About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process
|
| Presentation Title |
Microstructure-Aware Generative AI Model for Long-Term Spatiotemporally Consistent Prediction of Corrosion and Crack Evolution |
| Author(s) |
Lin Cheng, Yuhao Liu, Yao Fu |
| On-Site Speaker (Planned) |
Lin Cheng |
| Abstract Scope |
Corrosion and cracking in additively manufactured (AM) metals are governed by complex polycrystalline microstructures with multiscale interactions. Accurate long-term prediction is critical for reliability but remains challenging. Phase-field modeling offers high-fidelity simulation of corrosion and crack evolution, yet its high computational cost limits large-scale applications. We propose a generative AI framework that accelerates phase-field simulations by learning microstructure-dependent corrosion and crack dynamics. The model takes crystallographic microstructures as input and predicts full spatiotemporal evolution. A key innovation is time-embedding conditioning, enabling adaptive temporal modeling across corrosion- and crack-dominated regimes. We evaluate one-step, multi-frame, and sequence-to-sequence strategies, analyzing stability and accuracy. The model reproduces key phenomena, including corrosion front propagation, corrosion-to-crack transition, crack initiation, and branching, with strong consistency. It predicts full evolution for unseen microstructures within seconds, achieving orders-of-magnitude speedup while preserving essential physics, enabling rapid materials design and durability assessment. |