Metal halide perovskites have emerged as promising candidates for advanced photovoltaic and optoelectronic devices; however, their commercial deployment is significantly limited by the toxicity and chemical instability of lead-based compositions . All-inorganic lead-free alternatives, particularly cesium tin mixed-halide perovskites, (CsSn(ClxBryI1−x−y)3) provide a viable pathway to environmentally benign and thermally stable materials . However, identifying experimentally realizable compositions remains challenging due to the vast combinatorial compositional space and the persistent discrepancy between theoretical stability and practicalsynthesis.This study develops a physics-informed and explainable machine learning framework to predict the synthesizability and thermodynamic stability of mixed-halide tin perovskites . A Positive-Unlabeled learning strategy, utilizing transductive bagging and decision tree classifiers, estimates synthesis probability based on literature-derived data [5]. A 76-dimensional descriptorset, incorporating compositional, elemental, structural, and density functional theory (DFT)-derived thermodynamic parameters, is constructed . Gradient boosting regression modelsfurther predict macroscopic properties, while SHAP analysis provides physical interpretability of the prediction.The model identifies a limited subset of compositions with a high probability of successful synthesis and demonstrates that the advanced tolerance factor (𝜏), B-site atomic characteristics, and local lattice distortion are dominant stability descriptors [9, 10]. Machine-learned atomic potentials enable effective mapping of temperature-driven structural phase properties, overcoming the limitations of ab initio molecular dynamics . This framework bridges computational predictions and experimental feasibility, establishing a general data-driven methodology that accelerates the discovery pipeline for stable, non-toxic, lead-free perovskite materials.