pH prediction in commercial and experimental calcium silicate cements via material informatics

To develop and validate predictive machine learning model capable of estimating long-term pH profiles (up to 672 h) of calcium silicate-based cements (CSCs) using early-stage pH measurements (3 and 24 h).

pH and calcium ion release data from in vitro studies (2014 - 2024) were extracted and analysed using descriptive statistics and correlation metrics. Feature selection was conducted using Random Forest regressors to identify key variables. A hybrid stacked ensemble model was built, integrating Gradient Boosting Regressors (GBRs) as base models and sequential multilayer perceptron as a meta-model. Model calibration involved polynomial regression and residual correction with GBRs. Predictive performance was evaluated using MAE, RMSE, R², and k-fold cross-validation. Experimental in vitro validation was conducted using four commercial and three experimental CSCs (specimen surface area: 113, 169 and 220 mm²), comparing actual and predicted pH values at 72, 168, and 672 h.

The model achieved strong predictive accuracy (R² = 0.91, 0.89, and 0.85 for 72, 168, and 672 h) with consistent performance across validation folds. Residuals showed no systematic bias, and Bland-Altman plots confirmed agreement. Experimental validation demonstrated a strong correlation (R² > 0.80), with no statistically significant differences across time points or specimen surface areas. The model generalized well across commercial and experimental formulations.

The machine learning model was able to predict the alkalinity evolution of CSCs based on early pH measurements and specimen surface area. The approach reduces the need for prolonged testing and large specimen numbers, supporting biomaterials development and the design of next-generation endodontic materials.