About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Novel Strategies for Rapid Acquisition and Processing of Large Datasets From Advanced Characterization Techniques
|
| Presentation Title |
Discovering Hidden Fingerprints in Multimodal Process-Structure-Property Data via Joint Embedding |
| Author(s) |
Brad L. Boyce, Remi Dingreville |
| On-Site Speaker (Planned) |
Brad L. Boyce |
| Abstract Scope |
Combinatorial high-throughput datasets were collected for two manufacturing processes: (1) physical vapor deposition of thin nanocrystalline films, and (2) laser powder bed fusion additive manufacturing of a Kovar alloy. In both cases, large data collection campaigns led to substantial process-structure-property (PSP) data cubes. The experimental PSP data was fused with simulation predictions of the material deposition process. To analyze these large, complex datasets we employed several modern multimodal joint embedding techniques including an off-the-shelf algorithm, OpenAI’s CLIP, and custom Sandia-developed algorithm, a physics-informed multimodal autoencoder. In this presentation we will walk through several different analysis scenarios, ranging from Bayesian multi-objective optimization to aging predictions for as-deposited materials. Saliency assessment through Sobol sensitivity analysis reveals the relative importance of different information modalities in the prediction stream.
SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Additive Manufacturing, ICME, Machine Learning |