PSI - Issue 71

Akash S.S. et al. / Procedia Structural Integrity 71 (2025) 180–187

185

hyperparameters of the neural network architecture. The optimizer used is Adam. Mean squared error is chosen as the loss function and Mean Absolute error as the metric. On comparing the optimal models generated for each K fold validation dataset, one of them was selected based on the Mean Absolute Error. This neural network architecture is shown in Figure 3. The Keras Tuner performs multiple iterations with di ff erent hyperparameters and finds the optimal values. The trained ANN model has a mean absolute percentage error of 0.03 % on validating using thetestdataset. Thismodelisused for predicting uncertainty in the material properties.

Input Units: 3 Activation: relu

Units: 128 (+118 more) Activation: relu

Units: 384 (+374 more) Activation: relu

Units: 512 (+502 more) Activation: relu

Output Units: 4 Activation: linear

Fig. 3: Neural network architecture used in the model

5. Image processing for detecting fiber diameter distribution from microstructural images of Unidirectional composites The fiber diameters in unidirectional composites can be determined by analyzing the microstructure images ofthe material. Techniques like Synchrotron microtomography and computer tomography (CT) can be utilized for the acquisition of microstructure images. Microstructure image shown in Figure 4(a) is an example of such image obtained using synchrotron microtomography technique. This is a portion of a reconstructed µ CT slice of the Carbon Fibre- reinforced polymer (CFRP) obtained by G. Requena et al. (2009) using a resolution of (0.7 µ m) 3 / voxel (700x700 pixels / 490x490 µ m 2 ) looking onto the plane perpendicular to the fibres. G. Requena et al. (2009) had used CFRP made up of T-300-3k carbon fibers with a mean diameter of 7 µ m and epoxy matrix at a fibre volume fraction of 55 vol%. The ANN model was trained for this material in this case study. This image will be used to validate the accuracyof the model for a new set of fiber diameters.

(a) (b) Fig. 4: (a) Portion of a reconstructed µ CT slice of the CFRP using a resolution of (0.7 µ m) 3 / voxel (700x700 pixels / 490x490 µ m 2 ) looking onto theplane perpendicular to the fibres [Source: G. Requena et al. (2009)] (b) PDF of Fiber diameter distributions in the sub-images The microstructure image shown in Figure 4(a) was split into 32 sub-images. Image processing techniques were applied to detect the fibers and determine the fiber diameter distributions from the images. Detection of