| About this Abstract |
| Meeting |
2024 TMS Annual Meeting & Exhibition
|
| Symposium
|
Materials Informatics to Accelerate Nuclear Materials Investigation
|
| Presentation Title |
Optimizing Thermal Conductivity Prediction of Uranium Compounds using Balanced Multiclass Classification |
| Author(s) |
Yifan Sun, Masaya Kumagai, Yuji Ohishi, Eriko Sato, Masako Aoki, Ken Kurosaki |
| On-Site Speaker (Planned) |
Yifan Sun |
| Abstract Scope |
In the pursuit of advanced nuclear fuels, a multiclass classification model was applied to predict the thermal conductivity of uranium compounds. We used element properties (Magpie) and temperature as features, with thermal conductivity data, sourced from the Starrydata2 database, classified into three ranges. However, the class distribution in the Starrydata2 database is skewed towards materials with low thermal conductivity, heavily biasing the model’s predictions towards lower values. This imbalance posed a challenge in identifying nuclear fuels with high thermal conductivity. To mitigate class imbalance, we employed the Synthetic Minority Oversampling Technique and Random Undersampling to balance the training data. As a result, the bias towards underestimating thermal conductivity was reduced, with recall for classes 1 (5-15 W/mK) and 2 (15+ W/mK) increasing from 0.33 and 0.64 to 0.49 and 0.71, respectively. By enhancing the model’s recall, our chances of discovering fuels with high thermal conductivity have significantly improved. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Nuclear Materials, Machine Learning, Characterization |