Overview:
Our goal is to develop accurate and rapid computational models for predicting complex behaviors of advanced materials and structures with the help of artificial intelligence (AI) techniques. We are developing data-driven multiscale models to build an entire processing-microstructure-property-performance relationship. The research outcomes will help develop an AI-assisted engineering design system to recognize component features, attributes, and intended performances to make recommendations about the directions and parameters for manufacturing processes, materials selections and final structural design. |
Multiscale analysis of textile composites
Mechanical behaviors such as elastic, thermoelastic, and viscoelastic textile composites involving multiple scales and complex geometries can be challenging to predict. Mechanics of structure genome (MSG) has been extended to accurately capture the above mechanical behaviors. Compared with traditional representative volume element (RVE) analysis using 3D finite element analysis (FEA), the MSG-based models provide same accuracy at a fraction of computing costs. In addition, MSG models unifies micromechanics and structural mechanics so that the unnecessary scale separation assumptions can be removed for modeling thin and slender composite structures.
The developed functions have been implemented into an open source code in the cloud called TexGen4SC, which is freely accessible and can be launched through a web browser (https://cdmhub.org/tools/texgen4sc) without installation.
Mechanical behaviors such as elastic, thermoelastic, and viscoelastic textile composites involving multiple scales and complex geometries can be challenging to predict. Mechanics of structure genome (MSG) has been extended to accurately capture the above mechanical behaviors. Compared with traditional representative volume element (RVE) analysis using 3D finite element analysis (FEA), the MSG-based models provide same accuracy at a fraction of computing costs. In addition, MSG models unifies micromechanics and structural mechanics so that the unnecessary scale separation assumptions can be removed for modeling thin and slender composite structures.
The developed functions have been implemented into an open source code in the cloud called TexGen4SC, which is freely accessible and can be launched through a web browser (https://cdmhub.org/tools/texgen4sc) without installation.
Machine learning accelerated multiscale modeling
Taking advantages of the advanced machine learning models such as the state-of-the-art neural network model, a data-driven approach can accelerate multiscale analysis by replacing the expensive high-fidelity model with an efficient surrogate model. For example, a new failure criterion for fiber tows (i.e. yarns) is developed based on a micromechanical model using the mechanics of structure genome (MSG) and a deep learning neural network model. This failure criterion can be applied to yarns in mesoscale textile composites modeling while capturing the failure initiation at the fiber and matrix level.
Taking advantages of the advanced machine learning models such as the state-of-the-art neural network model, a data-driven approach can accelerate multiscale analysis by replacing the expensive high-fidelity model with an efficient surrogate model. For example, a new failure criterion for fiber tows (i.e. yarns) is developed based on a micromechanical model using the mechanics of structure genome (MSG) and a deep learning neural network model. This failure criterion can be applied to yarns in mesoscale textile composites modeling while capturing the failure initiation at the fiber and matrix level.
Multiscale machine learning discovered constitutive laws
In addition to constructing surrogate models to speed up the simulations, advanced neural networks are also good at approximating unknown physics thank to the universal approximation theorem. For multiscale modeling of advanced materials, a key question is that we don't know the nonlinear constitutive laws of constituents at the fine scales and it is very difficult to measure such laws from experiments. For example, what are the constitutive laws of yarns? what are the transverse properties of fibers? These are all the real questions to be answered. The critical challenge is that the data can be measured is usually at the structural level such as load-deflection curves while the data we want to use for neural networks is strain and stress in a constituent at the fine scale. This information mismatch can be solved by coupling neural networks into multiscale modeling.
In addition to constructing surrogate models to speed up the simulations, advanced neural networks are also good at approximating unknown physics thank to the universal approximation theorem. For multiscale modeling of advanced materials, a key question is that we don't know the nonlinear constitutive laws of constituents at the fine scales and it is very difficult to measure such laws from experiments. For example, what are the constitutive laws of yarns? what are the transverse properties of fibers? These are all the real questions to be answered. The critical challenge is that the data can be measured is usually at the structural level such as load-deflection curves while the data we want to use for neural networks is strain and stress in a constituent at the fine scale. This information mismatch can be solved by coupling neural networks into multiscale modeling.