Page 85 - MSAM-1-1

Materials Science in Additive Manufacturing A ML model for AM PSP of Ti64

represents the average RMSE, and corresponding standard stresses from XRD measurements, SFs and Taylor factors
deviations are presented for each feature combination. It captured from the EBSD map, and features from PCA of
should be noted that when all features are used, the standard the SEM images. In the XGBoost ML model, the weight
deviation for the XGBoost ML model is minimal (0.5 % (importance) of a feature will increase if it participates in
RMSE with ± 0.025% standard deviation). the creation of each tree during the forest building stage of
the ML model. Dhaliwal and Nahid (2018) reported that
4. Discussion when the tree is growing, for every gain of splits that use the
[43]
As shown in Figure 7, predictors trained with only feature, the importance of this feature increased . Hence,
machining parameters achieve about 90% accuracy in the feature score (F-score) is introduced to evaluate the
both XGBoost and linear models. However, when the SFs importance of a feature by calculating the number of times
and Taylor factors features are independently included in a feature appears in a tree and is presented in Figure 8.
the ML models, the XGBoost model’s accuracy increases As expected, machining parameters are the most
to 94.6%, with larger variance (±5%), and the linear important role in the S-P linkage construction as expected.
regression model was not improved with SFs and Taylor Based on the F-scores of the last model, where all features
factor. It is evident that the incorporation of residual stress are included, the surface residual stresses (RSs) have the
and crystal orientation increased the accuracy to predict highest impact among all material structure features,
machining behavior in a ML model. The larger variance which can be observed in Figure 8e, followed by SF in
shows that additional descriptors that represent other the prismatic slip systems and small strain Taylor factor
features of a microstructure should be considered. One calculated from EBSD measurements. PCs calculated from
possible reason is that the linear model is less capable the SEM microstructure data also show a positive effect
of synthesizing high-dimensional data. However, it is on predicting the specific cutting energy. If only SEM
evident that linear regression models are not sufficient information with the machining parameters was used to
to accurately predict the machining behavior of complex train the XGBoost model, PCs with higher variance (initial
microstructures produced through metal AM. It was also few PCs) could positively influence the prediction accuracy
observed that the accuracy of the ML model increased of specific cutting energy, as shown in Figure 8d. However,
to 97% when machining parameters and SEM spatial incorporating additional PCs of SEM microstructure
functions were considered to predict machining behavior. result in lower prediction accuracy. This could be due
This can be attributed to the relatively smaller grain size to the collinearity of SEM data with XRD and EBSD
and higher grain density in PBF AM systems, which are measurements. In addition, incorporating additional
captured in the SEM descriptors (2-point correlation PC elements of microscaled structural information may
function and CLDs). Finally, when all the design features not represent useful data for a specific prediction goal of
are included, the XGBoost ML model achieved an accuracy machining behavior. SFs and Taylor factors calculated
of 99.5% ± 0.025. from all three slip systems also possess some importance

It can be deduced that the XGBoost ML model in predicting the specific cutting energy. Since the Taylor
developed in this study is consistently more accurate when factors measurement highly depends on EBSD data quality,
compared to the linear regression model, which is widely zero solution in EBSD data could reduce the accuracy of
used in this field. The accuracy of the linear regression the Taylor factor.
model decreases in a high-dimensional dataset, such In summary, incorporating microstructure, SFs, and
as MP+SEM. Zhang et al. (2018) indicated that in the Taylor factors extracted from EBSD data, and residual
ML model, the overfitting problem might happen when stress calculated from XRD measurements were the most
the dataset has a large dimension of input factors with a effective in accurately predicting machining behavior
relatively small size of responses . Due to the smaller when compared with a baseline model where only
[42]
dataset volume (72 data samples) in both combinations of machining parameters were incorporated in both linear
machining parameters with residual stresses and EBSD, the regression and XGBoost model. It is also evident that
variance of testing accuracy is large for both ML models, a high-dimensional and large-volume dataset greatly
which could be attributed to overfitting in XGBoost model. improves the accuracy of prediction, but advanced ML
It is inferred that at higher dataset volume, the XGBoost approaches are necessary to handle such complex datasets.
model becomes more accurate in prediction performance. Due to the relatively small data volume of residual stress,
To isolate the individual effect, impact feature SFs, and Taylor factors, models for MP+EBSD, and
importance analysis was performed on all the features MP+XRD show a lower accuracy when compared with
considered in this XGBoost ML model, namely, residual model MP+SEM.

Volume 1 Issue 1 (2022) 13 https://doi.org/10.18063/msam.v1i1.6