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

Paper Number: 107538

107538 - A Combined Finite Element and Machine Learning Approach to Accelerate Calibration and Validation of Numerical Models for Prediction of Failure in Aerospace Composite Parts 

The need for an excessive number of costly experimental tests is a major struggle for the certification of composite parts in the aerospace structures industry. Although computational simulation, specifically the finite element method (FEM), has shown many benefits in accelerating the process, the time-consuming calibration process of the model is still an important shortcoming. Moreover, it often requires an engineer with years of experience in order to perform the calibration process successfully. On the other hand, failure analysis of composites requires an explicit FE model with various parameters to tune/calibrate, and thus, makes it even more complicated and expensive. In this research, an automated theory-guided machine learning (TGML) framework has been proposed, aiming to model the failure of composites and calibrate its parameters in an accelerated and low-cost manner. In this framework, FEM parameters are being calibrated using machine learning with respect to experimental test results. To evaluate the capabilities of the proposed framework, a single-edge notch tension (SENT) test model has been chosen as the case study here. The mentioned model with a certain geometry was further generated and imported into an originally developed program. In the first step, the model was simulated explicitly with predefined initial parameters and the result was compared with the experimental results to calculate the error. Consequently, Gaussian process regression (GPR) was utilized to automatically select new parameters considering initial values and the error resulting from the previous step. The program continues this process, repetitively, until the error becomes negligible. Finally, in order to verify and validate the process, the same program was used to predict the results for a center notch tension (CNT) test. In this study, it is demonstrated that the presented automated TGML framework, provides a robust and reliable calibrated model in a significantly accelerated manner, in comparison to the non-autonomous approach, and it requires only a few experimental test data. Hence, it reduces a substantial amount of experimental and computational costs, eliminates the need for high expertise, and saves valuable engineering time.

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 as a graduate research assistant in 2022. His research background includes composite structures and computational materials science.

Authors:

Amirali Eskandariyun University of Washington
Ashith Joseph University of Washington
Alexandru Stere The Boeing Company
Alan Byar The Boeing Company
Sergey Fomin The Boeing Company
Mohammed Kabir The Boeing Company
John Dong The Boeing Company
Navid Zobeiry University of Washington