Session: 03-06: Materials Development Using Artificial Intelligence

Paper Number: 121050

121050 - Accurate Prediction of Process-Induced Deformations in Composites Using Multi-Fidelity Simulation and Theory-Guided Probabilistic Machine Learning 

     Carbon fiber-reinforced polymer composites have gained significant popularity in high-performance aerospace applications thanks to their advanced and tailorable properties. Nonetheless, despite their widespread utilization, aerospace manufacturers regularly face several challenges, one of them being the prediction and control of process-induced deformations (PIDs) in composite parts. During autoclave processing of composites, residual stresses arise due to complex phenomena at different scales in the material and manufacturing environment. Upon demolding, some of these stresses may be alleviated through deformations such as alterations in enclosed angles at geometry transitions (i.e., spring-in or spring-out) or distortions of initially flat sections (i.e., warpage). Consequently, these PIDs may induce significant mismatches between components during assembly, prolong production timelines, and compromise the final aerostructure’s mechanical efficiency.

     Despite having a basic understanding of process-induced stresses and deformations, composites manufacturers often face significant hurdles when attempting to predict PIDs in industrial settings. These difficulties primarily stem from the inherent limitations of physics-based simulation tools in accurately modeling complex PID-contributing parameters such as tool-part interaction, interply-toughened microstructures, and others. Moreover, current simulation tools tend to disregard the influence of processing uncertainty and variability on PIDs, yielding deterministic predictions that do not align with real-world composites manufacturing scenarios. As a result, physics-based simulation tools often produce PID predictions with inevitable errors, and thus attempts to predict deformations are inefficient or unsuccessful. Therefore, improving these methods or developing strategies to harness their advantages while addressing limitations may be critical to help improve the understanding of and mitigate PIDs in composites.

     This paper presents a generalizable framework for the accurate evaluation and prediction of PIDs in composite parts. The method combines data from three sources to train probabilistic machine learning (ML) models. First, established closed-form theory is utilized to construct a theory-guided (i.e., physics-based) design architecture. Next, low-fidelity simulations consisting of 1D thermo-chemical and 2D thermo-mechanical analyses are employed to provide quick estimations of PIDs and inter-variable relationships in the design space. Throughout the low-fidelity evaluation, a Gaussian Process Regression (GPR) model is trained after each test (i.e., stepwise) until fitting accuracy and confidence bounds converge. Then, the solution scheme adaptively switches to utilizing high-fidelity 3D simulation data, where it is again iteratively updated until the convergence of both prediction accuracy and confidence. In this work, an established simulation method based on layer-wise models and the Carrera Unified Formulation (CUF) is utilized for the high-fidelity component. Finally, a small number of element-level experiments are conducted to calibrate and introduce uncertainty into the GPR model. At the conclusion of the solution scheme, a probabilistic model is built and validated by comparing its predictions to experimental results and its associated costs with traditional methods. The results in this paper include a case study involving L-shaped composite parts made from Toray’s T800S/3900-2B aerospace material system. The method proposed in this paper is a probabilistic, cost-efficient, and generalizable method to characterize, predict, and potentially mitigate PIDs in composites.

Presenting Author: Caleb Schoenholz University of Washington

Presenting Author Biography: Caleb Schoenholz is a Graduate Student at the University of Washington Department of Materials Science & Engineering and a part of the Composites Group led by Dr. Navid Zobeiry. Prior to joining UW, Caleb received his Bachelor of Science degree in Composite Materials Engineering from Winona State University. Caleb's research is broadly focused on process-induced residual stresses and deformations in advanced aerospace composites.