INTERPRETABLE MACHINE LEARNING FOR GEOMETRICAL VARIABLE SENSITIVITY ANALYSIS: APPLICATION TO AXIALLY COMPRESSED CYLINDRICAL BUCKLING

• Main Author: Geluntung Wafi, Muhammad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75151
• Institution: Institut Teknologi Bandung

<p align="justify"> For over a century of thin cylindrical shells buckling deterministic solution known has a diverge result to the experiment. The gap between classical buckling theory and experimental results drove the study to discover the reason behind this discrepancy. Further, the classical buckling theory might not fully capture the true complexity of the relationship between material properties and geometrical parameters with the critical buckling load. To that end, this final project combines finite element simulation and explainable machine learning to uncover such a relationship. This final project employs Gaussian Process Regression (GPR) with six input variables, namely, Young’s modulus, Poisson ratio, shell length, shell thickness, shell radius, and FTQC (Fabrication Tolerance Quality Class). The model achieved good prediction accuracy with ? 3% error averaged to the whole model and can continue for another step. Shapley additive explanation (SHAP) was introduced to explain the GPR model by letting the user examine every variable of the model and their interaction with another variable. The thickness of the cylinder was found to have a significant impact on the model, followed by Young’s modulus, shell radius, FTQC, shell length, and Poisson ratio. The outcome shows several aberrations to the classical theory, especially in how the variables interact and depart from linearity.