About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Autoregressive latent diffusion for 3D polycrystalline material dynamics |
| Author(s) |
Alexander Scheinker, Reeju Pokharel |
| On-Site Speaker (Planned) |
Alexander Scheinker |
| Abstract Scope |
Fast generative surrogate models present an opportunity to see a virtual view of a complex material’s internal state in real-time during experiments, providing useful information that can inform real-time experimental feedback tuning. In this work, we present results on development of an autoregressive latent diffusion model that rolls out predictions of all 24 fields (crystal orientations and micromechanical fields) of the time-varying 3D state of a polycrystalline material. We give an overview of using 3D variational autoencoders to compress large 3D volumes into much smaller latent representations and an overview of how a generative diffusion model is then used to learn the dynamics directly in the reduced latent space. We demonstrate the roll-out ability of this process for varying prediction lengths and quantify its ability to accurately predict the various fields for FFC structured materials. We also study the generalization ability of this approach to make predictions for unseen structures. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Mechanical Properties, Modeling and Simulation |