Harnessing deep learning for physics-informed prediction of composite strength with microstructural uncertainties

Document Type

Article

Publication Date

9-2021

Department

Department of Mechanical Engineering-Engineering Mechanics

Abstract

Representative volume elements (RVEs) are commonly utilized to analyze the effective properties of fiber reinforced composites based on their repetitive microstructures and the constituent fiber and matrix properties. Intrinsically, the randomness of fiber distribution exists in composites even though the manufacturing process is strictly controlled. Such microstructural uncertainties essentially render the composite strength stochastic and difficult to characterize. In this research, a physics-informed deep learning framework is developed to analyze the variation of the strength of composite material with microstructural uncertainties. A random fiber packing algorithm is employed to sample the RVE images that are subsequently subjected to composite progressive damage analysis using the finite element method. The input–output relations acquired from this first-principle analysis are used as training data to facilitate deep learning that is capable of directly predicting the composite strength based on the RVE image. Two neural network architectures, a customized convolutional neural network (CNN) and a VGG16 transfer learning neural network, are established, with the view to unleashing the power of deep learning with small data size. This new framework significantly expedites the uncertainty analysis. It can directly take the spatial uncertainties in RVEs into account, outperforming other uncertainty quantification approaches. Systematic case investigations are conducted, in which the statistical cross-validation confirms the validity of the method. Owing to the highly efficient emulation, one can further carry out convergence analysis for uncertainty quantification. The results clearly demonstrate the effectiveness and capability of the proposed new framework for composite strength prediction. This framework is generic, which can be potentially extended into uncertainty quantification of other composite properties.

Publication Title

Computational Materials Science

Share

COinS