ADVANCED MATERIALS & PROCESSES | JANUARY/FEBRUARY 2025 30 To accomplish this feat, the AI program can be pretrained on simulated CT scans of the CAD models of a manufacturer’s parts. Ideally, these scans will include realistic simulations of every possible defect type, such as round pores, ripped edge blowholes, and other irregularities. This data comes with the precise ground truth the models need for proper training. The AI model applies its learning to recognize patterns and features by looking at sample data for which the solution is known. In this way, AI can provide highly accurate, nearly real- time snapshots of what is happening on a particular production line. This capability supports confident decision- making about whether to accept or reject a part. This in turn informs production-variable changes, the results of which can be captured, collated, and statistically examined (Fig. 2). Note that an essential part of using AI in this way involves creating and curating the datasets that are being used to train the software. Providers are working closely with manufacturers on this effort, under non-disclosure agreements that safeguard proprietary information. At the same time, pilot programs are being conducted to enable these customers to teach their own AI systems with their in-house data. AI CASE STUDY iWP, a materials testing laboratory and service provider, is working to demonstrate the power and effectiveness of some of the latest deep learning software applications. The lab had already been using advanced segmentation tools for some time to locate regions of interest (ROIs) in CT scans of industrial parts and help identify the volumes of pores and particles that might affect part quality. For a broad range of industrial customers, iWP has been scanning everything from computer chips to cement. CT data analysis software can identify multiple different materials in a single part, such as rubber, metal, air, pores, and particles. applied to the resulting metrics enables manufacturers to achieve continuous improvements and deliver better quality products. AI ENTERS THE ARENA While artificial intelligence (AI) may raise controversy in a number of human endeavors, in the manufacturing world, industrial AI is demonstrating great promise and less debate. With NDT and particularly CT scan data analysis, deep learning—a subset of AI—goes far beyond mere optimization, dramatically augmenting and supporting decision-making regarding production. It is doing this by processing exponentially greater quantities of data than ever before and delivering highly accurate conclusions about product quality much faster and more precisely than humanly possible. As the saying goes, “garbage in, garbage out.” With a typical AI search through unregulated data on the internet, the conclusions can be far less than accurate, leading to faulty output and flawed decision-making. But in manufacturing, design, engineering, and analysis, data can be precisely sourced, controlled, segmented, and IT-protected. This restricts and focuses the power of industrial AI to the specific task at hand, generating highly reliable results at lightning speeds. Speed is the key here because— along with quality—it is the primary goal in series production. Advanced users of CT scan data analysis software (i.e., leading aerospace and auto- makers in the U.S. and Europe) are well aware of this. They have already begun working with providers to integrate AI capabilities into the software, in order to use the tools directly on their production lines as parts are being manufactured. How does this work? A high- resolution industrial CT scanner can take 20 to 60 minutes in a testing laboratory to generate a clear image of cavities in cast metal parts, pores in additively manufactured ones, or structural defects in battery anodes and cathodes. This certainly does not fit the definition of time efficiency. But a lower-resolution scanner set up on the production line can deliver scans (albeit noisy ones) in seconds rather than minutes. And AI can now be harnessed to process and make decisions based on these less-clear images as accurately as on high-resolution ones. CREATING AND CURATING DATA This is deep learning in action: trainable algorithms being taught to identify specific manufacturing flaws by interpreting what they look like even if they are fuzzy. Deep learning, based on artificial neural networks, uses its memory to compare what it sees against an existing database of all known identifiable defects, no matter whether the image is crisp or indistinct. Fig. 2 — AI capabilities are being developed to provide machine learning-based deep segmentation of 3D data that delivers precise analysis results quickly (here, anodes in an electric vehicle battery). From top, CT volume, segmentation, and analysis.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==