Session: 02-06-01: Machine Learning in Structural Dynamics and Aeroelasticity

Paper Number: 177517

177517 - Uncertainty and Sensitivity Analysis of Two-Dimensional Hierarchical Honeycomb Material Design 

Hierarchical structures are commonly used in both natural and engineered systems.  Applying hierarchy in honeycomb materials leads to a unique class of mechanical materials with remarkably high stiffness, strength, and energy absorption per unit weight. These properties are desirable and well-suited for lightweight aerospace and automotive structures. However, they are challenging to design and optimize because high-fidelity simulations, i.e., the Finite Element Method (FEM), are extremely computationally demanding. As a result, it is infeasible to optimally tailor hierarchical honeycomb materials for engineering application. To address this issue, the present research proposes an explainable machine learning-based framework to quickly and accurately predict the mechanical properties of 2D hierarchical honeycombs, thereby mitigating the reliance on high-fidelity simulations. Nevertheless, the consistency of these predictions remains a concern due to the inherent uncertainty in machine learning models. To overcome this problem, a deep ensemble of a Multi-Layer Perceptron (MLP) model is developed to predict the relative density, stiffness, and strength of the designs from five key geometric parameters while quantifying the uncertainty associated with each prediction. This approach can effectively quantify the data-dependent aleatoric uncertainty and model-dependent epistemic uncertainty. Results show that the proposed models achieve less than 1%, 1.5%, and 5% mean relative error in predicting relative density, stiffness and strength, respectively, with relatively narrow confidence intervals. For comparison, a Gaussian Process (GP) regression model is implemented as a probabilistic baseline. A dual sensitivity analysis is conducted to uncover the underlying design principles. SHapley Additive exPlanations (SHAP) are applied to identify important predictive features and enhance model interpretability, reveal complex nonlinear relationships and interactions between features. Simultaneously, global sensitivity analysis using the Morris’ method and Sobol’s indices is also performed to measure the impact that each input parameter has on the outputs. The insights obtained from these analyses provide valuable guidance for design optimization by identifying the geometric features that most significantly influence relative density, stiffness, and strength.

Presenting Author: Md Amanullah Kabir Tonmoy University of South Carolina

Presenting Author Biography: I am 3rd-year PhD student in the Department of Mechanical Engineering at the University of South Carolina. My current research focus is on surrogate modeling and surrogate-based/assisted optimization of lightweight material design and optimization. I am also working on the data-driven modeling of multiphase flow simulations.