About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing: Equipment, Instrumentation and In-Situ Process Monitoring
|
| Presentation Title |
AMDiffusion: Domain-Adaptive Diffusion Modeling for Causal Data Fusion in Additive Manufacturing |
| Author(s) |
Hyunwoong Ko, Fatemeh Elhambakhsh, Suk Ki Lee, Zhuo Yang, Yan Lu, Ho Yeung |
| On-Site Speaker (Planned) |
Hyunwoong Ko |
| Abstract Scope |
Digital twins (DTs) in additive manufacturing (AM) rely on understanding complex process-structure-property relationships. While fusing multimodal AM data can capture these dynamics, existing models struggle with pronounced spatio-temporal variations across different builds and domains. To address this gap, we introduce AMDiffusion, a novel domain-adaptive methodology combining diffusion-based generative modeling with causal graph networks. AMDiffusion uniquely fuses AM process parameters (PP) and in-situ process signatures (PS) while explicitly enforcing their physical causal dependencies, overcoming limitations of conventional data fusion. The framework jointly learns interactions between dynamic PP and PS features to generate physically consistent representations. We demonstrate AMDiffusion’s robust adaptability through two distinct case studies: laser powder bed fusion (LPBF) and aerosol jet printing (AJP). By capturing hidden causal linkages, AMDiffusion improves the simulation, prediction, and optimization of diverse AM operations, catalyzing advancements in predictive DTs. |
