Xin Liu2,Wenbin Yu1
Purdue University1,The University of Texas at Arlington2
Xin Liu2,Wenbin Yu1
Purdue University1,The University of Texas at Arlington2
Emerging artificial intelligence (AI) and machine learning (ML) techniques are changing the computational paradigm on the development of lightweight, high performance composite materials and structures. This presentation will introduce our recent works of ML-assisted modeling of composites. The presentation has two parts. The first part is the application of artificial neural network (ANN) models to reduce the computational costs of high-fidelity simulations. In this application, ANN models are trained to be efficient surrogate models for computationally intensive jobs such as multiscale modeling of textile composites, defect evaluation of non-crimp fabric (NCF) composites with fiber misalignment, and structural analysis of cylindrical shell buckling. Different physical constraints are imposed during the training to reduce the physically inconsistent predictions, minimize the required training data size, and customize the solutions with engineering practices. The second application is to learn unknown physics (e.g., constitutive model) with limited experimental data. A novel hybrid finite element-neural network computational framework is developed to enable training an ANN model with indirectly measurable data. This computational framework removes the requirements of training paired data and therefore many previously unlearnable physical laws can be learned within this computational framework. Moreover, the ANN models are fully constrained by known physics and therefore avoid physically inconsistent predictions. This computational framework has been applied to discover failure criteria of composites, nonlinear constitutive laws, and damage accumulation using limited measurable data.