| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Characterization of Minerals, Metals and Materials 2025: In-Situ Characterization Techniques
|
| Presentation Title |
Deep Learning Based Reconstruction From Sparse 2D Projection Datasets for In Situ Tensile Experiments |
| Author(s) |
Nathan S. Johnson, Orion Kafka, Hrishikesh Bale, Steve Kelly, Newell Moser, Jake Benzing, Jason Kilgore |
| On-Site Speaker (Planned) |
Nathan S. Johnson |
| Abstract Scope |
Laboratory X-ray characterization often suffers from lower flux and brightness compared to synchrotron sources, resulting in longer counting times for high-quality images. This is particularly challenging for in situ experiments requiring multiple datasets, often extending to days per sample. We present a deep-learning-based reconstruction method for sparse X-ray microscopy datasets. In an in situ tensile test on additively manufactured Inconel 718 dogbones, 1000 2D projections were collected in an hour. A commercial deep learning model (Zeiss DeepRecon) trained on this dataset generated high-resolution 3D reconstructions from only 100 projections, reducing acquisition time by 10x. The model accurately captured up to 80% of internal pores and maintained the morphology of the largest pores even with reduced data. This reduction from 1 hour to 6 minutes per load step significantly enhances the efficiency of in situ experiments, enabling higher quality data collection on more samples in less time. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Mechanical Properties, Characterization |