Modeling of titanium blast-finished surface roughness for dental implants using strictly grouped validation and SHAP-explainable machine learning

Predictable control of implant surface roughness is important because a consistent micrometre-scale topography may provide reproducible conditions for early osteoblast attachment and subsequent osseointegration. To address the low efficiency of traditional trial-and-error optimization and its limited ability to characterize the nonlinear relationship between sandblasting parameters and surface roughness (Ra) in pure titanium substrates for oral implants, this study developed a machine learning-based predictive model. Grit size (60–100 mesh), Pressure (0.2–0.4 MPa), Distance (50–100 mm), and Time (15–60 s) were selected as input variables. A total of 180 datasets were obtained by combining orthogonal experiments with random experiments. Ten machine learning algorithms, including ensemble learning, nonlinear, and linear models, were compared. To avoid the apparent data leakage caused by random splitting of repeated measurements, Leave-One-Group-Out Cross-Validation (LOGO-CV) was used to evaluate the model’s predictive performance and stability onunseen process combinations. The results showed that, within the investigated parameter range, the Ra values of the titanium surfaces ranged from 1.083 to 2.494 μm. This range lies between the approximate dimensions of osteoblast cell bodies and their filopodial extensions and may provide local anchorage sites for filopodia, thereby facilitating early mechanical interlocking at the bone–implant interface. Comparative analysis showed that the support vector regression (SVR) model achieved the best predictive performance. After Bayesian optimization, the optimized SVR model achieved an R² of 0.7538 and an RMSE of 0.1480 μm. Analysis of variance and SHAP-based interpretability analysis confirmed that grit size and blasting pressure had significant effects on Ra ( P < 0.05). The relative importance of the parameters was ranked as follows: grit size > blasting pressure > blasting time > blasting distance. Within the tested parameter range, the proposed model may facilitate sandblasting-parameter screening and improve the consistency of implant surface preparation, thereby providing process-level support for the reproducible production of surface conditions relevant to early osseointegration and implant stability and establishing a basis for subsequent biological validation and clinical translation.