About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Late News Poster Session
|
| Presentation Title |
H-36: Data-Efficient Deep Generative Modeling for Bidirectional Process–Structure–Property Relationships With Uncertainty Quantification |
| Author(s) |
Xiaofan Zhang, Junya Inoue, Satoshi Noguchi |
| On-Site Speaker (Planned) |
Xiaofan Zhang |
| Abstract Scope |
This study introduces a data-efficient deep generative framework that establishes bidirectional process–structure–property (PSP) linkages in structural materials. The model combines a Vector-Quantized Variational Autoencoder (VQ-VAE) and a Pixel Convolutional Neural Network (PixelCNN) to encode complex microstructures into discrete latent codes and generate realistic alloy-specific structures conditioned on processes or properties such as toughness. By combining with Bayesian inference, it also performs inverse prediction of processes or properties from given microstructures with uncertainty quantification. Compared with a conventional Convolutional Neural Network and Gaussian Process Regression (CNN+GPR) approach, the proposed framework achieves comparable or better accuracy using much smaller datasets while uniquely supporting both forward and inverse predictions. Visualization of the latent space confirms physically interpretable representations. Overall, this VQ-VAE + PixelCNN framework advances data-driven and uncertainty-aware materials design, moving beyond property prediction toward understanding, designing, and generating new materials with quantified confidence. |
| Proceedings Inclusion? |
Undecided |
| Keywords |
Machine Learning, Modeling and Simulation, Computational Materials Science & Engineering |