Researchers develop deep learning alternative to laser powder bed fusion monitoring

Many problems can occur with additively manufactured (AM) metals, and without in-situ process monitoring, defects can only be detected and characterized after the product is manufactured. Most commonly, manufacturers use high-speed cameras to monitor the shape of the melt pool and its changes during short periods of the laser powder bed melting (LPBF) process.

This requires expensive equipment, extensive memory storage (i.e. storing 20,000 to 30,000 high-resolution photos per second), and human effort to collect and classify the data. These ultimately increase the cost of online visual tracking and process analysis.

To achieve automated and cost-effective in-situ visual monitoring during metal AM, Carnegie Mellon University School of Engineering researchers use airborne acoustics or heat alone to capture and characterize melt pools in LBPFs. We have developed a deep learning approach that provides an alternative. emissions.

This team's method has recently been additive manufacturing journalenables manufacturers to capture critical melt pool geometries and predict transient melt pool fluctuations almost instantly.

“By taking advantage of the underlying physics of multimodal process signals and data-driven artificial intelligence, our pipeline enables engineers to create highly affordable and accessible sensors such as microphones and photodiodes. can be used to reconstruct important melt pool properties,” said Dr. Haolin Liu. .D. Candidate in Mechanical Engineering.

One clear advantage of this new approach is the potential to identify spatially dependent loss of fusion (LOF) defects in LPBF. One of the most typical process anomalies, LOF occurs when there is insufficient overlap of the melt pool as the laser passes through the powder bed.

The resulting unfused powder leaves large unfused gaps and residual porosity in the part that can significantly compromise the durability and other mechanical properties of the final product. Therefore, capturing these localized defects and melt pool fluctuations in real time is critical to producing consistently durable products.

The research team conducted a series of LPBF experiments to investigate various printing parameters of the titanium alloy Ti-6Al-4V (Ti-64). Aircraft acoustic, thermal, and high-speed imaging data were collected and synchronized from the pre-designed and completed structure for each corresponding process condition, successfully reconstructing the precise melt pool geometry. The researchers also tracked the vibrational behavior of the melt pool over short periods of time, just a few milliseconds. This approach also showed a promising ability to effectively detect local His LOF defects between two adjacent laser scan lines.

“This method enables monitoring of melt pools using low-cost sensors that can be installed on any laser powder bed AM machine. Generating artificial videos of high-velocity melt pools from acoustic and photodiode sensor data. is unique to the AM community,” said Jack Buse, professor of mechanical engineering and co-director of the NextManufacturing Center.

Additionally, the team's research has also led to an important step towards a deeper understanding of the physical correlations between multimodal field process signals.

“The cross-correlation between these signals has not yet been fully explored by the scientific community,” Liu said.

“Although our research focused on deep learning data-driven pipelines, it became clear that certain elementary relationships exist between acoustic properties, thermal radiation, and melt pool morphology. Its physics and mechanics require further scientific exploration and experimental investigation.”

“Many experts are aware of the interaction between acoustic radiation, thermal radiation, and the resulting melt pool dynamics in laser printing, but the exact relationship is still largely unknown.” Mechanical Engineering Professor Levent Burak Kara said.

“In this study, we established and demonstrated a data-driven predictive model that relates these three phenomena in a highly accurate and physically meaningful way.”

According to Anthony Rollett, professor of materials science and engineering and co-director of the NextManufacturing Center, acoustic behavior involves a significant physical interaction between the laser and the material.

“To our surprise, it revealed more than we expected and proved to be very useful in informing process-related quantities that could potentially impact manufacturing quality. .”

In the future, the team plans to explore more real-time monitoring applications leveraging acoustic and thermal emission data across materials other than Ti-64 and different platforms and AM processes.

“By deeper interpreting the potential of acoustic and thermal radiation, we hope to better understand the relationship between melt pool variability, keyhole vibrations, and other spatially dependent process features,” Liu he said.

“One day, we may be able to build fully functional digital twins with advanced surrogate models for other process characterization equipment, such as synchrotron X-ray equipment or entire AM processes.”