Data-Driven Discovery of Calcium--Phosphate Biomaterial Scaffolds Using Materials Informatics and Machine Learning Clustering

Soegianto Soelistiono (Department of Physics, Faculty of Science and Technology, Universitas Airlangga, Surabaya 60115)
Mohd Kamarulzaki Mustafa (Photonic Devices and Sensor Research Group, Faculty of Applied Sciences and Technology, Pagoh Education Hub, Universiti Tun Hussein Onn Malaysia, Muar,Johor 84600, Malaysia)

Indonesian Applied Physics Letters
Vol. 7 No. 1 (2026): Indonesian Applied Physics Letters - June 2026

The discovery of biomaterial scaffolds for bone tissue engineering remains challenging due to the vast compositional space and the limitations of conventional experimental approaches. In this study, a data-driven and AI-assisted framework is developed to identify promising calcium--phosphate biomaterial candidates using materials informatics. A total of 508 compounds were retrieved from a computational materials database and systematically screened based on thermodynamic stability, density, and electronic properties, resulting in 129 biomaterial-relevant candidates. An initial multi-parameter scoring model combined with unsupervised learning techniques reveals that electronic properties, particularly band gap, strongly influence material differentiation. However, this dominance introduces descriptor bias, potentially limiting the physical realism of the screening results. To address this limitation, a balanced multi-objective scoring model is introduced, incorporating density, formation energy, and energy above hull to achieve a more physically meaningful evaluation. The refined model consistently identifies Ca3(PO4)2 as a top candidate, in agreement with its well-established role in bone tissue engineering, thereby providing validation of the proposed approach. Comparative analysis between the initial and balanced models reveals significant ranking shifts, demonstrating that descriptor balancing substantially affects material prioritization. Furthermore, a trade-off is observed between clustering separability and physical interpretability, highlighting the limitations of purely statistical evaluation metrics in materials screening. Overall, this study demonstrates that integrating materials informatics with machine learning enables efficient and scalable biomaterial discovery. More importantly, it shows that correcting descriptor bias through multi-objective balancing is essential for achieving reliable and physically meaningful results. The proposed framework provides a reproducible pathway for identifying next-generation scaffold materials for biomedical applications.