Machine learning applications to bioactive glasses: From traditional statistical models to cutting-edge deep learning paradigms

The challenges associated with the complexity of bioactive glasses (BGs) make traditional trial-and-error approaches impractical for developing compositions that are balanced with respect to bioactivity, degradation, mechanical reliability, therapeutic functionality, manufacturability, and translational safety. This article reviews the impact of machine learning (ML) on reshaping BG research and transition from empirical optimization to data-driven and physics-informed materials design. The article outlines the descriptors, endpoints, and validation needs associated with ML-enabled BG modelling with focus on the transition from statistical/shallow-learning approaches to advanced paradigms, including: physics-informed neural networks (PINNs); convolutional (CNNs), recurrent (RNNs), and graph neural networks (GNNs); generative adversarial networks; and self-supervised training approaches. The article highlights how ML is being applied to BGs for composition design and structure-property-function prediction. The main translation barriers for BGs are also identified, including: fragmented datasets; inconsistent metadata; limited prospective validation; poor extrapolation beyond compositional domains; black-box ML model behavior; and inadequate integration with multimodal experimental, simulation, and biological data. Finally, ML-related emerging opportunities are presented and discussed. By merging algorithmic advances with glass science and biomaterials requirements, this review provides a roadmap for shifting ML from a retrospective prediction tool to a rational design engine for next-generation BGs.