Machine Learning-Based Prediction of Magnetic Properties from Hysteresis Curves: A Comparative Study of Random Forest, Gradient Boosting, Xgboost, Lightgbm and Support Vector Regressions

Abstract

This study investigates the application of machine learning regression models for predicting magnetic properties directly from hysteresis curves. A dataset of experimentally measured hysteresis loops was preprocessed and used to train five algorithms: Random Forest, Gradient Boosting, XGBoost, LightGBM, and Support Vector Regression (SVR). A dataset of 561 samples, each containing 800 resampled features and five target variables, was divided into training and testing sets to evaluate model performance using R², MAE, and RMSE metrics. Results revealed that Gradient Boosting and XGBoost consistently achieved superior accuracy, with Gradient Boosting excelling in intrinsic coercivity (R² = 0.959) and XGBoost dominating in remanence (R² = 0.921), coercivity, and residual induction. Support Vector Regression demonstrated specialized strength in saturation moment (R² = 0.753), though it underperformed in intrinsic coercivity. The study recommends prioritizing Gradient Boosting and XGBoost for multi-target regression tasks.