Page 45 - IJAMD-2-2

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

developed as a forward prediction tool to precisely forecast and potentially misaligned with experimental results.
shape deformation from a given material distribution; To address these challenges, ML approaches have gained
this was then coupled with GD and EA to form an inverse prominence. ML models can learn intricate correlations
design loop. By optimizing the voxel-level material layout, among variables, predict optimal parameters, minimize
the proposed method effectively controlled 3D shape defects, and enhance print quality.
transformations in more complex configurations. The An augmented ML framework integrating mechanistic
results highlight how ML-driven strategies can replace or modeling and domain knowledge has also been proposed
augment high-cost finite element simulations, significantly to mitigate the lack of fusion in laser PBF (LPBF) processes
accelerating the design of 4D-printable smart polymer (Figure 5A). By computing five dimensionless variables
136
structures (Figure 4D). from process parameters and material properties and

3.2. Metals applying decision tree and linear regression models,
lack of fusion defects were predicted with over 90%
Metals exhibit a regular atomic lattice structure held accuracy. The resulting index and process maps facilitate
together by robust metallic bonds, which impart notable quantitative defect prediction and alloy-specific parameter
physical properties such as high density, strength, and optimization.
electrical/thermal conductivity. 117,118 Moreover, techniques
such as heat treatment, alloying, and post-processing While many ML applications focus on real-time
enable the tailored adjustment of these physical and monitoring and control during printing, model-driven
chemical properties, facilitating their application as approaches for pre-print process parameter optimization
essential materials in aerospace, medical, and automotive are also gaining traction. A representative example is the
industries. 119-123 Recently, to achieve complex geometries integration of high-throughput LPBF experiments with
and weight reduction, the integration of metal materials ensemble ML models to optimize process parameters
137
with 3D printing technology has garnered significant for 316L stainless steel (Figure 5B). By simultaneously
attention, with two principal AM processes widely fabricating 54 samples under varied laser power and
employed. scanning speed, and training ML models on porosity,
hardness, and corrosion data, researchers successfully
For enhanced geometric complexity and reduced predicted optimal parameters that significantly reduced
weight, metals are increasingly used in 3D printing porosity (<0.1%) while enhancing tensile strength and
technologies. 124-127 A representative metal 3D printing corrosion resistance. The trained model also demonstrated
method, PBF, selectively melts and solidifies the powder transferability to AlSi7Mg, indicating the framework’s
using a high-energy source. 128-130 Typically, fine metal potential for cross-material optimization and accelerated
powders (10–60 µm) such as aluminum, stainless steel, process design.
nickel, or titanium are used, enabling the production of
high-precision parts. Another method, DED, deposits Mitigating defects that arise during printing is crucial
metal powder (or wire) by melting it with a high-power for ensuring the reliability and performance of the final
laser or a plasma arc. 43,131-133 While the alloys commonly parts, and research has been actively conducted to address
utilized in PBF can also be employed in DED, the powder issues related to thermal fluctuations. In particular, deep
used in DED is generally larger (around 50 – 150 µm). reinforcement learning has been applied to PBF processes
Compared to PBF, the DED process may have lower to regulate laser speed and power in real time, thereby
dimensional accuracy but fewer constraints on part size maintaining a stable melt pool shape and preventing
and faster production speeds, making it highly attractive overheating. 138,139 Such methods effectively reduce defects
for industrial applications. A summary of representative and improve print quality. Meanwhile, for DED processes,
ML models and their applications in metal-based 3D ML-based thermal characteristic prediction has been
printing is provided in Table 2. investigated. Using XGBoost and LSTM algorithms,
researchers developed a model that analyzes the
3.2.1. Process optimization relationships between process parameters and temperature,
140
Various defects can arise in metal AM (both PBF and thereby predicting temperature distributions. This
DED), such as excessive porosity, balling, lack of fusion, strategy effectively controls temperature variations, leading
spattering, and warping. These defects occur due to the to improved microstructure formation and mechanical
complex interplay among multiple process parameters, properties.
including laser power, scanning speed, and powder feed Furthermore, a vision-based real-time defect detection
rate. 122,134,135 Conventional optimization typically relies technique was introduced using an autoencoder-based
on trial and error, which can be time-intensive, costly, ConvLSTM model. This model autonomously detects
141

Volume 2 Issue 2 (2025) 39 doi: 10.36922/IJAMD025130010