Page 46 - IJAMD-2-3

International Journal of AI for
Materials and Design Biomimetic ML for AFSD aluminum properties

A Table 4. Optimal hyperparameters for predicting von Mises
stress in additive friction stir deposited aluminum‑based
walled structures
Algorithms Best max Best min Best min Best N
depth samples split samples leaf estimators
GA-DT 10 2 1 -
GA-RF 5 3 1 93

B Table 5. Optimal hyperparameters for predicting
logarithmic strain in additive friction stir deposited
aluminum‑based walled structures
Algorithms Best max Best min Best min Best N
depth samples split samples leaf estimators
GA-DT 5 2 1 -
GA-RF 11 2 1 23

Table 4, it is observed that the GA-DT model favors a
deeper tree with a maximum depth of 10, while the GA-RF
Figure 6. Convergence curves for GA-DT and GA-RF models in model uses shallower trees with a maximum depth of 5.
predicting von Mises stress in the additive friction stir deposition process. This suggests that for von Mises stress, the DT benefits
(A) GA-DT, (B) GA-RF. Both models show rapid improvement in initial
generations, with GA-RF exhibiting slightly better convergence in later from more complex decision paths, while the RF achieves
stages. better results with simpler individual trees, leveraging
Abbreviations: GA-DT: Genetic algorithm-decision tree; GA-RF: Genetic the power of ensemble learning. Both models prefer a
algorithm-random forest. small minimum number of samples to split an internal
node (2 for GA-DT, 3 for GA-RF) and the minimum
A possible number of samples at a leaf node (1 for both),
indicating that fine-grained decision-making enhances
model performance. The GA-RF model uses 93 estimators
(trees), a relatively high number, implying that ensemble
diversity significantly contributes to its predictive power
for von Mises stress. Table 5 shows a reversed trend for
tree depth in the context of logarithmic strain prediction:
the GA-DT model uses shallower trees (depth of 5), while
the GA-RF model uses deeper trees (depth of 11). This
B implies that logarithmic strain prediction benefits from
different model architectures compared to von Mises
stress prediction. Nevertheless, both models maintain
a preference for a small minimum number of samples
to split (2) and at leaf nodes (1), consistent with the von
Mises stress prediction. Notably, the GA-RF model uses
fewer estimators (23) for logarithmic strain prediction
compared to von Mises stress, suggesting that fewer but
more complex trees are more effective for this particular
prediction task.
Figure 7. Convergence curves for GA-DT and GA-RF models in Tables 6 and 7, along with Figures 8 and 9, provide a
predicting logarithmic strain. (A) GA-DT, (B) GA-RF. The convergence comprehensive overview of the performance metrics for
process shows that predicting strain is more challenging, with erratic the GA-DT and GA-RF models in predicting von Mises
patterns in the GA-RF model compared to von Mises stress predictions.
Abbreviations: GA-DT: Genetic algorithm-decision tree; GA-RF: Genetic stress and logarithmic strain, respectively, for AFSD
algorithm-random forest. aluminum-based walled structures.

Volume 2 Issue 3 (2025) 40 doi: 10.36922/ijamd.5014