Session: 12-05-02: Data-Enabled Predictive Modeling, Scientific Machine Learning, and Uncertainty Quantification in Computational Mechanics

Paper Number: 97070

97070 - Cnn-Based Surrogate for the Phase Field Damage Model: A Study on Its Generalization Across Microstructure Parameters for Composite Materials 

The phase field damage model within the finite element framework has been widely used to predict fracture in homogeneous quasi-brittle materials; however, it is computationally expensive for predicting fracture in heterogeneous composite material microstructures. While computational multi-scale mechanics models are generally popular for evaluating composite material performance for a given microstructure configuration, the need for optimizing design-space microstructure parameters has led researchers to pursue computationally inexpensive surrogate models based on micromechanics and, more recently, machine learning approaches. While machine learning approaches are more matured in the fields of computer vision, natural language processing, recommendation algorithms, their applications in the field of computational solid and fluid mechanics is relatively recent. Advances in scientific machine learning approaches and studies focusing on their generalization are necessary to replace established numerical methods, such as the finite element method, to address complex problems in computational mechanics.

In this talk, we present a study assessing the generalization of a convolutional neural network (CNN) based surrogate for the phase field model to predict the 2D damage field image at maximum/peak load, given the image of an arbitrary 2D microstructure of a unidirectional fiber-reinforced composite. First, we will briefly discuss the procedure to generate training and test data from synthetic microstructures spanning the microstructure parameters of volume fraction and fiber radius, using the phase field damage model and the finite element method. Next, we will describe the image-to-image CNN-based surrogate model to predict the 2D damage field and an alternative CNN-based regression model to predict peak load, respectively. Our main finding is that a direct prediction of peak load from the microstructure image using the regression model is not viable. Instead, the indirect prediction of the peak load by correlating the integral of the 2D damage field obtained from the surrogate model seems viable. Finally, we will describe a few case studies to assess the capability of the surrogate model to predict damage and peak load and to interpolate over fiber radii and volume fraction, and accuracy metrics to assess the generalization of the surrogate model when using different training datasets. Through these comprehensive studies involving training and testing with datasets spanning the microstructure parameters (i.e., fiber radii and volume fraction), we find that the surrogate model is able to accurately predict the ranking of peak load and damage integral in 80% of the cases, and the predict peak load is generally with  of the least squares fit values. Therefore, we believe that the surrogate model generalizes reasonably well and can be useful in fast screening of composite microstructures to identify optimal design-space parameters.

Presenting Author: Ravindra Duddu Vanderbilt University

Presenting Author Biography: Ravindra Duddu got his B. Tech in Civil Engineering from the Indian Institute of Technology Madras. He obtained his M.S. and Ph.D. in Civil and Environmental Engineering from Northwestern University, respectively. After that he worked as a postdoctoral research at the University of Texas at Austin Institute for Geophysics and Columbia University in the City of New York. Currently, he is an Associate Professor of Civil and Environmental Engineering, Mechanical Engineering, and Earth and Environmental Sciences at Vanderbilt University. His research interests are in the general area of computational solid mechanics with an emphasis on fracture mechanics and multi-physics modeling of material damage evolution. Specific application interests include: fracture of glacier ice and sea ice, delamination of fiber reinforced composites, and corrosion of metal alloys. He is a member of American Society of Civil Engineers, American Society of Mechanical Engineers, American Geophysical Union, and United States Association for Computational Mechanics.

Authors:

Yuxiang Gao Vanderbilt University
Matthew Berger Vanderbilt University
Ravindra Duddu Vanderbilt University