Page 119 - MSAM-4-3

Materials Science in Additive Manufacturing Sustainable manufacturing composite material optimization

while the remaining 22% may be attributed to factors To address the trade-off between mechanical
such as microstructural inconsistency, filament quality performance and energy consumption, an AI-MCDM
variation, or limitations in the modeling process. The approach using Fuzzy AHP-TOPSIS was employed to
MAE was 0.87 MPa, indicating that predictions by AI identify the optimal printing parameters. The AI model
models differed from experimental results by an average of utilized historical FDM process data to predict the
0.87 MPa. The MAPE was 2.1%, indicating a relatively low optimal layer thickness, infill density, and shell thickness,
average percentage error and demonstrating the model’s prioritizing mechanical strength over energy efficiency
strong predictive performance. These results highlight based on predefined objective weightings. Table 3 presents
the potential of the AI model in predicting mechanical the AI-MCDM-optimized parameters of FDM for impeller
properties within FDM, while also highlighting that production, balancing mechanical performance and
additional training data, sensor feedback, and parameter energy efficiency under different objective weightings. The
tuning for RL could further reduce prediction errors and Fuzzy AHP method was used to determine the relative
enhance the overall optimization accuracy toward real- importance of criteria under two optimization scenarios:
time adaptive control of FDM process parameters. (i) mechanical performance (60%) versus energy efficiency
The precision of AI model predictions for flexural (40%); and (ii) energy efficiency (60%) versus mechanical
strength, compression strength, and wear resistance performance (40%). In (i), where mechanical performance
was determined using scatter plots and standard error was prioritized, the optimal parameters were 0.2 mm layer
metrics. For flexural strength, the model demonstrated thickness, 70% infill density, and 1.2 mm shell thickness,
moderate accuracy with an RMSE of 2.57 MPa, R of resulting in a tensile strength of 41.5 MPa and energy
2
0.68, MAE of 1.77 MPa, and MAPE of 4.06%. These consumption of 145 Wh. In (ii), where energy efficiency
results indicate that while the model generally followed was emphasized, the model suggested 0.3 mm layer
experimental trends, small variations were present, thickness, 50% infill density, and 1.0 mm shell thickness,
likely due to material anisotropy and interlayer bonding with a resultant tensile strength of 38.0 MPa and reduced
differences in FDM (Figure 3). In contrast, the prediction energy consumption of 125 Wh. These results demonstrate
of compression strength displayed high accuracy, with the capability of the AI-MCDM approach to effectively
an RMSE of 1.37 MPa, R of 0.95, MAE of 1.33 MPa, and balance strength, wear resistance, and sustainability in
2
MAPE of 2.38%. These values suggest that the trained FDM-based impeller production.
model closely replicated the load-carrying behavior of Table 4 presents a comparison between AI-MCDM-
the material under compression (Figure 4). For wear optimized parameter selection and conventional FDM
resistance, the model had slightly lower predictive parameter selection. The analysis revealed that the
accuracy, with an RMSE of 0.048 mm /N·m, R of 0.84, AI-based approach outperformed manually selected
3
2
MAE of 0.037 mm /N·m, and MAPE of 5.17%. This parameters. Specifically, AI-optimized parameters
3
decrease in performance may be attributed to the improved tensile strength by 7% and wear resistance by
complex nature of tribological interactions and surface 11%, while maintaining only a moderate increase in energy
wear mechanisms (Figure 5). Across all tests, scatter consumption of 3.5%. In contrast, conventional parameter
plots revealed that AI-optimized parameter selection selection, in accordance with standard FDM protocols,
effectively predicted mechanical properties, reducing resulted in suboptimal mechanical performance and
reliance on trial-and-error methods in FDM-based inefficient material deposition, resulting in unnecessary
impeller production. However, further improvements, energy consumption. The use of the Fuzzy AHP-TOPSIS
such as incorporating real-time sensor feedback, enlarging ranking technique enabled optimal selection of process
the training dataset, and applying RL for adaptive control, parameters from real-time measurements, effectively
can enhance the model’s accuracy, particularly in wear reducing the need for trial-and-error experimentation. In
resistance predictions where surface roughness and addition, the AI model continuously refined its predictions
frictional force introduce high variability. based on sensor feedback from the 3D printer, thereby

Table 3. AI‑MCDM‑optimized FDM parameters for impeller production

Objective weighting (%) Layer thickness Infill Shell thickness Tensile Energy Wear rate
(mm) density (%) (mm) strength (MPa) consumption (Wh) (mm /N·m)
3
Mechanical (60); energy (40) 0.2 70 1.2 41.5 145 0.55
Mechanical (40); energy (60) 0.3 50 1.0 38.0 125 0.61
Abbreviations: AI-MCDM: Artificial Intelligence–multicriteria decision-making; FDM: Fused deposition modeling.

Volume 4 Issue 3 (2025) 11 doi: 10.36922/MSAM025200033