Search

Machine learning is a powerful tool that can predict the iron loss of electromagnetic steels with high accuracy. However, there is no theoretical explanation as to why machine learning can make such accurate predictions; it is a black box. Therefore, it is unclear how to select an algorithm for predicting iron loss behaviors among many machine learning algorithms available. In this study, we used approximately 10000 iron loss data for electromagnetic steels under stress and 19 machine learning algorithms to evaluate and compare the learning algorithms in terms of prediction accuracy, robustness, the number of data required, avoidance of overfitting, and agreement with experimental results. As a result, we conclude that LightGBM is the best algorithm for predicting the iron loss properties of electromagnetic steels. Although the discussion in this study has not yet led to a theoretical breakthrough in the field of machine learning, we hope that it will provide an effective guideline for selecting the learning algorithm in materials science. [ABSTRACT FROM AUTHOR]

At present, almost 50% of electrical power is consumed by motors. Thus, increasing the efficiency of motors is an important issue. To achieve more efficient operation, it is vital to improve the accuracy of input data for motor loss design. In this study, we focused on the iron loss of electromagnetic steels, which is assumed to account for a large proportion of motor losses, and examined whether the effect of stress on the iron loss and its frequency dependence could be predicted with high accuracy by machine learning. First, experimental iron loss data are obtained at flux densities of 0.1-1.7 T, frequencies of 50-3000 Hz, and applied stresses from 200 to 200 MPa. No significant deterioration in iron loss behavior is observed in specimens subjected to 3% and 10% pre-strain by tensile loading. These data show that the effect of stress on iron loss varies significantly depending on the excitation conditions. The complex iron loss behaviors are the result of interplay between magnetic wall movement and magnetic domain rotation during the magnetization process. As simple regression of the magnetization process is difficult, we apply three machine learning algorithms to the experimental dataset. The results show that the LightGBM algorithm produces the most accurate predictions of the experimental iron loss values. The contributions of the explanatory variables are found to be consistent with empirical knowledge. This study demonstrates the potential for machine learning to enable improve the accuracy of iron loss data input to motor loss design. [ABSTRACT FROM AUTHOR]