Accurate prediction of electrical conductivity for CaF₂-based multicomponent melts is hindered by conventional "composition → property" direct mapping approaches that ignore the mechanistic role of microstructure. This study proposes a "composition → microstructure → conductivity" paradigm that explicitly models microstructure as the physical bridge linking composition to transport properties. The framework integrates: (1) a classification-based hybrid sampling strategy achieving 96.5% compositional space coverage with only 62 molecular dynamics simulations; (2) a progressive microstructure prediction pipeline that iteratively predicts 22 structural descriptors (average R² = 0.924); and (3) a physics-constrained neural network embedding Arrhenius temperature dependence and network former/modifier effects as regularization constraints. The model achieves R² = 0.810 and MAPE = 9.71% for conductivity prediction—a 6.2% mean error within the industrially critical 1.5–4.0 S/cm range. SHAP interpretability analysis reveals that CaF₂ influences conductivity through a balanced dual mechanism: 57.5% direct carrier effect and 42.5% indirect microstructure modification, independently validated by molecular dynamics simulations. The paradigm is generalizable to other oxide–fluoride melt systems.
Chen et al. (Wed,) studied this question.