تركيب محاسبات نظريه تابعي چگالي و يادگيري ماشين براي خوشه‌بندي مواد پادفرومغناطيس

With the rapid growth of public materials datasets an‎d the emergence of various feature generation tools, extracting meaningful physical patterns from these datasets by using artificial intelligence techniques has become an important issue in materials science. Machine learning approaches have been increasingly applied to various classes of materials—such as semiconductors, solar cells, catalysts, an‎d ferromagnetic materials—highlighting the growing an‎d ongoing expansion of these techniques across different domains of materials science. However, antiferromagnetic materials appear to have received limited attention in this context. In this research, an integrated, reproducible, an‎d interpretable framework for the unsupervised clustering of antiferromagnetic materials is presented. Its innovation lies in the systematic connection established between feature selec‎tion an‎d the physical eva‎luation of clusters during the assessment phase. To eva‎luate stability an‎d generalizability, two independent datasets were used: one containing 980 materials an‎d an extended set containing 1,732 materials. Materials were described using four parallel an‎d comparable feature sets, an‎d the effect of different scaling methods was systematically examined. Subsequently, eight clustering algorithms with different mechanisms—including centroid-based, density-based, graph-based, an‎d self-organizing map methods—were implemented an‎d compared using intrinsic eva‎luation metrics such as silhouette score, Davies–Bouldin index, an‎d Calinski–Harabasz index. In the present study, feature selec‎tion is not merely a preprocessing step but rather a decision-making pillar in unsupervised machine learning problems. Based on this premise, a silhouette-driven strategy was developed an‎d employed for feature selec‎tion. The results indicate that substantial dimensionality reduction can be achieved alongside simultaneous improvement in clustering quality an‎d preservation of structural stability. Additionally, Neel temperature regression was introduced as a complementary metric to eva‎luate the physical homogeneity of clusters independently of geometric separation. This research provides a general an‎d reproducible framework for analyzing unlabeled data in materials science, which can serve as a basis for extracting physical knowledge, rapid material screening, an‎d guiding experimental studies in complex systems.

جواد هاشمي فر

مجتبي اعلائي

اسماعيل عبدالحسيني سارسري , حامد بخشيان