X-ray computed tomography for additive manufacturing: improved non-destructive evaluation using deep learning

Abstract

X-ray computed tomography is a popular non-destructive evaluation method for additive manufacturing, to provide confidence in part quality and to quantify the defects in critical parts, for example to ensure they are smaller than some size limit or less than some maximum volume fraction, or below the detection limit in critical areas. The method is key to process optimization, quality control and is undisputed as one of the best tools for evaluating AM parts. In recent times, deep learning methods have become popular in various fields, and in the context of CT for additive manufacturing, there are some advancements that are highly valuable for this application. In this talk, two methods are discussed in some detail. The first is image enhancement using deep learning. This can be achieved by different methods, one of which is called “super-resolution”. In this approach, a poor resolution and high resolution scan of the same object is used to train a model to improve the quality of poor resolution scans of similar parts. This allows enhanced contrast on poor resolution scans, which allows the user to save time in scanning or scan at larger voxel size while getting more reliable results similar to a high resolution or longer scan time. This will be demonstrated using a lattice structure sample, which is often used in AM medical implants, but often cannot be scanned at high resolution due to size limitations on the object size vs the lattice feature size. A second method involves teaching a deep learning model to segment AM porosity. The challenge with AM pores are that they are small and often near the voxel size of the scan, making their contrast insufficient for a good manual segmentation. There is also often a challenge with image artifacts due to material density or scan quality issues. These issues make the quantification of porosity challenging, especially when high throughput is required. Deep learning segmentation models can be developed to perform this task despite poor contrast and despite image artifacts, providing superior results in comparison to traditional “thresholding” methods. Results will be demonstrated on using such a model “out of the box” as a pre-trained model, as well as using this model as a starting point for adding training to make a stronger model. Overall these techniques can assist in improving the use of CT for AM in terms of reliability and ease of use, and has great potential for automation of the image analysis workflows involved in using CT (since deep learning models, once trained, do not require any human input).