Efficient Magnetic Property Prediction and Process Optimization by Combining Generative Adversarial Networks and Auto Machine Learning in Thin-Gauge Fe-Si Alloy Sheet
Source abstract
Abstract
In order to greatly increase the energy efficiency of the motor core of electric vehicles, commercial 0.5 mm non-oriented Si steel was re-rolled to a thin-gauge of 0.1 mm, which cannot be accomplished with commercial rolling. To maximize the flux density, a key factor affecting energy efficiency, the hysteresis AI model was optimized based on the heat treatment conditions. GAN, a data augmentation technology, and deepfake technology were used. The effectiveness of GAN and AutoML were explored as a way of increasing the consistency of the hysteresis AI model with fewer experiments. The AutoML model with FLAML based on the tree model was effectively improved, and the model performance was effectively improved on 4000 data per condition, similar to one condition of the existing data without data augmentation. The R2 score was 0.9827 and the MSE was 0.0213 for various temperature and time conditions, which is less than half of the MSE of 49% compared to the MSE before applying GAN. The heatmap analysis of the flux density model showed that the maximum flux density of B8 was 1.50T under these conditions, which was about 25% better than the lowest flux density (1.14T). The value was then confirmed by actual experiment. A maximum (1.50T) of B8 in 0.1 mm re-rolled Si steel was obtained, demonstrating that the combination of GAN and FLAML can improve the consistency of continuous data material models such as hysteresis curves when experiments are insufficient. This has sufficient potential for developing magnetic data GAN technology.
Source: crossref · Quality: baseline
Research context
OpenAlex topics: Magnetic Properties and Applications · Welding Techniques and Residual Stresses · Microstructure and Mechanical Properties of Steels
SciVX discovery tags are terminology-based. OpenAlex topics are source-provided classifications.
References (37)
Show source-provided references
1. Wang · 2022 · 9983407
2. 10.1016/s0304-8853(03)00142-210.1016/s0304-8853(03)00142-2 ↗
3. 10.1016/s1006-706x(10)60170-710.1016/s1006-706x(10)60170-7 ↗
4. 10.1016/j.matchar.2023.11282110.1016/j.matchar.2023.112821 ↗
5. 10.1016/j.mtcomm.2023.10570810.1016/j.mtcomm.2023.105708 ↗
6. 10.1016/j.scriptamat.2022.11514510.1016/j.scriptamat.2022.115145 ↗
7. 10.1016/j.commatsci.2024.11279310.1016/j.commatsci.2024.112793 ↗
8. Choudhary · 2022
9. 10.1016/j.jmat.2017.08.00210.1016/j.jmat.2017.08.002 ↗
10. 10.3365/kjmm.2023.61.5.37910.3365/kjmm.2023.61.5.379 ↗
11. 10.3365/kjmm.2024.62.11.92010.3365/kjmm.2024.62.11.920 ↗
12. Karras · 2006
13. Cui · 2022
14. Singh · 2020
15. Wang · 2023
16. 10.1016/j.commatsci.2023.11266110.1016/j.commatsci.2023.112661 ↗
17. 10.1016/j.jma.2023.12.00710.1016/j.jma.2023.12.007 ↗
18. Momin · 2024 · 4270
19. Wu · 2021 · 1654
20. Lu · 2020 · 123
21. 10.1007/s00170-021-07037-y10.1007/s00170-021-07037-y ↗
22. 10.1016/j.pmatsci.2024.10133210.1016/j.pmatsci.2024.101332 ↗
23. 10.1016/j.measurement.2020.10813510.1016/j.measurement.2020.108135 ↗
24. Wang · 2023 · 110093
25. Roussel · 2022 · 204801
26. Van de Wiele · 2010 · 103902
27. Abert · 2014 · 103902
28. Yamamoto · 1974 · 2
29. Javidi · 2016
30. Wen · 2025 · 015111
31. Li · 2017 · 123
32. Yu · 2025 · 4567
33. Wang · 2021
34. 10.1016/j.jobe.2023.10807110.1016/j.jobe.2023.108071 ↗
35. 10.1145/342262210.1145/3422622 ↗
36. 10.1109/tkde.2021.313019110.1109/tkde.2021.3130191 ↗
37. 2022