Integrating artificial intelligence with experimental and computational materials science to predict crystallization in multicomponent glasses

Crystallization profoundly influences the performance of glasses, governing their stability in nuclear waste forms, manufacturability in industrial systems, and functionality in glass-ceramics. However, the predictive models remain elusive due to the structural complexity of multicomponent glasses. Here, we establish a data-driven framework that integrates systematic experimental datasets, atomistic simulations and machine learning to predict crystallization in multicomponent alkali aluminoborosilicate glasses designed in the primary crystallization field of nepheline (NaAlSiO4). Molecular dynamics models, validated against multinuclear MAS NMR spectra, reveal medium-range order and cluster motifs that are structurally similar to nepheline. From this, we derived a structural descriptor—the mean cumulative displacement (MCD)—that quantifies structural similarity between the glass and nepheline crystal. When combined with optical basicity, MCD enables robust separation of crystallizing and non-crystallizing glasses. A support vector machine (SVM) classifier trained on MCD and OB achieved ≥94% accuracy across 88 glass compositions, yielding interpretable decision boundaries that link chemistry, structure, and crystallization behavior. While the present study focuses on nepheline crystallization, the methodology provides a structure-based framework that could be extended to other crystalline phases and glass families in future work.