Large Parallel Adaptive Multiscale Modeling of Interfaces

The computationally demanding nature of multiscale modeling (e.g., Computational Homogenization (CH)) necessitates the development of parallel and adaptive strategies for industrial-scale applications. Accordingly, this dissertation introduces an adaptive and parallel multiscale strategy for interface modeling. It adopts an adaptive approach, selecting between two microscale models through an offline database. This database utilizes nonlinear classifiers based on Support Vector Regression (SVR) constructed from microscale sampling data to serve as a preprocessing step for multiscale simulations. A co-designed parallel network library, facilitating seamless model selection, integrates tailored communication layers to ensure scalability, which is essential in parallel computing environments. This work presents a novel multiscale solver capable of executing high-fidelity, large-scale engineering simulations. The implementation of the solver is verified and validated through the application, demonstrating its ability to capture the physics observed in experimental data at both macro and micro scales. This is illustrated through the analysis of the failure of a large wind turbine blade.

多尺度建模（例如计算均质化（Computational Homogenization，CH））因其计算复杂度极高，亟需面向工业级应用开发并行自适应策略。据此，本论文提出了一种面向界面建模的自适应并行多尺度策略。该策略采用自适应方法，通过离线数据库在两种微观尺度模型间进行选择。该数据库依托由微观尺度采样数据构建的支持向量回归（Support Vector Regression，SVR）非线性分类器，作为多尺度模拟的预处理环节。本研究协同设计了一款并行网络库，可实现模型选择的无缝衔接，其集成了定制化通信层以保障可扩展性——这在并行计算环境中至关重要。本研究提出了一款新型多尺度求解器，可实现高保真、大规模工程模拟。通过应用实例对该求解器的实现进行了验证与确认，结果表明其能够准确捕捉宏观与微观尺度下实验数据所反映的物理规律。该验证过程可通过大型风力机叶片失效分析实例予以展示。