Machine learning predictive models for initial dissolution rate of silica-based glass composition from 1 to 35 components

Predicting the initial dissolution rate of silica-based glass compositions is challenging due to high compositional complexity, nonlinear behavior, and limited data. This study applies machine learning (ML) to model initial dissolution rates ( r 0 ) for multicomponent silica-based glasses containing up to 35 components. Model interpretability was examined through prediction profiling and feature importance analysis. We distinguish two evaluation strategies: interpolation within seen compositions, used for interpretation and prediction of unseen compositions, used to assess true generalization. Results show that extreme gradient boosting (XGBoost) performs better for predicting unseen compositions within the training range compared to other ML models. Under stratified validation, neural network (NN) models fit the high-dimensional composition space most completely and provide the most reliable basis for interpreting oxide effects. The important analysis showed that pH, temperature, SiO 2 and B 2 O 3 are the most dominant features. By separating interpretation from generalization and quantifying both, this study clarifies the key factors controlling dissolution behavior and provides a robust framework for data-driven prediction of r 0 in silica-based glasses.