Design of Artificial Neural Networks for Damage Estimation of Composite Laminates: Application to Delamination Failures in Ply Drops

• Published in: Composite structures Vol. 304; no. 1; p. 116320
• Main Authors: Mendoza, Arturo, Friderikos, Orestis, Trullo, Roger, Baranger, Emmanuel
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.01.2023; Elsevier
• Subjects: Artificial Neural Networks; Computer Aided Engineering; Computer Science; Damage mechanics; Engineering Sciences; Manifold learning; Mechanics; Mechanics of materials; Neural and Evolutionary Computing; Neural architecture learning; Nonlinear finite element analysis; Damage mechanics; Nonlinear finite element analysis; Manifold learning; Artificial Neural Networks; Neural architecture learning; Artificial Neural Networks (ANN)
• ISSN: 0263-8223; 1879-1085
• DOI: 10.1016/j.compstruct.2022.116320

This work presents a data-driven approach based on Artificial Neural Networks (ANN), that benefits from parametric non-linear finite element analyses, in order to provide a “cheaper” numerical alternative to the more expensive experimental testing of advanced composite laminates. The chosen subject of study is the damage evolution and associated delamination failures of ply drops. As such, physical-based modeling of these phenomena are used as training data for obtaining the most suitable ANN capable of estimating the damage evolution using only the prescribed initial conditions (i.e., laminate geometry, material orientation, applied loading). Additionally, this work aims at providing first-hand experience into the procedure that leads to obtaining such ANN. In order to do so, we detail each of the required steps, such as choosing a network architecture, defining a custom loss function, as well as deciding on the better learning parameters. Each stage on this process is guided by intuitive statistical analyses and simple criteria, such as selecting the “simpler” model over whenever a more complex one “improves” on the performance but only by a narrow margin. Indeed, for every step, multiple numerical experiences are provided so that the reader can also get a deeper understanding on the inner-workings of these ANN models. With this goal in mind, we employ dimensionality reduction techniques (PCA and t-SNE) to also propose a geometrical visualization of the nonlinear transformations performed by the ANN. Finally, additional tests regarding the network ability to generalize to unseen data showed that the optimal well-trained ANN is accurate and robust enough for near real-time predictions of the various damage evolution patterns, and outperforms other data-driven methods under comparison.