An accurate classification of material stability often requires fusing multiple features under uncertainty. Dempster–Shafer (D–S) theory is a powerful framework for multi-source information fusion under uncertainty. However, its effectiveness critically depends on the quality of basic probability assignments (BPAs), which are typically assigned heuristically. To overcome this limitation, we proposed a classification model that integrates genetic algorithm (GA)-optimized kernel density estimation (KDE) with a weighted D–S fusion strategy. The GA automatically selects the optimal kernel function and bandwidth for KDE, enabling data-driven and accurate BPA construction without manual parameter tuning. The proposed method is first validated on benchmark datasets (Iris, Wine, Ionosphere, and Hepatitis), achieving competitive or superior performance compared to the existing methods. More importantly, when applied to predict the thermodynamic stability of double perovskite halide materials, our method achieves 93.7% accuracy and 85.3% precision under 10 × five-fold cross-validation, substantially outperforming CatBC (76.9% precision) and XGBC (75.0% precision). Notably, the model maintains robust performance even for compositions containing chemical elements absent from the training set, demonstrating strong transferability. These results highlight the potential of our GA-KDE-DS framework as a practical tool for accelerating the discovery of novel functional materials under limited data and uncertain conditions.
Liang et al. (Thu,) studied this question.