PNNL

The Science

Evaluating new ideas and potential improvements for manufacturing is typically too expensive and time consuming for most companies to tackle. As a result, available data is scarce. Researchers at the Pacific Northwest National Laboratory (PNNL) are hoping to mitigate these obstacles by applying conditional generative adversarial networks (GANs) to scanning electron microscope (SEM) imagery from shear assisted processing and extrusion (ShAPE), an emerging advanced manufacturing process.

Researchers use GANs, a machine learning (ML) model where two neural networks compete, to improve the accuracy of their predictions. As generative models, GANs can also create new data samples that resemble the training, or sample, data. SEM imagery is used to closely assess the surface of materials used in manufacturing, enabling the characterization of microstructures from a topological perspective.

Combining the use of GANs, SEM imagery, and ShAPE, a team of researchers at PNNL is looking to establish reliable frameworks and approaches that make research and development improvements across material systems and advanced manufacturing processes easier to implement.

The Impact

In 2021, manufacturing contributed $2.3 trillion to U.S. gross domestic product (GDP), which is about 12 percent of total U.S. GDP and more than 14 million workers. With such a large impact on the economy, the U.S. Department of Energy has invested in a national plan to revitalize American manufacturing.

The work by the team at PNNL takes a critical step in leveraging ML to accelerate advanced manufacturing research and development, specifically for manufacturing techniques without access to efficient, first-principles simulations.

Summary

Material microstructures play a key role in identifying parameters in manufacturing and a material’s performance. The team at PNNL developed generative models trained on SEM images of the aluminum alloy AA7075, which is commonly used in aircraft and bikes due to its high strength and toughness, tubes manufactured using the ShAPE technology. ShAPE-synthesized parts have unique microstructures with minimal porosity and unparalleled performance.

They generated realistic images conditioned on temper and either experimental parameters or material properties, ultimately comparing synthetic images with experimental ones. The team's ML solution can be integrated into the development cycle by allowing users to immediately visualize the microstructures that can result from particular process parameters or desirable properties.

The researchers found that the synthetic image distributions do not uniformly align with experimental SEM images. The dissimilarity between experimental and synthetic images was consistent with the small number of unique experiments present in most advanced manufacturing datasets.

This work forms the technical backbone for a fundamentally new approach to understanding manufacturing processes. In the future, the team plans to use differentiable data augmentation, developments in GAN regularization, and invertible neural networks to better leverage limited advanced manufacturing data and increase model sample efficiency.

PNNL Contact

Sylvia Howland
Pacific Northwest National Laboratory
Sylvia.Howland@pnnl.gov, (509) 371-6016

Funding

The research project was funded by PNNL’s Mathematics for Artificial Reasoning in Science (MARS) initiative, which focuses on machine learning to solve important scientific problems and national security challenges.