Department Seminar of Hadas Hochster Artificial Neural Networks as Surrogate Models for the PHFGMC Nonlinear Micromechanical and Failure Analysis

School of Mechanical Engineering Seminar
Wednesday, December 29, 2021 at 14:00   
Wolfson Building of Mechanical Engineering, Room 206

Hadas Hochster

Msc student of Prof. Rami Haj-Ali

Artificial Neural Networks as Surrogate Models for the PHFGMC Nonlinear Micromechanical and Failure Analysis

Multilayered composites are widely used in the aircraft industry due to their lightweight and high strength.  However, a major disadvantage of these composites is their complex multi‐mode interactive process of failure. In this study, discrete multi-axial failure points of IM7/977-3 unidirectional laminate are generated using the cohesive parametric high-fidelity-generalized-method-of-cells (Cohesive-PHFGMC) micromechanical model. In general, cohesive elements or surfaces are used to model their adhesion and separation according to mixed-mode traction-separation law.  Therefore, the progressive local-global failure initiation and propagation are fully captured.  In the Cohesive-PHFGMC model, cohesive elements are discretely embedded in the matrix phase allowing the damage growth in arbitrary un-specified paths.  The model calibration is based on test data performed for unidirectional coupons. The calibrated Cohesive-PHFGMC model is used to simulate failure under bi-axial loads using refined unit-cells.  The simulated failure points are used to generate continuous failure envelopes for the composite under general multi-axial loading. Predicted failure envelopes are compared to bi-axial experimental results from the literature.  Good results are shown for the new simulated failure envelopes that can be used as an alternative to current composite failure theories and can capture additional modes of failure.

The second part of this study includes using Artificial Neural Networks (ANNs) to capture the multi-axial effective stress-strain responses for the composite based on pre-simulations of finite strain paths by the PHFGMC micromechanical model. The new ANN are examined in their ability to predict the PHFGMC response for stress-strain paths that were not used in the training process. The trained surrogate ANN models are later integrated as material models in the Abaqus commercial FE explicit code.  Low-Velocity Impact (LVI) analyses were conducted using the integrated ANN with 3D-shell layer-by-layer elements. The results are compared to those using standard nonlinear material models.  It is demonstrated that the newly proposed ANN surrogate modeling can be an efficient computational tool for future multi-scale analysis with highly refined models at both the micro and macro scales.

Join Zoom Meetin https://us02web.zoom.us/j/82108132163?pwd=Z2h4UzNzUS9mbXplT0lMU1pZenFEQT09