Machine learning prediction for magnetic properties of Sm-Fe-N based alloys produced by melt spinning

Abstract

In this paper, we applied machine learning (ML) to the materials processing of Sm-Fe-N magnetic powders produced by melt spinning. Neural Networks (NN) were used to generate ML models. Data were collected from relevant papers and patents, amounting to more than 800 data entries. These data were composed of chemical compositions, melt spinning process parameters and heat treatment parameters as the input layers, and magnetic properties as the output layers. ML models were developed by adjusting hyperparameters. Subsequently, agreement between the derived and actual values was found, and tended to reduce as the values moved further from the area with a large amount of data. The ML models were utilized to predict the properties of some test Sm-Fe-Co-N samples with Nb, Ti, or Zr additions. The predicted data were roughly in agreement with the actual data, except for the samples which had extrapolated data.