Session: 01-05-01: Applications of Artificial Intelligence/Machine Learning for Aerospace Structures

Paper Number: 121597

121597 - A Novel Machine Learning Framework for Digital Estimation of Allowables in Aerospace Composites 

Certification of composite aerostructures typically follows the established building-block approach of analysis supported by testing. These comprehensive testing campaigns span from coupon-level to sub-component and component. Different environmental conditions, testing geometry, loading conditions, lay-up, and material architecture are often considered. To account for material variability and processing uncertainties, tests from different batches are repeated to establish statistical significance and set A-basis and B-basis design allowables. However, establishing these statistical values is expensive, time-consuming, and limits the exploration of the design spectrum, for example, in terms of lay-up or material architecture.  Additionally, the high cost associated with this process inhibits the qualification and certification of new material systems and designs. 

While the failure simulation of composites has progressed significantly in the past decade, current approaches are inherently slow and ignore variability and uncertainties in deterministic methods. The calibration process for high-fidelity simulation models is non-trivial. Although low-fidelity and reduced-order modeling can address some of these difficulties, the accuracy and applicability of such models have been questioned. Given the multi-scale nature of composites, efforts were undertaken to develop multi-scale simulation approaches for failure analysis. However, such models often neglect the effect of processing and process-induced defects on material performance and design allowables. Machine learning (ML) methods have been explored to assist in the certification process and quantify uncertainty in recent years. However, success has been limited due to difficulties associated with generating the necessary data for training ML models with high-fidelity and multi-scale simulation models.

 In the proposed framework, to consider the effect of material and processing uncertainties on allowables, we combine stochastic process simulation with stochastic failure analysis, by considering effects of processing-induced defects directly on failure. To enable machine learning analysis, we employ both low-fidelity and high-fidelity FE simulations, as well as multi-scale simulations. Low-fidelity simulations enable the generation of large amounts of data, while limited high-fidelity simulations are used as ground truth. To consider effects-of-defects on performance, we employ a multi-scale simulation approach. This framework allows for training inverse surrogate ML models with limited data from stochastics FE analysis. These models can be used for quantifying uncertainty in material and process using a Monte Carlo approach. The estimated uncertainty can then be used to digitally arrive at allowables.

In a case study, Hexcel AS4/8552 is chosen for virtual fabrication and testing. The proposed framework is used to estimate allowable as a function of a several uncertainties, including heat transfer uncertainties, variation of residual stresses, and volume fraction variation. For validation, our digitally estimated design allowables are compared with publicly available experimental data.

Presenting Author: Amirali Eskandariyun University of Washington

Presenting Author Biography: Amirali Eskandariyun is a Ph.D. student in the University of Washington's materials science and engineering program. He has joined UW Composites Group, led by Prof. Navid Zobeiry, as a graduate research assistant in 2022. His research interests includes composite structures, computational mechanics, and machine learning.