Accelerated Pyrolysis Analysis of Thick High-Temperature Composite Parts Using Theory-Guided Probabilistic Machine Learning and Finite Element Analysis

Abstract

Fabrication of advanced composites for high-temperature applications typically involves a complex, multi-step process. This process includes an initial lay-up step, curing, high-temperature pyrolysis to transform the cured resin into a carbonized amorphous structure, several resin-backfilling steps to fill voids and cracks formed during pyrolysis, additional pyrolysis-densification cycles to further increase the carbon content, and a final graphitization step to achieve the desired crystalline structure of carbon atoms. The processing parameters in each step, as well as the lay-up and overall thickness of the part, directly impact the kinetics of reactions, phase transformations, and the resulting end-part properties. During pyrolysis, the resin undergoes a complex system of degradation reactions affected by heat and mass transfer through the part thickness. To achieve optimal performance of high-temperature composites, it is crucial to accurately establish the relationship between processing conditions, part geometry, and end-part properties. Current process optimization relies heavily on time-consuming and resource-intensive testing campaigns to characterize pyrolysis kinetics, followed by trial-and-error efforts. This is further complicated by complex temperature cycles, gradients of pyrolysis rate across thick composite parts, uncertainties, and variabilities of the material and process. This paper introduces a novel framework combining theory-guided probabilistic machine learning (ML) and finite element (FE) process simulation to address these challenges. Limited targeted experiments are used to characterize pyrolysis kinetics using an in-house developed ML code while quantifying uncertainty at each testing step. The kinetics model is then used by an FE model to simulate heat transfer and rate of pyrolysis, allowing the gradient of degree of pyrolysis across the thickness of a laminate to be simulated. For validation, laminates of varying thicknesses were fabricated and pyrolyzed using different temperature cycles. Imaging techniques including light microscopy and scanning electron microscopy (SEM) were utilized to capture the resulting microstructures. An image analysis code was developed to detect cracks and voids to identify potential gradients and patterns. This, along with Raman spectroscopy, was used to estimate the degree of mass loss across the thickness of different parts and compared with predictions of the combined ML-FE framework. This framework significantly reduces trial-and-error experimental efforts and enables pyrolysis process optimization of thick composites to achieve desired properties, such as yield, porosity, and laminate permeability.