3D Software

ORNL develops Peregrine AI software for real-time monitoring of metal 3D printing

Oak Ridge National Laboratory (ORNL) researchers have developed Artificial Intelligence (AI) software capable of monitoring the metal 3D printing process in real-time. 

Nicknamed Peregrine, the algorithm assesses the quality of parts during production in a cost-effective alternative to characterization equipment. The program is part of ORNL’s broader “digital thread,” which closely tracks and analyzes data through each step of the manufacturing process. Within the factories of the future, the ORNL team believes their algorithm could be utilized as a quality control method for self-correcting machines.

Researcher Chase Joslin using the Peregrine software to monitor and analyze a component being 3D printed. at ORNL. Photo via Luke Scime/ORNL.
Researcher Chase Joslin using the Peregrine software to monitor and analyze a component being 3D printed. at ORNL. Photo via Luke Scime/ORNL.

Automating metal 3D printing 

During powder-based metal 3D printing, a number of issues can arise which negatively impact upon the features of the end-use printed part. The uneven distribution of powder or binding agent, spatters, insufficient heat, and porosities can lead to defects on the surface of each layer. Many of these printing issues go undetected using conventional monitoring techniques, leading to suboptimal parts or the component being scrapped altogether. 

Although some production anomalies happen very quickly and can be difficult to prevent, others are more predictable and large enough to make layer-wise detection possible. 

“One of the fundamental challenges for additive manufacturing is that you’re caring about things that occur on length-scales of tens of microns and happening in microseconds,” said Luke Scime, Principal Investigator for Peregrine. “Because a flaw can form at any one of those points at any one of those times, it becomes a challenge to understand the process and to qualify a part.”

According to the ORNL team, prior attempts to monitor 3D printing errors in real time have not been automated enough to allow their widespread adoption in factory settings. In terms of analyzing the data collected during the production process, previous approaches have also focused on comparing the finished part to its 3D model. The benefit of a comparative approach is that it’s simple, but it’s also time limiting, and only identifies one ‘generic flaw’ in a printed object.

A number of abnormalities were observed by the ORNL team using their ML algorithm. Image via the Additive Manufacturing journal.
A number of abnormalities were observed by the ORNL team using their ML algorithm. Image via the Additive Manufacturing journal.

ORNL’s Peregrine software

In order to overcome the limitations observed in earlier research, the ORNL team developed a novel Convolutional Neural Network (CNN) architecture. The computer vision technique uses a custom-designed algorithm to carefully examine the pixel values of images taken during the printing process. 

In practice, when an anomaly is detected during printing, the software is able to alert the machine’s operator that adjustments need to be made. Modelled on the human brain, Peregrine is also capable of analyzing and sharing images of surface visible defects across multiple 3D printers. As a result, using the ORNL algorithm, one system is able to learn from the printing errors encountered by another.  

“Capturing the information creates a digital ‘clone’ for each part, providing a trove of data from the raw material to the operational component,” explained Vincent Paquit, who leads part of ORNL’s Imaging, Signals and Machine Learning Group. “We then use that data to qualify the part and to inform future builds across multiple part geometries and with multiple materials.”

Testing ORNL’s deep learning method

In order to test Peregrine’s deep learning algorithm, the researchers provided it with a data set for evaluation. Objects were 3D printed across eight different Laser Powder Bed Fusion (LPBF) and binder jetting machines. Two 8-bit images were captured per camera, for each layer, and on every system, with one immediately following powder fusion or binder deposition, and the other after powder spreading. 

The performance of the algorithm was then measured based on the amount of inter-machine transfer learning that took place. Evaluating the data shared between a GE Additive ConceptLaser M2 and an ExOne Innovent system, the team found that 16 percent of testing was lost over 10,000 test batches. Although the algorithm had, therefore, demonstrated its ability to share information between systems, a significant amount of ambiguity remained about what is labeled as an anomaly. 

For instance, parts printed using an Arcam Q10 additive manufacturing system reported a true-porosity ratio of 78.4 percent, but if those within the space of two pixels are discounted, that rises to 89.5 percent. The algorithm’s performance was also found to be strongly dependent on its input data. If the anomaly in the printed part isn’t clear in the captured images, it’s unlikely to be flagged by the software. As a result, refining the lighting and imaging configurations during testing would likely have improved Peregrine’s performance. 

Overall, the algorithm achieved segmentation times of 0.5-2.4 seconds per layer, making it rapid enough for applications in serial manufacturing. The software also demonstrated the ability to pass significant amounts of data from machine to machine, across different 3D printing technologies. In future research, the team concluded that they will explore different imaging wavelengths, lighting conditions, and spatially mapped sensor modalities. 

3D printed components for the prototype reactor. Photo via Britanny Cramer/ORNL/US Dept. of Energy.
ORNL has used its algorithm to optimize the production of its 3D printed reactor (pictured). Photo via Britanny Cramer/ORNL/US Dept. of Energy.

Nuclear applications of ORNL’s algorithm 

Over the last six months, Peregrine has been tested on hundreds of builds at ORNL, including as part of its 3D printed Transformational Challenge Reactor (TCR) program. ORNL’s TCR project aims to build an eco-friendly microreactor by 2023, using a set of integrated sensors and controls to capture data. 

The ORNL research team have deployed their newly-developed algorithm to optimize the process, lowering the lead times and costs associated with building the reactor core. 

“Establishing correlations between the signatures collected during manufacturing and performance during operation will be the most data-rich and informed process for qualifying critical nuclear reactor components,” said Kurt Terrani, TCR program director. “The fact that it may be accomplished during manufacturing to eliminate the long and costly conventional qualification process is the other obvious benefit.”

Fully-automated 3D printing 

In addition to its potential nuclear applications, the Peregrine software could also provide automated quality control within self-sustaining 3D printing factories. A number of companies have developed their own ‘factory of the future’ projects in recent years, with the aim of fully-automating the manufacturing process. 

EOS, Daimler, and Premium AEROTEC have successfully piloted their NextGenAM project in a German plant. The self-sustaining factory’s software has a digital twin feature, which makes it possible to “copy and paste” the same system of machines in order to increase capacity. All stages of the process were automated, including powder handling, print bed removal, post-processing, and quality inspection.

Automotive manufacturer BMW is developing its own self-regulating processes in the form of its IDAM and Polyline projects. As part of its IDAM venture, BMW aims to digitize the production of at least 50,000 components and over 10,000 individual and spare parts per year. Within its Polyline program, the company is working with 15 partners to develop digitally linked process steps, and a consistent quality assurance methodology.

The researchers’ findings are detailed in their paper titled “Layer-wise anomaly detection and classification for powder bed additive manufacturing processes: A machine-agnostic algorithm for real-time pixel-wise semantic segmentation,” which was published in the Additive Manufacturing journal. The report was co-authored by Luke Scimea, Derek Siddel, SethBaird, and Vincent Paquita. 

Nominations for the 2020 3D Printing Industry Awards are still open, let us know who is leading the industry now.

The fourth edition of the 3D Printing Industry Awards Trophy Design Competition is now underway. Enter your design for the chance to win a CraftBot Flow 3D printer. 

To stay up to date with the latest 3D printing news, don’t forget to subscribe to the 3D Printing Industry newsletter or follow us on Twitter or liking our page on Facebook.

Are you looking for a job in the additive manufacturing industry? Visit 3D Printing Jobs for a selection of roles in the industry.

Featured image shows researcher Chase Joslin using the Peregrine software to monitor and analyze a component being 3D printed. at ORNL. Photo via Luke Scime/ORNL.