|About this Abstract
||2018 TMS Annual Meeting & Exhibition
||Computational Design and Simulation of Materials (CDSM 2018): Computational Design of Materials
||Computational Thermodynamic and Machine Learning Approach
to Accelerate the Design of High-temperature Alloys
||Dongwon Shin, Sangkeun Lee, Yukinori Yamamoto, Michael P Brady
|On-Site Speaker (Planned)
We present a data-driven theoretical framework that has potential to significantly reduce the number of prototypical alloys for experimental validation. We exploit machine learning techniques to predict creep of alumina-forming austenitic (AFA) alloys with and without microstructural descriptors. First, we train machine learning models to predict rupture time of AFA alloys as a function of composition, stress, and temperature. We have used experimental alloy/creep dataset consistently collected over a decade at ORNL. Next, we exercise a high-fidelity CALPHAD database in a high-throughput manner to rapidly populate various microstructural features. We then fuse the dataset with over two hundred features to augment ‘raw’ alloys dataset followed by extensive correlation analysis to select features to be used within the training. We demonstrate that trained machine learning models can predict creep properties of hypothetical AFA alloys in a high-throughput manner.
Research was sponsored by the LDRD Program of Oak Ridge National Laboratory.
||Planned: Supplemental Proceedings volume