Page 42 - IJAMD-2-2

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

demonstrated a 99.03% F1-score, highlighting the high top-down setup places the laser or projector above the resin,
accuracy of the proposed detection method (Figure 3B). lowering the platform with each layer, which can disturb
Another study leveraged ML to predict the printability of the resin surface level and expose the newly formed layer to
86
bioink formulations. A total of 210 bioink compositions ambient oxygen, thereby inhibiting photopolymerization.
were experimentally constructed from 16 biomaterials, It also tends to require larger vats and more resin volume.
and decision trees, RF, and ANN were compared. Notably, By contrast, bottom-up printing positions the light source
the RF model achieved the highest precision (90.6%) below the vat and raises the platform upward. While this
and accuracy (88.1%), thus providing reliable bioink reduces some top-down constraints, separating each
recommendations, while ANN offered better recall (87.3%) cured layer from the transparent vat window can induce
and demonstrated superior extrapolation capability. These mechanical stress. Several ML-driven approaches have
findings illustrate how data-driven methods can overcome been proposed to optimize bottom-up printing. 53,99,100
the limitations of purely experimental approaches, enabling One group developed a real-time monitoring system to
rapid identification of optimal formulations. detect part detachment during the upward movement of
99
In vat photopolymerization, process parameters–such the build platform in DLP. Using strain gauges to track
as light intensity, exposure time, print speed, and build platform deformation and a photodetector to measure
orientation–must be considered alongside resin properties UV exposure; they applied an RF classifier along with an
(viscosity, photoinitiator concentration, and reactivity). 84,87-90 exponentially weighted moving average p-control chart to
When printing solid-particle–filled resins in DLP, light detect part detachment in real time and halt the process
scattering may occur as particles interact with the curing automatically. The system achieved an F1-score of 99.03%
beam. 91-93 A similar phenomenon can emerge when cells are and demonstrated robust process stability by stopping
included in a hydrogel-based bioink due to refractive index printing after repeated detections. Another study proposed
differences between cells and the surrounding gel. 67,91,92 an ML-driven predictive model to optimize idle times
Scattering diverts light away from the intended target during the bottom-up approach, where the platform was
region, compromising pattern fidelity. Recent research raised, fresh resin flowed in, and the platform descended
100
used a deep learning-based correction method to mitigate again. Introducing the concept of a prediction feature
cell-induced light scattering and enhance print quality in region allowed them to estimate wait time effectively
DLP-based bioprinting. A master-slave neural network without resorting to complex fluid simulations. A multi-
92
based on U-Net was employed to automatically generate layer perceptron model was trained to identify geometry-
optimized correction masks that suppress the scattering specific idle times, yielding a 47% reduction in average
effect. This approach can significantly reduce iterative trial- waiting time and a 25% overall reduction in total print
and-error. Another study addressed grayscale DLP printing time compared with conventional methods. Similarly, for
optimization by building a recurrent neural network (RNN) bottom-up SLA, a deep learning model has been employed
75
with LSTM layers to learn the relationship between per- to predict layer-wise stress distributions. Training
pixel grayscale values and the final deformed structure. datasets built from FEA simulations fed into a basic CNN
52
Coupling this with finite element modeling (FEM) enabled and a 2-Stream CNN architecture achieved up to a 55%
faster and more accurate deformation prediction, while reduction in prediction error compared to a conventional
an evolutionary algorithm (EA) optimized grayscale neural network (NN) model with high accuracy even in
distribution to achieve target deformation patterns. The non-uniform geometries.
proposed ML–EA hybrid approach significantly reduced ML has also been leveraged for high-precision vat
computational time compared with conventional FEM- photopolymerization (Figure 3C). In one DLP study, an
centric optimization, maintaining comparable precision. RF model was trained to learn the relationships between
In a separate effort to improve geometric fidelity, boundary UV exposure time, light intensity, layer thickness, and
distortion in vat photopolymerization was mitigated using final print error. This approach maintained an average
an ML-driven model that predicts and compensates for printing error below 2.3 µm. Moreover, it demonstrated
101
boundary shifts during layer projection. More recently, a minimum feature size of ~20 µm in complex triply
94
ML-guided temperature prediction models have been periodic minimal surface structures, highlighting the
coupled with numerical simulation to develop optimized model’s capacity for generalization to intricate geometries.
control schemes for DLP printing, improving thermal In addition, ML-based surrogate modeling has been
uniformity and curing accuracy. 95 applied in projection two-photon lithography to optimize

In SLA and DLP, the printing apparatus can be arranged polymerization dynamics and improve printability, further
in either top-down or bottom-up configurations. 96-98 The enhancing precision in microscale fabrication. 102

Volume 2 Issue 2 (2025) 36 doi: 10.36922/IJAMD025130010