guest ::
TY - THES AU - Danesh, Hooman TI - Computational homogenization and data-driven surrogate modeling of mechanical metamaterials PB - Rheinisch-Westfälische Technische Hochschule Aachen VL - Dissertation CY - Aachen M1 - RWTH-2025-10667 SP - 1 Online-Ressource : Illustrationen PY - 2025 N1 - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University 2026 N1 - Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2025 AB - Mechanical metamaterials provide unique opportunities to tailor material properties for specific engineering applications. However, their practical use is limited by the challenges of modeling their complex and multiscale behavior. Traditional homogenization methods are computationally expensive, limiting the efficiency of simulation, prediction, and design processes. This thesis is driven by the need to create efficient, accurate, and user-friendly computational tools that integrate physics-based homogenization with data-driven approaches to enable rapid property prediction, inverse design, and uncertainty quantification, ultimately promoting the widespread use of metamaterials in engineering design. The present thesis develops an integrated framework that enhances computational efficiency and accessibility in metamaterial design. It introduces a two-scale homogenization approach using truss elements to model elastoplastic truss-based lattices, significantly reducing computational costs while maintaining accuracy. Additionally, it leverages fast Fourier transform (FFT)-based homogenization to efficiently generate datasets for metamaterial unit cells, enabling data-driven surrogate modeling. With their real-time predictive capabilities, the framework integrates the established surrogate models into a user-friendly graphical interface that supports property prediction and inverse design of auxetic metamaterials, making exploration and design accessible to both experts and non-experts. Finally, transitioning from deterministic to probabilistic machine learning approaches, the thesis establishes uncertainty-aware and data-efficient reduced-order structure-property linkages for stochastic metamaterials. This cumulative thesis compiles the publications of the author (and his co-authors) on physics-based computational homogenization and data-driven surrogate modeling of various metamaterial classes, including truss-based, auxetic, and stochastic structures. Following an exploration of the motivation, a review of the current state of the art, and addressing open gaps and research questions, four published articles are presented. The first article investigates the limitations of computational homogenization approaches, particularly FE<sup>2</sup> and FFT-based methods, addressing issues such as the lack of scale separation in 3D-printed structures and infinite stiffness contrasts that impede conventional FFT solvers. The second article develops a nonlinear two-scale homogenization framework for elastoplastic truss-based lattices, utilizing truss elements at the microscale with combined nonlinear exponential isotropic and linear kinematic hardening laws, accurately capturing elastic and plastic responses across various loading conditions, validated against direct numerical simulations. The third article focuses on data-driven surrogate modeling for auxetic metamaterials with orthogonal voids, using FFT-based homogenization for efficient dataset generation and training surrogate models to predict elastic properties in real time while supporting inverse design through an intuitive graphical interface. The fourth article explores probabilistic machine learning for stochastic metamaterials, employing statistical correlation functions and principal component analysis (PCA) for dimensionality reduction, Gaussian process regression (GPR) for robust property prediction with uncertainty quantification, and an active learning strategy to minimize training data, enhancing efficiency for random unit cell designs. LB - PUB:(DE-HGF)11 DO - DOI:10.18154/RWTH-2025-10667 UR - https://publications.rwth-aachen.de/record/1023683 ER -