A Generic Framework for Multiscale Modeling in Computational Mechanics

Crone, Joshua (DEVCOM Army Research Laboratory)

Co-Authors:
Wilson, Zachary
Leiter, Kenneth
Becker, Richard
Knap, Jaroslaw

Category:
Computational Structure Mechanics

Hierarchical multiscale modeling (HMM) is an approach to developing high-fidelity material models by combining various "at-scale models" that capture salient phenomena at disparate spatial and/or temporal scales. HMM can provide reliable models for novel materials where experimental data is limited or nonexistent. HMM may also be the only choice when the structure-property relationships of a material are too complex to be captured analytically or empirically. However, integration of disparate at-scale models into robust finite element analysis (FEA) tools is time consuming and labor intensive. Furthermore, the resulting material models are computationally expensive, if not intractable.

DEVCOM Army Research Laboratory (ARL) has developed a Hierarchical MultiScale (HMS) framework with an Abaqus Vectorized User MATerial (VUMAT) interface to significantly simplify the labor-intensive process of developing and integrating multiscale models. This is accomplished by abstracting out many of the computational science tasks ubiquitous across HMM, such as scheduling, load balancing, and monitoring of at-scale evaluations, as well as communication between scales. The framework efficiently utilizes high performance computing (HPC) resources and machine learning (ML) to reduce computational costs. The VUMAT interface ensures seamless portability of multiscale models across any commercial or government FEA tool that supports the VUMAT interface.

In this presentation, we will provide an overview of the HMS-VUMAT capabilities and demonstrate the ease in developing new multiscale material models. We will highlight the flexibility of the framework through applications ranging from architected materials and polycrystalline metals to soils and energetics. Finally, we will demonstrate the tremendous computational saving provided by the ML-based surrogate models, which can efficiently predict the response of lower scale models at points in the input space that are the same as, or sufficiently close to, points that have already been sampled.