Session: 03-07-01 Materials for Extreme Environment I

Paper Number: 105683

105683 - A Novel Framework for Accelerated Characterization of Pyrolysis Kinetics of High-Temperature Composites Using Theory-Guided Probabilistic Machine Learning 

Establishing the relationship between processing conditions and end-part properties is essential for the efficient manufacturing of advanced composites and for ensuring optimal performance. Depending on the complexity of the manufacturing process, the effect of processing parameters in each step, as well as the cumulative effects of all parameters on the final part properties, must be well understood. For fiber-reinforced polymer composites for high-temperature applications, this manufacturing process typically includes multiple complex stages such as a lay-up step to deposit the material onto a mold; a curing stage in which the polymeric reaction of the matrix occurs; a high-temperature pyrolytic process through which the cured polymer is converted into an amorphous structure with high carbon content; several resin backfill steps to fill any voids and cracks formed during the pyrolysis step, followed by pyrolytic process repeats to increase the carbon content of the material; and a graphitization stage so that amorphous carbon structure transforms into the desired crystalline structure. The processing parameters in each stage will directly impact the kinetics of the corresponding reactions and their phase transformations, and hence the end-part properties. During the high-temperature pyrolytic process, in particular, a polymer resin undergoes multiple complex degradation reactions. The extent and evolution of these reactions, along with the final yield and laminate permeability depend on the temperature cycle used during the process, as well as on other factors including geometry and lay-up. Characterizing the pyrolysis kinetics to identify the optimal temperature parameters typically requires an exhaustive testing campaign to study the effect of different temperature cycles on the final part properties. Given the type of temperature cycles of interest, this may involve considering many combinations of multi-heating ramps and holds, for which the effects of the heating rate at each heating step, and hold temperature/time at each dwell step, are carefully analyzed.  Each of these tests can take a significant amount of time; thus resulting in an intractable search space bounded by the industry's knowledge and experience. This approach severely limits the understanding of correlations between processing parameters and the material’s final properties, and hence hinders the success of any optimization task. Although established literature succeeds in characterizing the pyrolysis kinetics for simple dynamic temperature cycles, it fails to account for the effect of more complex cycles that are of interest to industry on properties such as final yield or permeability. This paper aims to address the above challenges by introducing a novel probabilistic machine-learning (ML)-based framework for the accelerated characterization of the pyrolysis kinetics. Using process-specific and theory-based transformations of limited and noisy experimental data, this framework can iteratively characterize pyrolysis kinetics subjected to complex temperature cycles, while quantifying uncertainty in each step, and guiding the experimental approach to minimize uncertainty. In this study, for a given polymer-based composite, the degradation reactions and fraction of volatile components of the material are studied using thermogravimetric analysis techniques. Results are analyzed using theory-guided Gaussian Process Regression (GPR), a Bayesian probabilistic approach to regression, to accurately characterize the process kinetics. In a step-by-step guided effort, experiments are conducted to characterize the pyrolysis kinetics of complex temperature cycles. The framework is then successfully used for the identification of the optimal parameters to achieve the desired yield while satisfying constraints related to both temperature and time. Using this approach, the experimental effort is considerably reduced while developing an accurate pyrolysis kinetics model for the material and establishing optimal processing parameters.

Presenting Author: Paulina Portales Picazo University of Washington

Presenting Author Biography: Paulina Portales is a Ph.D. student in the Materials Science & Engineering department at the University of Washington. She's part of Dr. Navid Zobeiry's research group and has focused her research on the accelerated characterization of high-temperature composite materials using Machine Learning (ML)-based methods.
Before joining UW, she completed her BS in Chemical Engineering at ITESM in Mexico, and spent a year doing research focused on non-invasive diagnostic techniques at the Boston Children's Hospital.

Authors:

Paulina Portales Picazo University of Washington
Alexander Gray University of Washington
Navid Zobeiry University of Washington