Symposium CH03: Advanced Characterizations of Heterogeneities in Electrochemical Energy Storage Materials
Symposium EN03: Beyond Li-Ion Batteries—Low Cost Alternatives Based on Other Chemistries

Symposium X—MRS/The Kavli Foundation Frontiers of Materials

Tuesday Symposium X 2_270x180Jian Cao, Northwestern University

Physics-based AI-assisted Property Control in Metal Additive Manufacturing

Written by Sophia Chen

During Tuesday’s Symposium X, mechanical engineer Jian Cao of Northwestern University described research efforts to incorporate machine learning and simulation to improve the technology’s consistency in a talk titled “Physics-based AI-assisted Property Control in Metal Additive Manufacturing.”

In metal additive manufacturing, machines build a metal component layer by layer out of metallic powder or wire and fuse the layers using heat from lasers, electron beams, or other sources. The technology, popularly known as metal 3D printing, dates back to the 1990s, when engineers first used the 3D printed components for rapid prototyping and testing.

Today, engineers use the technology to create components for aerospace, biomedical, and automobile applications. In the last decade or so, metal 3D printing has begun to play a role in final production. However, scalability poses an issue, as the printed components lack consistency. One prototype may have vastly different mechanical properties than another ostensibly identical component.

Cao’s group studies how machine learning and physics simulations could help the metal additive manufacturing process be more consistent. Her talk centered on “how mechanics and AI work together for manufacturing process, design, modeling, and control,” she said. The physics and AI-based modeling predicts how the powder or wire melts and cools, which in turn determines the component’s mechanical properties. Based on those predictions, the researchers can adjust the manufacturing process to achieve the mechanical properties they want.

Tuesday Symposium X_800 wide

For example, Cao’s group recently devised a process that used both machine learning and physics to reduce the number of holes that form between metal layer. The holes constitute an undesirable property known as porosity that weakens the material. One way to control the material’s porosity is to make the melt pool—the area of the component that the laser melts—as consistently as possible.

The physics-based strategy involved numerically simulating the melt pool. Porosity is related to the geometry of the melt pool, Cao explained. To accurately simulate melt pools, the researchers first conducted high-speed x-ray imaging of the 3D printing process at the Advanced Photon Source at Argonne National Laboratory in Illinois. Specifically, they studied a metal 3D printing process known as directed energy deposition. In their setup, focused heat from a laser melts metal powder.

Cao’s team studied how the laser scanning speed, the rate of deposition, among other variables, affected the printed component’s porosity. To study these variables more efficiently, they developed a high-throughput setup that could easily adjust the parameters of the 3D printing process. Using these x-ray images, they characterized the geometry of the melt pool and the pores that formed. In addition, they simulated the depth of the melt pool. The researchers found that by controlling the laser’s intensity, they could produce a melt pool with more consistent depth, and thus control the porosity.

They then took samples out of the materials produced in this process and correlated the material’s mechanical properties with how it cooled, also known as its thermal history. This was the AI part of the process, as they used a purely data-driven algorithm known as a random forest without incorporating any physics knowledge. Using this information, the researchers created a type of AI model called a neural network which they combined with a physics model that could control the material’s thermal history.

In the future, the field needs to work on bridging designers with manufacturers, said Cao.  In particular, she highlighted the need for databases between the two groups. “Materials science has been doing pretty well to generate common databases,” she said. “On the manufacturing side, not really. This is something that we need to catch up.”

In addition, the field needs to continue developing new techniques to measure and characterize the process, an area Cao refers to as functional metrology. They need new ways of determining what kinds of flaws are permissible, beyond studying the melt pool geometry and the texture of the surface.

Symposium X—MRS/The Kavli Foundation Frontiers of Materials features lectures aimed at a broad audience to provide meeting attendees with an overview of leading-edge topics.


The comments to this entry are closed.