual strains in ABS parts fabricated
via FDM, offering a physics-based understanding of
Polymers are macromolecules composed of covalently internal stresses through thermo-optic and strain-optic
bonded repeating units, forming long chains that exhibit coefficients. 78
diverse rheological properties. Such chain structures endow
polymers with low density, flexibility, and–in certain cases– A common approach involves in situ monitoring of
viscoelastic behavior. Their physicochemical properties image data. 51,77-82 For instance, capturing the top view of an
34
can also be tuned during synthesis and processing to FDM part printed with ABS or PLA at certain build stages
meet specific requirements. Owing to these adaptable and then applying image-processing techniques combined
characteristics, polymers are widely employed as elastomer with SVM to detect the presence or absence of defects
matrices in soft robotics, 55-57 soft actuators, 58,59 and soft has been reported. Similarly, CNN have been employed
79
sensors. 8,60,61 to classify defects from top-view images of partially
80
Because of their tunability and ease of processing, completed prints. Moreover, FDM processes are highly
polymers are well-suited for various 3D printing sensitive to environmental conditions, which may drift
81
techniques. Material extrusion techniques like FDM over time (so-called “drift” in process parameters). Such
62
typically use thermoplastic filaments such as ABS and domain shifts can degrade the performance of previously
PLA, while DIW prints polymer-based materials, trained diagnostic models. To mitigate this, a deep
63
including hydrogels and elastomers. Recent studies learning-based defect detection system has been developed
37
54
have leveraged ML-based approaches to optimize DIW that utilizes top-view images of PLA-printed samples.
parameters, especially for bioinks and hydrogels used in The training data were balanced using a conditional
biomedical applications. Key rheological behaviors–such GAN (CGAN), and domain adversarial neural networks
37
as apparent viscosity, shear-yielding, and shear-thinning– (DANN) were employed to adapt to parameter drift. This
are typically considered for DIW formulations. 8,34,64,65 A system achieved a classification accuracy of 91.01% in
summary of recent ML applications and model types used detecting various FDM defects. Nevertheless, relying solely
for polymer-based 3D printing is presented in Table 1. on top-view images can pose challenges in detecting flaws
on vertical surfaces. Complementing purely data-driven
In addition, vat photopolymerization techniques such approaches, hybrid models that incorporate physics-based
as SLA and DLP enable high-resolution printing for knowledge into ML frameworks have also been proposed.
biomedical 66,67 and electronics 68,69 applications. During the Kapusuzoglu and Mahadevan demonstrated a physics-
46
printing process, photosensitive resins must maintain a informed ML model for fused filament fabrication, which
low viscosity to accommodate continuous layer-by-layer integrates thermal and mechanical principles to improve
fabrication and be formulated to exhibit optimal reactivity prediction accuracy and reduce the data dependency often
within the targeted wavelength range of the photoinitiation seen in conventional ML models.
system. While acrylate- and epoxy-based monomers
are commonly employed, 68-71 other photosensitive Beyond defect detection, real-time process
formulations may be used for specialized applications. For compensation offers a more proactive strategy. One study
example, photocurable hydrogels are frequently considered used a CNN-based classifier to analyze nozzle-near images
for bioprinting. 67,71,72 in real-time, detecting under-extrusion or over-extrusion
and automatically adjusting extrusion flow rates. By
82
3.1.1. Process optimization incorporating a feedback loop, the system reduced
Despite the advantages of polymer-based 3D printing, corrective action times to 8.6 s for under-extrusion and
process parameters can significantly influence printing 9.8 s for over-extrusion–faster than human intervention.
quality and yield. In FDM, parameters such as nozzle Another work further automated process correction by
51
temperature, layer height, printing speed, build modifying G-code based on defect types. Using a YOLO
orientation, and infill settings often affect part integrity, object detection model, under- and over-extrusions were
potentially causing interlayer delamination and uneven identified in real-time, and flow rates or recent G-code
stress distributions. 73-77 With the rapid development commands were updated accordingly. A streamlined

Volume 2 Issue 2 (2025) 32 doi: 10.36922/IJAMD025130010