Python script for predicting glass forming ability (GFA) - Physics-informed machine learning of glass-forming ability - integrating thermodynamic limits and quantum signatures - research data

Accelerated discovery of oxide glasses is hindered by the combinatorial complexity of multicomponent systems and the inability of standard "black-box" Machine Learning (ML) models to respect thermodynamic phase boundaries. In this work, a Physics-Informed Indirect Framework is introduced, shifting the paradigm from direct regression of stability indices to the prediction of fundamental characteristic temperatures (Tg, Tx, Tliq). This approach is empowered by a novel hierarchical feature engineering architecture that fuses statistical entropy metrics with quantum signatures (orbital s/p/d/f occupancies), enabling the distinction between network formers and modifiers at a sub-atomic level. Validated on the complex Na2O - B2O3 - SiO2 system, the immiscibility gap was autonomously identified by the model as a mathematical singularity in the derived Hruby parameter (KH) - a feat unachievable by direct models, which smoothed over this "forbidden zone." Furthermore, the non-linear Mixed Alkali Effect (MAE) was successfully captured, and "stability hallucinations" in data-sparse regions were eliminated by the indirect framework through the use of the liquidus temperature as a physical safety fuse. Finally, it was confirmed via SHAP analysis that predictive decisions are driven by enthalpy and orbital topology rather than spurious correlations. This methodology bridges the gap between data science and glass physics, offering a robust tool for navigating the stability landscape of disordered materials.