About this Abstract |
Meeting |
1st World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022)
|
Symposium
|
First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022)
|
Presentation Title |
Generating Realistic Material Microstructures Using Conditional GANs for Advanced Manufacturing |
Author(s) |
Scott Howland, Kevin Loew, Henry Kvinge, Xiaolong Ma, Joshua Silverstein, Nicole Overman, Md. Reza E Rabby, Scott Taysom, Tianhao Wang, WoongJo Choi, Scott Whalen, Keerti Kappagantula, Luke Gosink, Tegan H. Emerson |
On-Site Speaker (Planned) |
Tegan H. Emerson |
Abstract Scope |
Advanced manufacturing technologies enable material scientists to produce materials with increasingly attractive physical properties. Because they are novel, intuitions for what experimental parameters produce superior materials don’t necessarily transfer from older processes. And since advanced processes are commonly limited in scale, it is cost and time-prohibitive to conduct extensive research into how experimental parameters influence production outcomes. A machine learning pipeline used to extrapolate from a limited set of experiments could help practitioners judge what future experiments might yield the best-performing materials, accelerating innovation and decreasing cost. As a step towards such a pipeline, we explore the use of generative adversarial networks (GANs) conditioned on practitioner-specified experimental process parameters to generate plausible microstructure imagery. We show that GANs are able to produce realistic microstructures for a variety of temper conditions, propose next steps for incorporating more conditioning information, and show how this approach could fit into an end-to-end scientific pipeline. |
Proceedings Inclusion? |
Undecided |