Agrin Febrian Pradana, Intan Septia Sari, Andreas Federico, Rio Sudwitama Persadanta Kaban, Yusril Yusuf, Donanta Dhaneswara, Muhammad Taufik, Hartatiek, Iis Sopyan, Jaka Fajar Fatriansyah
The ability to fine-tune and control the crystallinity of hydroxyapatite (HAp) is essential for its clinically applications. One of the effective methods to alter HAp crystallinity is by introducing ion doping. In this research, machine learning (ML) methods of K-nearest neighbor (KNN), CatBoost, XGBoost, and artificial neural networks (ANN) were used to predict the crystallinity of HAp under influence of various doping ions of strontium (Sr2+), zinc (Zn2+), silver (Ag+), fluoride (F−), nickel (Ni2+), Iron (Fe3+), Erbium (Er3+), boron (B3+), aluminum (Al3+), barium (Ba2+), tungsten (W6+), magnesium (Mg2+) and sintering temperature and time. Although pH, precursor type, and synthesis method are important, the pre-screening model built is focused on doping ion and sintering parameters due to the complete data availability. The results reveal that CatBoost and XGBoost performed well in prediction performance and their reliability, with maximum R2 scores of 83.9 % (k-fold of 83.0 %, k = 5) and 95.6 % (k-fold of 81.9 %, k = 5), respectively. Further tests of Diebold-Mariano showed that the there are no significant difference of prediction performance between XGBoost and CatBoost. However, XGBoost is preferable over CatBoost due to its higher k-fold performance than CatBoost and its success is attributed to its ability to handle small, bimodally distributed datasets. The feature analyses demonstrated that Mg2+ ion doping significantly reduces HAp crystallinity. Our finding also showed that the addition of Mg2+ and Zn2+ ion doping to Hap is preferable to control crystallinity which was confirmed by DFT study. These findings validate the use of ML models for pre-screening ion-doped HAp, offering an efficient tool for optimizing its properties for clinical applications. Despite these promising results, the model has limitations originating from the small, bimodal dataset and the exclusion of other crucial synthesis parameters, which may not capture the full complexity of ion doping effects on HAp. © 2025 The Authors
Department of Metallurgical and Materials Engineering, Faculty of Engineering, Universitas Indonesia, Kampus UI, Jawa Barat, Depok, 16424, Indonesia; Department of Physics, Faculty of Mathematics and Natural Sciences, Universitas Gadjah Mada, Yogyakarta, 55281, Indonesia; Department of Chemistry, Faculty of Mathematics and Natural Sciences, Universitas Sumatera Utara, Sumatera Utara, Medan, 20155, Indonesia; Department of Physics, Faculty of Mathematics and Natural Science, Universitas Negeri Malang, Jl. Semarang 5, Malang, 65145, Indonesia; National Research and Innovation Agency (BRIN), Jalan M.H. Thamrin No. 8, Jakarta, 10340, Indonesia
