Calcium silicate-based cements (CSCs) are used in root-end filling, apexification, furcation repair, and vital pulp therapy to facilitate tissue vitality and promote cell differentiation and regeneration [1], [2]. The CSCs promote dental pulp stem cells differentiation by inducing the expression of key markers and pathways that regulate mineralization and tissue regeneration. These outcomes are connected to CSCs’ reactions that generate calcium silicate hydrates (C-S-H) and calcium hydroxide (Ca(OH)₂) [3]. These compounds regulate Ca²⁺ and OH⁻ release, progressively alkalizing the microenvironment, influencing calcium phosphate solubility, and promoting hydroxyapatite precipitation, which is essential for tissue regeneration [4], [5]. However, excessive or prolonged microenvironment alkalinity may induce the activation of apoptotic pathways and the accumulation of reactive oxygen species, affecting the bioactivity and clinical performance of CSCs [6], [7].

Modulating CSCs’ alkalinity over time is key for tissue responses. The alkalinity usually fluctuates between 6.0 to > 12.0 and is strongly influenced by the microenvironment and CSCs' composition [8]. For instance, higher contents of tricalcium silicate (C₃S) and dicalcium silicate (C₂S) enhance Ca²⁺ and OH⁻ release, increasing pH and maintaining an alkaline environment [9]. Conversely, additives such as titanium and zirconium dioxide (TiO₂ and ZrO₂) can dilute reactive phases or alter hydration dynamics. Also, excess aluminum-rich phases can alter hydration pathways [10], [11]. These mechanisms ultimately reduce Ca(OH)₂ formation, lower the availability of Ca²⁺ and OH⁻ ions and diminish the alkalizing effect [12]. Beyond composition, the ratio between material quantity and environmental volume also influences alkalinity. Such variations affect cell responses and increase the complexity and workload of research and development, thereby hindering formulation optimization and limiting the comparability of results [1]. Establishing a material’s pH profile is time-consuming and labor-intensive, as it typically requires multiple pH measurements over an extended period—commonly up to 28 days. Notably, the alkalinity profile of CSCs is highly sensitive to compositional variations, making formulation design experimentally demanding and time-consuming [13]. Therefore, if several compositions or experimental parameters need to be tested—often necessary to simulate clinical variability (e.g., varying specimen sizes or extended timepoints)—such testing further increases the time and resource required. Unfortunately, formulations can only be re-optimized and tested after the pH cycle profiling has been completed.

The sequential and labor-intensive process of establishing the pH profile of CSCs hinders the development of biomaterials and highlights the need for innovative research and development workflows. This requires predictive frameworks capable of modeling alkalinity behavior to facilitate rational design of compositions and expedite material development. Materials informatics (MI) emerges in this context, leveraging data-driven methodologies, including machine learning and statistical modeling, to optimize materials’ compositions and predict their properties and performance [14]. By leveraging multidimensional datasets, experimental data, computational simulations, and multivariate descriptors, MI reduces reliance on extensive empirical testing, streamlines formulation screening, and enables the design of biomaterials with predicted properties [15]. This capability is particularly relevant for CSCs, since pH varies dynamically and is interdependent on multiple factors that cannot be modeled using conventional techniques. Therefore, a machine learning model that can accurately predict long-term alkalinity profiles (up to 672 h) holds great potential to redude the need for extended measurements, streamlining workflows and accelerate CSCs development.

This study aimed to develop and validate a machine learning model using a stacking ensemble approach, integrating gradient boosting regressors (GBR) as the base model and a sequential multilayer perceptron (SMP) neural network as the meta-model, to predict the long-term alkalinity profile of CSCs using early-stage pH values (3 and 24 h) and specimen surface area as input features. We hypothesized that a data-driven model trained on such indicators could accurately predict pH evolution up to 672 h across diverse formulations, reducing the need for extended testing exercises.