About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Uncertainty Quantification in Ultra-High Temperature Materials Manufacturing
|
| Presentation Title |
Autonomous Materials Characterization Through Simulation to Experiment Analysis with Continual Deep Learning |
| Author(s) |
Zihan Wang, Hyun Sang Park, Ali Rachidi, David Elbert, Todd Hufnagel, Wei Chen |
| On-Site Speaker (Planned) |
Zihan Wang |
| Abstract Scope |
Autonomous laboratories for materials discovery require learning systems that can interpret characterization data reliably, adapt to evolving experimental conditions, and continuously incorporate new information. X-ray diffraction (XRD) is a key characterization technique for probing crystal structure and phase evolution, but its interpretation remains challenging due to the complexity and variability of diffraction patterns. We present a continual deep learning framework for 2D XRD image classification in autonomous laboratory settings. The framework combines a spectral-normalized neural Gaussian process (SNGP) for uncertainty quantification with a domain adaptation strategy that bridges simulated and experimentally collected XRD data under realistic conditions. By updating the model incrementally as new data arrives, the method reduces the need for full retraining, improving computational efficiency and memory usage while enabling faster adaptation to newly observed materials. Results on sequential datasets show strong classification performance and a meaningful latent feature space. |