Page 119 - MSAM-3-2

Materials Science in Additive Manufacturing Hybrid lattice structures design with AI

Figure 4. Examples of hybrid lattice configurations (red color represents P-Honeycomb cell and blue color represents G-Honeycomb cell).

Figure 7 provides a visual representation of the 3,000 randomly reduction. Around 20 epochs into training, the model loss
created hybrid lattices within the property space. Specifically, for the validation set reaches a state of convergence.
the elastic modulus along the X- and Y-directions ranged from Subsequently, both the training and validation set
10 MPa to 24 MPa, while Poisson’s ratio spans from 0.22 to losses become below 0.05 after 80 epochs, indicating
0.34. Remarkably, the hybrid lattice designs demonstrated the robust prediction capabilities regarding the properties of
capability to exhibit isotropic behavior across the entire range hybrid lattices. To validate the training outcomes further,
of elastic modulus values while also offering a high degree the target and predicted properties for the validation set
of anisotropy within the dataset (Figure 7C). Furthermore, were extracted. Figure 9 compares the target and predicted
it is observed that for a given elastic modulus along the values of E , E , ν and ν for the hybrid lattices within
y
x
yx
xy
X- or Y-direction, the lattice could show varying Poisson’s the validation set. Overall, a good agreement was observed
ratios (Figure 7D). The randomly generated hybrid lattice between the actual and predicted properties. However,
structures show a broad spectrum of mechanical properties. it is notable that the model slightly underestimated
Consequently, the dataset generated for the training of the the elastic modulus of the hybrid lattice. Moreover, the
BPNN can provide comprehensive insights into the properties predictive accuracy varied across different properties: the
and behaviors of the hybrid lattices. model performed better in predicting the modulus along
the Y-direction compared to the X-direction, whereas it
3.2. Training and validation of BPNN
demonstrated higher accuracy in predicting the Poisson’s
A total of 3000 random hybrid lattices were generated, ratio along the X-direction. The trained model was saved
with 80% of the dataset (2,400) allocated for training and for future predictions of properties based on the lattice
the remaining 20% (600) for validation. The MSE loss configuration.
function was employed to quantify the model prediction
error. Early stopping was implemented to improve training 3.3. Performance of BPNN
efficiency and prevent overfitting. Figure 8 illustrates the The performance of trained BPNN was tested by the
evolution of model loss for both the training and validation dataset prepared using the homogenization method. The
sets throughout the training process, demonstrating the randomly generated lattice patterns were simplified to
convergence status of the model. Notably, as the total binary matrix and input into the trained BPNN. Elastic
iterations increase, the model loss exhibits a significant modulus along the X- and Y-direction and Poisson’s ratio

Volume 3 Issue 2 (2024) 6 doi: 10.36922/msam.3430