A Deep Learning-Driven Fast Scanning Method for Micro-Computed Tomography Experiments on CMCs

Background: In-situ micro-computed tomography (µCT) technology is an attractive approach to investigate the evolution process of damage inside ceramic matrix composites (CMCs) during high-temperature service. The evolution process is highly time-sensitive under temperature-induced loads, and fast scanning is very necessary for in-situ µCT tests. Objective: The objective of this work is to provide a fast scanning method for in situ µCT tests on CMCs with complex microstructures by the innovation of a reconstruction algorithm. Method: To overcome the severe degradation of the reconstructed image quality resulting from sparse CT scans, a deep-learning-based multi-domain sparse reconstruction method was proposed. Three sub-networks including the projection-domain, image-domain, and fusion network were constructed in the multi-domain method to make full use of the information from the projection and image domain. Results: The proposed deep-learning-based sparse reconstruction method provided satisfactory µCT images on C/SiC composites with acceptable quality. The scanning time was reduced by 6 times. All selected evaluation metrics of the proposed method are higher than those of other single-domain methods and traditional iterative method. The segmentation accuracy of the µCT images obtained by the proposed method can meet the subsequent quantitative analysis. An in-situ tensile test of CMCs is conducted to further evaluate the performance in the practical application of in-situ experiments. The results indicate that the weak and thin micro-cracks can still be effectively retained and recovered. A detailed workflow to implement the method generally is also provided. Conclusions: Based on the deep-learning-based multi-domain sparse reconstruction method, the process of in-situ µCT tests can be greatly accelerated with little loss of the reconstructed image quality.