پيش‌بيني دماي كوري مواد مغناطيسي با يادگيري ماشين: مهندسي توصيف‌گرها، شبكه‌هاي عصبي گراف و اهميت داده‌هاي پالايش‌شده

Materials with magnetic properties play a fundamental role in modern technologies, including data storage, spintronic systems, sensors, an‎d electromagnetic devices, an‎d hold a special position in materials science an‎d engineering. One of the key parameters determining the performance of these materials is the Curie temperature, which indicates the phase transition from the ferromagnetic to the paramagnetic state an‎d depends directly on the chemical composition, electronic structure, an‎d atomic interactions. Accurate prediction of this quantity has always been a major challenge in designing new magnetic materials. In this study, a machine learning-based framework is proposed to predict the Curie temperature, an‎d the effects of data quality, feature engineering, an‎d advanced model applications are systematically investigated. The results show that using refined an‎d reliable data has a significant impact on improving the performance of prediction models. In the feature extraction process, two main approaches, including composition-based descriptors an‎d graph neural networks, are employed. Graph neural networks, by modeling compounds as feature-based graphs of elements, enable learning of combinatorial interactions an‎d nonlinear dependencies among elements, an‎d, in combination with ensemble learning methods, enhance the predictive power of the models. Feature importance analysis indicates that features related to the electronic structure, especially ionization energies, play a decisive role in predicting the Curie temperature. Overall, the results of this study demonstrate that integrating high-quality data, targeted feature engineering, an‎d advanced machine learning models can accelerate the design an‎d discovery of new magnetic materials with desired Curie temperatures.

مجتبي اعلائي , فرهاد شهبازي دستجرده

جلال ذهبي , حميده شاكري پور , علي صادقي