Deep learning makes CT X-ray inspection of 3D printed parts faster, more accurate

A new deep learning framework developed at the Department of Energy’s Oak Ridge National Laboratory is speeding up the inspection of further manufactured metal parts using X-ray computed tomography, or CT, while increasing the accuracy of the results. Reducing costs in terms of time, labor, maintenance, and energy is expected to spur the expansion of additive manufacturing, or 3D printing.

ORNL research leader Amir Ziabari said: “Scanning speed reduces costs significantly. “And the quality is higher, so post-processing analysis becomes much simpler.”

The framework has been incorporated into the software used by trading partner ZEISS in its machines at DOE’s Manufacturing Demonstration Facility in ORNL, where the companies hone their 3D printing methods.

ORNL researchers have previously developed technology that can analyze the quality of a part while it is being printed. The addition of high image accuracy after printing provides an additional level of confidence in additive manufacturing while potentially increasing output.

“With this, we can inspect each part from the 3D printer,” says Pradeep Bhattad, ZEISS business development manager for additive manufacturing. “Currently CT is limited to prototyping. But this one tool could push additive manufacturing towards industrialisation.”

CT X-rays are important to certify the durability of a 3D printed part without damaging it. The process is similar to a medical CT X-ray. In this case, an object placed inside the cabinet is slowly rotated and scanned at each angle with strong X-rays. Computer algorithms use the resulting two-dimensional projection overlay to create a 3D image that displays the density of the object’s internal structure. CT X-rays can be used to detect defects, analyze damage or certify that a product conforms to intended composition and quality.

However, X-ray CT is not widely used in additive manufacturing because current scanning and analytical methods are time consuming and inaccurate. Metals can perfectly absorb lower-energy X-rays in the X-ray beam, producing image accuracy that can be doubled if the object has a complex shape. Errors in the image can obscure cracks or voids that the scan is intended to reveal. A trained technician can correct these problems during the analysis, but the process is time- and labor-intensive.

Ziabari and his team developed a deep learning framework to quickly provide a clearer, more accurate reconstruction and an automated analysis. He will present the process his team has developed during the International Conference on Image Processing of the Institute of Electrical and Electronics Engineers in October.

Training a supervised deep learning network for CT often requires many expensive measurements. Because metal parts poses additional challenges, getting proper training data can be difficult. Ziabari’s approach takes a giant leap forward by generating actual training data without requiring extensive tests to collect it.

A generic adversarial network, aka GAN, is used to aggregate creating a realistic-looking dataset for training. neural network, which leverages physics-based simulation and computer-aided design. GAN is a machine learning class that uses neural networks to compete with each other like in a game. Ziabari says it’s rarely used for practical applications like this.

Because this X-ray CT frame needs scans with fewer angles to achieve accuracy, it reduced the time it takes to take pictures by six times, says Ziabari — from about an hour to 10 minutes or less. Working quickly with too few viewing angles will often add significant “noise” to a 3D image. But the ORNL algorithm taught on the training data overcame this, even enhancing its ability to detect minor flaws by a factor of four or more.

The framework developed by Ziabari’s team will allow manufacturers to quickly refine their builds, even with design or material changes. With this approach, sample analysis can be completed in one day instead of six to eight weeks, Bhattad said.

“If I can quickly and cost-effectively inspect the entire part, we can be 100% confident,” he said. “We are working with ORNL to make CT a trusted and accessible industry testing tool.”

ORNL researchers evaluated the performance of the new framework on hundreds of samples printed with different scanning parameters, using complex, dense materials. These results are very good, and ongoing trials at MDF are working to verify that the technique is equally effective with any metal alloy, Bhattad said.

That’s important, because the method developed by Ziabari’s team can make it a lot easier to certify parts made from new metal alloys. “People don’t use novel materials because they don’t know the best printing parameters,” says Ziabari. “Now, if you can characterize these materials quickly and optimize the parameters, that will help move these new materials to additive manufacturing.”

In fact, Ziabari said, the technology could have applications in many fields, including defense, auto manufacturing, aerospace and electronic printing, as well as evaluating the durability of vehicle batteries. electricity.