Page 104 - MSAM-2-3

Materials Science in Additive Manufacturing Validation of a novel ML model for AM-PSP

fitted with the normal curve and calculated the mean value 87% prediction accuracy, which decreased from 90%
µ and variance σ , which show the grain size distribution. when only L-PBF materials were considered. However,
2
An ANOVA test was applied to analyze the microstructure when more feature information was added, the prediction
difference based on Gaussian coefficients. Table 7 shows results improved. When SFs and TFs were considered in
that materials selection, that is, AM processing condition, is the MP+EBSD condition, the XGBoost model’s accuracy
statistically highly significant (P = 0.000), which proves that increased to 95.5% with a large variance, while the
there is a significant difference among the microstructure linear regression also increased to 90.5%. Through all
information. This also shows that the LPBF_XZ samples five conditions, the XGBoost model proved superior to
have a significant difference compared with all other linear modeling. To analyze the importance of all features
planes. In addition, the DED_XZ plane is the only surface included in the new PSP linkages, feature importance
showing a positive mean value, indicating that the average analysis was employed to determine the feature impact.
grain size in the DED sample is larger than the L-PBF In the feature analysis, in addition to the machining
material. The reason for the large negative mean value in parameters, all residual stresses play important roles in
the L-PBF sample could be due to the considerable number
of tiny discontinuous β grains growing along the α or α’ the XGBoost model training. This indicates that the near-
grain boundary. Therefore, the SEM features in the testing surface residual stress heavily affects the specific cutting
set could be outside of the training boundary. This leads to energy when machining AM Ti-6Al-4V samples. From
poor results in model prediction. To better understand the this, SFs on basal slip systems, prismatic slip systems,
SEM microstructure among these materials, the normal and TFs form EBSD data measurement are shown in
distribution was used to represent CLDs, and the power Figure 10.
functions were used to represent 2-point correlation The high-dimension SEM microstructure information
functions. Tables 8 and 9 show a pairwise comparison significantly improves the PSP linkage accuracy. When
analysis based on the materials’ CLDs and 2-point combining all features, the high-dimension SEM data do
representative coefficients. not show a significant impact on the model. This may need
The second condition could solve the above problem. future in-depth research to reveal the reason, obtain a
In Figure 9B, when the model training from 80% of all better method to reduce dimension and extract additional
data points was collected, the MP condition shows an microstructure data.

Table 7. ANOVA of SEM microstructure information among AM Ti‑6AL‑4V
Variables Degree of freedom Sum of squares Adj SS Mean of square F P‑value
Materials 7 1328590541 1328590541 189798649 11.20 0.000
Number 89 1274031245 1274031245 14314958 0.84 0.839
Residual error 623 10555977115 10555977115 16943783 \ \
Total 719 13158598901 \ \ \ \
Abbreviations: Adj SS: Adjusted sum of squares; SEM: Scanning electron microscope.

Table 8. Comparison of P-values for AM Ti‑6Al‑4V CLDs coefficients
CLDs coefficient EB‑PBF EB‑PBF LPBFHT LPBFHT LPBFNHT LPBFNHT DED DED
XY XZ XY XZ XY XZ XY XZ
EB-PBF XY \ 0.680 0.348 0.000 0.005 0.000 0.933 0.014
EB-PBF XZ \ \ 0.819 0.000 0.021 0.000 0.759 0.109
LPBFHT XY \ \ \ 0.000 0.019 0.000 0.508 0.004
LPBFHT XZ \ \ \ \ 0.172 0.264 0.000 0.000
LPBFNHT XY \ \ \ \ \ 0.038 0.010 0.001
LPBFNHT XZ \ \ \ \ \ \ 0.000 0.000
DED XY \ \ \ \ \ \ \ 0.096
DED XZ \ \ \ \ \ \ \ \
Abbreviations: CLDs: Chord length distributions; DED: Directed energy deposition; EB-PBF: Electron beam powder bed fusion; LPBF: Laser
powder bed fusion.

Volume 2 Issue 3 (2023) 12 https://doi.org/10.36922/msam.0999