Oak Ridge National Laboratory has developed a novel deep-learning framework that accelerates the process of examining additively built metal components using X-ray computed tomography, or CT, while improving the accuracy of the results. The growth of additive manufacturing, also called 3D printing, is expected to be helped by lower costs for time, labor, maintenance, and energy.
Amir Ziabari, the principal researcher at ORNL, remarked, “The scan speed greatly cuts costs.” “As a result of the improved quality, post-processing analysis becomes simpler.”
The framework has already been included in the software used by commercial partner ZEISS on its machines at ORNL’s Manufacturing Demonstration Facility, where companies develop 3D-printing techniques.
Previously, ORNL researchers have created equipment that can analyze the quality of a printed item as it is being produced. Adding a high level of image accuracy after printing makes people feel more confident in additive manufacturing and could increase output.
“With this, we can inspect every 3D-printed part,” said Pradeep Bhattad, business development manager for additive manufacturing at ZEISS. “CT is currently confined to prototyping. But this one tool can push the industrialisation of additive manufacturing. ”
X-ray CT scanning is crucial for validating the structural integrity of a 3D-printed component without destroying it. The method is comparable to that of medical X-ray CT. In this instance, a cabinet-enclosed object is slowly rotated and scanned from all angles using strong X-rays. Using the resulting stack of two-dimensional projections, computer algorithms generate a three-dimensional depiction of the density of the object’s underlying structure. X-ray computed tomography (CT) can be used to find flaws, figure out why something failed, and make sure that the composition and quality of a product match its specifications.
However, X-ray CT is not widely employed in additive manufacturing since existing scanning and analysis techniques are time-consuming and imprecise. Metals can completely absorb the lower-energy X-rays in the X-ray beam, resulting in picture errors that are amplified if the item has a complicated geometry. The resulting image faults may mask fractures or pores that the scan is meant to reveal. A skilled technician can make adjustments for these problems during analysis, but it takes a long time and is hard work.
Ziabari and his colleagues developed a deep-learning framework that enables faster, more accurate reconstruction and automated analysis.In October, he will discuss the method his team created at the International Conference on Image Processing, sponsored by the Institute of Electrical and Electronics Engineers.
Typically, training a supervised deep-learning network for CT needs numerous costly measurements. Because metal parts provide extra problems, it can be difficult to obtain the necessary training data. Ziabari’s method advances the field by producing realistic training data without requiring lengthy experimentation.
Using physics-based simulations and computer-aided design, a generative adversarial network (GAN) method is employed to generate synthetically a realistic-appearing data set for training a neural network. GAN is a type of machine learning that employs neural networks to fight against one another in a game-like manner. According to Ziabari, such practical uses have been rare.
According to Ziabari, because this X-ray CT framework requires fewer scan angles to attain accuracy, it has decreased imaging time by a factor of six, from around one hour to ten minutes or less, according to Ziabari. Working so quickly with so few viewing angles would often result in a 3D image with severe “noise.” But the ORNL method taught on the training data fixes this, making it four or more times easier to find small flaws.
The framework created by Ziabari’s team would enable manufacturers to rapidly fine-tune their creations, even as designs or materials are altered. With this method, sample analysis can be finished in a single day as opposed to six to eight weeks, said Bhattad.
“If I can quickly and cost-effectively inspect the entire item, then we have complete faith,” he added. We are working with ORNL to make CT a reliable and easy-to-use inspection tool for the business world.
Researchers at ORNL examined the performance of the new framework by printing hundreds of samples with varying scan parameters and complex, dense materials. Bhattad stated that these results were favorable and that ongoing tests at MDF were confirming the technique’s efficacy with any type of metal alloy.
This is significant because the approach developed by Ziabari’s team could make it much simpler to verify metal alloy components. Zibari stated, “People do not use novel materials because they do not know the optimal printing parameters.” Now, if you can quickly figure out what these materials are and how to change their properties, they could be used in additive manufacturing.
Ziabari says that the technique can be used in many fields, including defense, auto manufacturing, aerospace, electronic printing, and nondestructive inspection of electric car batteries.
