About this Abstract |
| Meeting |
MS&T25: Materials Science & Technology
|
| Symposium
|
Materials Informatics for Images and Multi-Dimensional Datasets
|
| Presentation Title |
Bidirectional Prediction of Microstructure–Property/Process Relationships in Advanced Structural Materials Using Deep Generative Models |
| Author(s) |
Xiaofan Zhang, Junya Inoue, Satoshi Noguchi |
| On-Site Speaker (Planned) |
Xiaofan Zhang |
| Abstract Scope |
Our research presents a unified deep learning framework to establish both forward and inverse process-structure-property (PSP) linkages using a single model. The framework learns a shared latent representation of microstructures that connects processing conditions and/or material properties by combining a vector-quantized variational autoencoder with a pixel convolutional neural network. This shared space enables not only the generation of statistically equivalent microstructures from given processing parameters or properties, but also the identification of suitable microstructure features to achieve target conditions. The methodology captures the stochastic nature of heterogeneous microstructures while preserving physical interpretability. Demonstrated on advanced structural materials, the framework significantly enhances the efficiency and flexibility of data-driven materials design. It underscores the potential of integrating generative models with materials science principles to support both predictive modelling and inverse optimization within a single, coherent system, offering a scalable solution for addressing inverse design challenges in complex microstructural systems. |