Page 71 - Read Online

Gao et al. J Mater Inf 2023;3:6 https://dx.doi.org/10.20517/jmi.2023.03 Page 9 of 15

Validation and prediction for the selected ML model
Based on the selected MLPReg model, the model hyperparameters were optimized by a random search
strategy with a 10% testing dataset. The structure of the optimized MLPReg model was 5 × 100 × 200 × 3 (1
input layer with 5 features, 2 hidden layers with 100 and 200 neurons, and 1 output layer with 3 output
features). The detailed optimization results are displayed in Supplementary Table 3. The comparison
between the predicted and experimental properties of the trained MLPReg model is shown in Figure 5.
When the predicted result is very close to the actual value, the data point should be in the vicinity of the line
y = x. The results show that the predicted UTS, YS, and EL are in good agreement with the experimental
results on both the training and testing set. The average R score of training and testing sets can reach 0.8962
2
and 0.8985.

Besides, different operators and experimental environments can affect the mechanical properties of the
target alloy. In order to test the generality of the model and further verify the model accuracy, 3
representative alloys were selected from each of the three systems [5,48,58,59,78,79] (i.e., Al7SiMg, AlSi10Mg,
Al12Si) from the cleaned dataset as the second test set. The composition and manufacturing parameters of
those 9 alloys are shown in Supplementary Table 4. The comparison of predicted and experimental results
of the second test set is also shown in Figure 5. The results showed that the selected ML model can predict
the UTS, YS, and EL of all the alloys simultaneously, especially reproducing the mechanical properties of the
9 alloys in the second test dataset well. Therefore, it indicates that the presently selected ML model can
simultaneously predict the UTS, YS, and EL of all the SLMed Al-Si-(Mg) alloy with high accuracy and will
be utilized to design the alloy with high performance further.

Based on the trained and validated MLPReg model, the mechanical properties of alloys over a wide range of
composition and process parameters can be predicted. A method for randomly generating composition and
processes over the parameter space was used to create the big input dataset in this work. The range of Si is
between 4.1wt.% and 12.1wt.% with Δw(Si) = 1wt.%, and that of Mg varies from 0wt.% to 0.6wt.% with Δw(
Mg) = 0.1wt.%. The value of Ed varies from 26 J/mm to 171 J/mm with a step of 5 J/mm . Supplementary
3
3
3
Table 2 shows that over 95% of the high-performance alloys were prepared using 67°, 73°, or 90° as the layer
rotation angle. To achieve the best performance and increase the computational efficiency, three values (i.e.,
67°, 73°, and 90°) were selected as candidates to make predictions. Figure 6 shows the predicted results
based on the trained MLPReg model for random 11,300 data (combinations of compositions and processes)
in SLMed Al-Si-(Mg) alloy. The plot has been colored with Quality Index (QI). The strength-ductility trade-
off relationship indicates that monolithic materials have a limitation in achieving simultaneously high
strength and elongation. In Figure 6, the alloys with high performance (i.e., high QI) usually presented a
high strength with medium elongation.

To further validate the accuracy of the model, the predicted data with 7 wt.%, 10 wt.%, and 12 wt.% Si were
extracted respectively and compared with the experimental results, as shown in Figure 7. The results
indicated that all the experimental data locates in the region of the predicted results, inferring that the
model can not only describe the properties of the existing experimental data very well but also predict the
mechanical properties of the alloys over the unknown region(s).

Alloy design with desired mechanical properties
Based on the accurate prediction results over wide composition and process parameter space, one can
design the SLMed Al-Si-(Mg) alloy with desired machinal properties. However, how to design an alloy with
simultaneously high strength and ductility needs to be considered. The multi-objective to single-objective
optimization strategy, which converts a weighted linear combination of multi-objective performance into a
[82]
single objective for optimizing the design for a single objective quantity (Ashby’s method ), is one of the