School of Mechanical Engineering Seminar
Monday November 27.11.2023
Wolfson Building of Mechanical Engineering, Room 206
Low-Velocity Impact Analysis of Laminated Composites: Multi-Scale Nonlinear Micromechanics and Surrogate Neural Networks
Or Benson
M.Sc. research under the supervision of Prof. Rami Haj-Ali Tel-Aviv University, Department of Mechanical Engineering
Low-velocity Impact (LVI) of laminated composite structures can lead to severe growing hidden damage states. This study introduces a new multi-scale framework for the mechanical progressive damage analysis of laminated composite structures during and post-LVI. The central hypothesis is that refined nonlinear micromechanical models can be integrated into concurrent progressive damage analysis and yield good prediction ability compared to present macro-damage homogenized anisotropic theories. To this end, we demonstrate that the Parametric High Fidelity Generalized Method of Cells (PHFGMC) can provide refined predictive nonlinear micromechanical behavior of unidirectional and woven laminates. The PHFGMC can solve for both the local and overall nonlinear and damage responses in heterogeneous multiphase composites. Implementation of PHFGMC involves imposing periodicity conditions within a repeating unit cell (RUC), divided into subcells representing distinct fiber and matrix phases. Maintaining continuity in average traction and displacement across these subcells ensures equilibrium, enhancing computational efficiency compared to conventional Finite-element (FE) methods.
Massive computational effort is needed to integrate the progressive-damage PHFGMC model in multi-scale LVI analysis of laminated structures. This multiscale refined framework will require repeated computations for the RUC responses at thousands (and often more) integration points at all time increments. In order to avoid the massive (if not possible) computational effort, a new AI-ANN approach is introduced whereby different ANN classes are trained off-line for different nonlinear PHFGMC simulations for one RUC under different discrete potential mechanical stress-strain and histories or loading paths. The trained ANNs are examined in their ability to accurately predict the PHFGMC responses for cases not preview-to during the training. Incorporating the trained PHFGMC-ANN surrogate micromechanics is then used for the LVI simulations to allow for accurate predictions of local and global responses and to navigate the complex damage modes of the material that develop during impact events.
The IM7/972 and M21 laminated and woven carbon/epoxy composite material systems are considered as these are two common advanced composites currently employed in aerospace structures. An extensive overview of the material system is provided, introducing PHFGMC-ANN in modeling nonlinear behavior and damage. The seminar will discuss different selections of ANN models and their architecture. The research introduces the application of Long Short-Term Memory (LSTM) ANNs trained by the PHFGMC for composite material modeling. A foundational understanding of LSTM is provided, outlining its seamless integration process, encompassing data collection, sequence encoding, LSTM architecture, training, prediction, and analysis. The thesis emphasizes the benefits of this innovative integration approach and how it can be further expanded and implemented in multi-scale analysis.
https://tau-ac-il.zoom.us/j/86497933118
