| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Fracture in Metals: Insights from Experiments and Modeling Across Length and Time Scales
|
| Presentation Title |
3D Surrogate Model Training Using Active Learning with Elasto-Viscoplastic FFT Simulations of Pore Morphologies from Laser Powder Bed Fusion of Ti64 |
| Author(s) |
Daniel Diaz, Xingyang Li, Elizabeth Holm, Anthony Rollett |
| On-Site Speaker (Planned) |
Daniel Diaz |
| Abstract Scope |
Additive manufacturing is revolutionizing the way we manufacture products, but properties are limited by porosity produced during processing. To better understand the relationship between pore morphologies and fatigue we use an elasto-viscoplastic FFT simulation called MASSIF to calculate stress and strain field outputs and fatigue harmfulness rankings. To overcome the computational costs of simulating loading on hundreds of thousands of pores, we developed a 3D Deep Learning surrogate model that directly predicts the MASSIF stress and strain field outputs given a pore morphology. In our approach we incorporate Active Learning with Monte Carlo Dropout to train the model by selectively labeling pores from the unlabeled pool with MASSIF as the oracle labeler. This approach prioritizes labeling pore instances that the model is most uncertain of while using clustering to maintain diversity in the growing training dataset. Our work opens the capability for complex physical simulation on big data. |