Page 116 - MSAM-4-3

Materials Science in Additive Manufacturing Sustainable manufacturing composite material optimization

(iii) Output layer: Four neurons corresponding to
predicted values of tensile strength, flexural strength,
compression strength, and wear rate; a linear activation
function was applied to accommodate the continuous
nature of the output values.

2.2.2. Dataset and training strategy
The model was trained on a dataset of over 500, comprising
both experimentally generated data and literature-reported
FDM process records. Each record included different
parameter combinations along with corresponding
mechanical test results. The dataset was divided into
80% for training and 20% for testing to ensure unbiased
performance evaluation. Input and output values were
normalized to enhance convergence. Figure 2. Scatter plot comparing artificial intelligence (AI)-predicted and
experimentally measured tensile strength. The blue “×” mark represents
2.2.3. Training parameters and tools an individual data point comparing predicted and actual results. The
red dashed line (y = x) denotes perfect prediction; data points aligning
Training was performed using the mean squared error with this line indicate accurate AI estimations, while deviations reflect
(MSE) loss function and the Adam optimizer, with a prediction errors. This comparison illustrates the effectiveness of the AI
learning rate of 0.001. The model was trained over 100 model in estimating mechanical performance
epochs with a batch size of 32, implemented in a Python-
based TensorFlow/Keras environment.

2.2.4. Model performance and validation
The performance of the trained ANN was assessed using
several metrics:
(i) Root MSE (RMSE) ranged between 0.85 – 2.57 for
different mechanical properties.
(ii) Mean absolute error (MAE) and mean absolute
percentage error (MAPE) indicated consistent and
reliable prediction consistency,
(iii) R reached up to 0.95 for compression strength,
2
reflecting an excellent model fit.
Scatter plots of experimental and predicted outcomes
(Figures 2-5) further support the accuracy and aptness
of the model for optimizing FDM parameters in impeller Figure 3. Scatter plot comparing artificial intelligence-predicted and
experimentally measured flexural strength
production.
2.3. MCDM framework decision-making. Fuzzy AHP was used to determine the
relative weights of various criteria based on expert opinion
The MCDM framework serves as the backbone for FDM and historical data. By integrating linguistic variables and
process optimization in impeller production, aiming to fuzzy logic, it mitigates uncertainty in the decision-making
balance mechanical performance, energy consumption, process and facilitates prioritization of key attributes, such
and sustainability. This framework utilized Fuzzy AHP
and TOPSIS for the sequential weighing and ranking of as tensile strength, flexural modulus, wear resistance,
different process parameters, such as layer height, infill and surface roughness, to ensure a holistic evaluation of
density, nozzle temperature, print speed, and cooling impeller performance. After the weighting phase, TOPSIS
rate – each of which influences the quality and strength of was used to rank different printing parameter settings
the 3D-printed impellers. Given the intricate trade-offs based on closeness to the ideal solution that optimizes
33
involved in impeller design – where mechanical strength performance and energy efficiency. The implementation
and durability must be maximized while minimizing of AI-based MCDM models significantly enhances the
material usage and energy consumption – a hybrid Fuzzy system’s predictive capabilities and supports real-time
AHP-TOPSIS model was employed to enable intelligent adjustments to the FDM setup through sensor feedback

Volume 4 Issue 3 (2025) 8 doi: 10.36922/MSAM025200033