This project aimed to broaden the knowledge of high-temperature polymers and composites when exposed to elevated temperatures and an oxygen-containing atmosphere. The main accent has been on thermosetting polyimide resins reinforced with carbon fibers. When subjected to harsh atmospheric conditions, such as elevated temperatures and oxygen, polymer resins can undergo thermo-oxidative degradation, often resulting in weight loss and a surface layer with altered properties. High-temperature composites could experience such environments during operation. Therefore, it is crucial to understand how exposure to it could affect their performance. To simulate such an environment in the lab, the materials are aged in a controlled manner in a furnace or other equipment. The ageing of polyimide composites in this project was often performed at temperatures at or above 288 degrees Celsius for extended periods of up to 1500 hours.

The first part of the project, and the first article, delved into the effect of different layups and thicknesses of the carbon fiber bundles on the thermo-oxidative behaviour of two composite materials made with the same thermosetting polyimide. Modelling the desorption during the initial stages of the ageing, showed that it exhibited a Fickian behaviour. X-ray computed tomography experiments were used to investigate the ageing behaviour of the materials and revealed that the satin weave composite formed a network of cracks, voids, and delaminations, that progressed with the ageing time, while the damage in the material made of thin plies was in the form of delaminations at the edges. The analysis of the tomographic datasets was performed using Otsu’s thresholding method for semantic segmentation of the defects within the materials.

In an attempt to counter the crack formation on the surface of the satin weave composite observed during the first study, a new polyimide formulation was developed by the manufacturer. The amount of internal crosslinkers was reduced, aiming to increase the toughness of the resin after curing. The second article compares neat resin samples of the original and newly developed formulations with the help of a three-point bending test, differential scanning calorimetry, dilatometry, weight loss, light optical microscopy and nanoindentation experiments. Samples were aged up to 1500 hours in ambient air. The results showed that while there were hints of a slight increase in the fracture toughness of the new formulation, the glass transition temperature had decreased, compared to the original resin.

The two formulations were further investigated and compared with the help of thermogravimetric analysis in the fourth paper. Experiments were performed in isothermal and non-isothermal conditions for more robust results. It was found that the thermal oxidation of the two materials follows an autocatalytic model. The study highlights the importance of using both isothermal and non-isothermal data in the pursuit of more precise and robust analysis and modelling of the thermal oxidation of high-temperature polymers. Based on the results, a diagram, predicting the weight loss at specific times and temperatures, was created for each material.

An alternative way of studying crack formation within challenging polymer composite tomographic datasets was presented in the fourth article. Instead of using a thresholding method, such as the

previously used Otsu’s in the first study, in this case, a deep learning model was applied to the datasets to follow the progressive micro-cracking within the composite during a series of thermo-mechanical loadings. In contrast to a global thresholding method, which segments all defects within the dataset, the deep learning model, Attention U-Net, made it possible to create a more straightforward and robust way of performing segmentation on transverse cracks. The model was compared to and outperformed both Otsu’s method and a conventional U-Net.

The previously developed methodology for semantic segmentation and the obtained results on transverse cracks were applied in a practical case in the fifth article, where the developed damage prediction model assumes that transverse cracks in thick plies span through the whole width of the specimen. The tomography and deep learning methodology helped shed light on the nature of the cracks and showed that previous assumptions, based on edge observation with light optical microscopy, should be taken as a conservative estimation.