## Determination of composite material finite width correction factors using machine learning strategiesPaper: icaf2023 Tracking Number 9 PPT: icaf2023 presentation Session: Session 1: Digital engineering I Room: Theatre room: plenary Session start: 11:10 Mon 26 Jun 2023 freedyuval@gmail.com Yuval FreedAffifliation: Israel Aerospace Industries Topics: - Advanced materials and innovative structural concepts (Genral Topics), - Digital Engineering (Genral Topics)
Abstract:
Aviation products made of composite materials are designed as damage tolerant. That is, it is assumed that the structure contains some sort of imperfection, either induced during manufacture, assembly or service, that may remain undetected throughout the entire lifetime of the aircraft. A common industry practice for determination of strength allowables is usage of open hole specimens (per ASTM D6484) that represent such imperfections. Strength allowables obtained from such standard specimens strongly depend upon the specimen geometrical dimensions (especially the ratio between the hole diameter and the specimen width). As opposed to metallic structures, for structures made of composite materials the laminate layup also plays a major role affecting the strength allowables. When the design introduces a short edge distance or layup which is significantly different from that of the test specimen (which is usually a quasi-isotropic), finite width correction factors should be applied to the damage tolerance strength allowables to ensure the structural integrity of the composite part. Machine Learning (ML) algorithms have been used to study different characteristics of composite materials in the past few years. The main advantage of using such approaches over standard numerical analyses is their capability to efficiently perform regression based on relatively small dataset. ML approaches can also deal with large data very efficiently. While there are many ML strategies available, they all similar to each other in their regression process. As a first step, the algorithm is 'trained' with respect to a given dataset. In this process, the algorithm establishes correlations between the different data points with respect to predefined characteristics. It is strongly recommended to use a physical-based model for the training procedure, especially if the training process is based on relatively small dataset. Once these relations are established, the entire investigated domain can be spanned using regression. Several machine learning algorithms were utilized in this study to obtain finite width correction factors for IM7/8552 unidirectional composite material. The proposed methodology also includes a procedure to determine the number of required training data points using the Gaussian Process Regression (GPR) algorithm. Finite width correction factors carpet plots were produced, allowing the designer to easily obtain the required correction factor for a given layup and geometrical characteristics. |