Page 34 - MSAM-1-4

Materials Science in Additive Manufacturing Process optimization of SEBM IN718 via ML

A B

Figure 13. The effect of (A) beam current and (B) scan speed on relative density.

set. The composition and preprocessing of data set or the
inappropriate hyper-parameters could result in overfitting.
Compared with GPR, SVR had better R²and MSE in the
test-set, but the robustness and generalization were worse.
Therefore, the results of GPR model were better in this
study to represent the predicted value of relative density in
parameter space. However, it does not mean that any data
set related to SEBM process is applicable to GPR model.
Complex machine learning model does not necessarily
have better learning performance. Model selection is based
on the performance, interpretability, complexity of model,
size, dimension of data set, and training time and cost. For
a new data set, especially when the feature number is large,
it is necessary to conduct performance tests on different
machine learning algorithms to select the appropriate
Figure 14. Lack-of-fusion defects inside Inconel 718 sample with porous algorithm, such as ten-fold cross-validation. SVR and GPR
surface, 7.5 mA and 6 m/s. chosen here are suitable for few-shot learning, because the
data are limited in this study. Another feature deserved
are generally irregular, and the defect direction may be that more attention of GPR model is that the prediction
either vertical to the powder layer or parallel to the powder error can be directly obtained. Updated GPR model is also
layer. When the input energy is appropriate, the surface
morphology is even. Although the surface is even and free convenient for small databases obtained in engineering
of defects, the properties of the Inconel 718 fabricated by application. When more measured data are obtained, the
SEBM vary greatly when specific parameters are different. new data can be further added to the data set, and the
Therefore, the internal defects can be roughly judged by model can be trained again to enhance the prediction
the surface morphology, while the mechanical properties performance and accuracy.
need further evaluation. The SEBM processing window established through
machine learning model with limited data has larger
4.2. Machine learning models for optimizing SEBM parameter selection range than before [15,25,48] , which
processing window is conducive to selecting the appropriate processing
SVR and GPR models have exhibited good ability of parameters and controlling microstructure. High beam
learning and prediction in this study. As shown in current and scan speed could improve production
Figure 6, SVR model incorrectly predicted that Inconel efficiency. Moreover, using machine learning to predict,
718 fabricated in the region with low beam current the surface integrity is meaningful. There is a high
and high scan speed had high relative density. Error in relative density area at high-energy density, but this area
prediction may be caused by overfitting. Overfitting is not suitable to manufacture due to the severe loss
led to wrong prediction in the test-set in spite of good of dimensional accuracy. Four new samples that were
performance of the machine learning model in the train- manufactured according to processing window proved

Volume 1 Issue 4 (2022) 11 https://doi.org/10.18063/msam.v1i4.23