Perpustakaan Digital ITB

This study develops a machine learning model to predict the elastic properties of unidirectional fiber-reinforced composite materials using microstructural images of the composite. Traditionally, the properties of these composites are analyzed using micromechanics-based finite element methods, which are time-consuming and resource-intensive. The deep learning approach was chosen for its faster processing time, self-learning capabilities, and potential for high accuracy. This research implements Resnet (residual network) 101 that consist of 101 convolutional layers for stable but complex prediction results. Multiple random fiber RVE is generated using python script and its elastic properties is calculated using finite element micromechanics. The RVE and its elastic properties data is then used as a dataset to train and validate the machine learning model. Results indicated that the deep learning model provided accurate predictions with only minor differences from simulation outputs. The runtime for machine learning predictions was significantly shorter than for FEM simulations. Each elastic property model had distinct characteristics, with longitudinal property being the most accurate, followed by transverse, and shear properties. It is found that the more nonlinearity present in the properties the more difficult it is for the ML model to gain accurate predictions.