| Title |
Magnetic Property Prediction and Heat Treatment Optimization of Ultra-Thin Non-Oriented Electrical Steel via DDPM-Based Data Augmentation and AutoML |
| Authors |
성민수(Minsu Sung) ; 오지은(Jieun Oh) ; 김성진(Sung-Jin Kim) ; 이수석(Sooseok Lee) ; 박세민(SeMin Park) |
| DOI |
https://doi.org/10.3365/KJMM.2026.64.7.670 |
| ISSN |
1738-8228(ISSN), 2288-8241(eISSN) |
| Keywords |
Diffusion model; Data augmentation; Non-oriented electrical steel; Magnetic properties; AutoML |
| Abstract |
Optimizing the magnetic properties of 0.1 mm ultra-thin non-oriented electrical steel is essential for improving electric vehicle (EV) traction motor efficiency. However, traditional experimental trial-and-error methods incur massive time and costs and suffer from limited data availability. To overcome these limitations, this study aims to confirm the effectiveness of data augmentation without additional experiments, enhancing modeling accuracy for magnetic property optimization using limited data. An integrated artificial intelligence (AI) pipeline combining Denoising Diffusion Probabilistic Model (DDPM)-based data augmentation with the FLAML framework was investigated for predicting the steel's magnetic flux density (B) and hysteresis loss (Wh). A systematic three-step validation strategy was conducted, which included initial performance evaluation, augmentation scale optimization, and final interpolation performance evaluation on specific evaluation conditions within the explored domain. Consequently, the boosting-based LightGBM was chosen as the final model for magnetic flux density prediction using large-scale augmented data and a 300-second computation time. For hysteresis loss prediction, the ascending and descending branches of the loop were trained separately, selecting Extra Trees and CatBoost as optimal algorithms to minimize prediction errors. Based on the selected models, a grid search within the process parameter space identified 1,050°C for 16 minutes as the optimal process among candidates satisfying the target magnetic flux density (B25 >= 1.5T). This condition achieved a magnetic flux density of 1.6143 T while minimizing energy loss to 409.18 J/m3. These predictions showed an error of less than 2% against actual experimental measurements. In this study, the data-driven optimization framework significantly reduces the number of experiments required to develop high-performance electrical steel compared to conventional trial-and-error approaches, demonstrating its potential for broad application in alloy design and microstructure control. |