Analysis and optimization of non-linear mechanical properties of structural composite materials using machine learning-based algorithms

Abstract

Owing to its heterogeneous microstructure, structural composite materials are well-known for their outstanding combination of mechanical properties, including strength and toughness, while also maintaining a lightweight nature. However, the complex microstructure of the composite materials results in complicated deformation and failure mechanisms, making it challenging to develop analytical formulas for their non-linear mechanical properties over time. Despite the significant demand for structural composites across various industries, the absence of analytical formula hinders their real-life application, as the optimal designing of the composite microstructure for higher performance is difficult to achieve. In this thesis, we propose machine learning-driven frameworks designed to assist in the analysis and optimization of the structural composite materials for higher non-linear mechanical properties. In the first part of this thesis, we propose an optimization framework that can be used for mechanical properties directly related to the performance of composites, such as toughness, strength, and lightweight characteristics. This study efficiently derives the optimal designs by applying Bayesian optimization to a nacre-inspired composite, a representative biomimetic composite material. In the second part of the thesis, we devise a method to determine the fatigue strength, a property directly related to the reliability of composites, more quickly than conventional fatigue tests. By utilizing this fatigue strength measurement method, we can accumulate fatigue strength data of materials more efficiently, making it possible to apply machine learning-based optimization algorithms to the reliability of materials as well.