Conference EMMC20
Machine learning-based multiscale fracture modelling
Prof. Julien Yvonnet
Université Gustave Eiffel, MSME, 5 Bd Descartes, F-77454 Marne-la-Vallée Cedex 2, France
- webpage
In this lecture, we present recent work on multi-scale modeling of anisotropic fracture in quasi-fragile materials. Anisotropic fracture can be induced by microstructures where specific orientations occur, as in composites, architected or natural materials (bone, wood). Predicting the macroscopic fracture behavior of these materials from knowledge of their microstructure (Representative Volume Element - RVE) alone is a difficult task, but one that could be extremely useful to engineers for virtual testing and design. Two machine learning-based strategies are presented to solve this problem. Firstly, an acceleration of FE2-type simulations, involving nested two-scale nonlinear simulations, is achieved by means of a clustering technique. The method is based on a machine-learning type method for classification, where integration points of the macroscopic mesh are clustered to detect and avoid self-similar nonlinear problems to be solved on the attached RVEs. Secondly, an anisotropic damage model called the DDHAD (Data-Driven Harmonic Analysis of Damage) model is built sequentially on the basis of fracture simulations within the RVE and using harmonic damage analysis. An optimal number of macroscopic internal variables is defined as the reduced coefficients related to the spherical harmonic expansion terms of the orientation-dependent scalar damage functions. A machine learning strategy is then developed to learn the evolution laws of these macro internal variables through off-line cracking simulations on the RVE. Applications to highly anisotropic composite and architected materials are presented.