Adaptive and Parallel Multiscale Framework for Modeling Cohesive Failure in Engineering Scale Systems

TL;DR

This work addresses the high computational cost of multiscale modeling for cohesive interfaces in large-scale engineering systems by introducing an adaptive framework that selects between two microscale models (FM and TM) via an offline SVR-based database and a dedicated parallel multiscale network (multiscale_net). The method leverages computational homogenization with Hill-Mandel energy equivalence to couple macro and micro scales across curved interfaces, while dynamically distributing computation across HPC resources. It is verified on a curved double-cantilever beam and validated on the National Rotor Testbed blade, achieving high accuracy (differences of a few percent) and substantial HPC speedups, demonstrating practical scalability for industrial problems. The results suggest that expanding the offline model library and incorporating additional reduced-order or phenomenological models could further boost performance without sacrificing predictive fidelity.

Abstract

The high computational demands of multiscale modeling necessitate advanced parallel and adaptive strategies. To address this challenge, we introduce an adaptive method that utilizes two microscale models based on an offline database for multiscale modeling of curved interfaces (e.g., adhesive layers). This database employs nonlinear classifiers, developed using Support Vector Machines from microscale sampling data, as a preprocessing step for multiscale simulations. Next, we develop a new parallel network library that enables seamless model selection with customized communication layers, ensuring scalability in parallel computing environments. The correctness and effectiveness of the hierarchically parallel solver are verified on a crack propagation problem within the curved adhesive layer. Finally, we predict the ultimate bending moment and adhesive layer failure of a wind turbine blade and validate the solver on a difficult large-scale engineering problem.

Figures (21)

• Figure 1: National Rotor Testbed 13 m wind turbine blade. (from Murray et al. murray2019fusion)
• Figure 2: Kinematics of multiscale cohesive modeling in a finite strain setting. $\Omega_0^\pm$ are the macroscale domains, while $\Theta_0$ represents the microscale domain.
• Figure 3: Dependency trees of the adaptive multiscale network and parallel communication patterns.
• Figure 4: Schematic of the workload balance through inter-domain communication for computational jobs. In this example, five jobs are processed by three servers between time steps $t_n$ and $t_{n+1}$.
• Figure 5: Schematics of the multiscale DCB test. (a) Dimensions and boundary conditions, (b) RUC at the adhesive layers.