作者
Narasak Pandech,Lappawat Ngamwongwan,Sirichok Jungthawan,Wutthigrai Sailuam,Ittipon Fongkaew
摘要
Cationic doping of hydroxyapatite (HAp, Ca 10 (PO 4 ) 6 (OH) 2 ) is widely used to tune its mechanical, electronic, and thermodynamic properties for biomedical applications, yet periodic-table-wide screening has been impractical by first-principles methods alone. Here we employ cross-domain machine learning, in which Random Forest models trained on the AFLOW and MatBench databases predict the bulk modulus, band gap, and formation energy for 58 non-radioactive dopant elements substituted at the Ca site, with undoped HAp as the reference composition (bootstrap of 0.81 to 0.95 on the training data). Transferability to HAp is tested by density functional theory (DFT): 58 relaxed 44-atom cells (29 dopants at both Ca sites, 10 mol%) supply the site preference, substitution energy and band gap that a composition-only model cannot resolve, and elastic tensors for ten doped cells, each dopant at its preferred site and in the lowest-energy spin state found, show that none of the ten dopants tested stiffens HAp at 10 mol%: every one of them lies between 70.6 and 80.1 GPa against 80.0 GPa for the host, with Zr, Mg, Sr and Sc within 0.4 GPa of it. The ML model predicts 101 to 107 GPa for all of them, a systematic offset of GPa, and does not reproduce the DFT ordering (Spearman , ); Ti, which it ranks first, softens the lattice, and charge-compensated supercells soften Ti and Zr further (53.4 and 73.7 GPa). The DFT stiffness tracks the distortion of the dopant’s oxygen polyhedron ( ) rather than its formal charge or electronegativity, which composition-only features cannot represent. The raw bulk-modulus ranking therefore requires DFT confirmation, whereas the prediction that Mg and Sr leave the host stiffness intact is confirmed. A biomedical figure of merit incorporating toxicity and formation energy ranks Sr first across 37 weighting scenarios, followed by Mg , Ba and, more sensitively, Y and Sc ; DFT confirms that Sr, Mg and Sc preserve the host stiffness. SHAP analysis identifies electronegativity as the model’s universal descriptor.