Revealing time-regime-dependent defect formation pathways in perovskite single crystals through physics-informed machine learning.

The mechanisms governing defect formation in perovskite single-crystal growth remain elusive due to the complex interplay of kinetic, thermal, and compositional factors, which vary across different growth methods and durations. Here, we develop an explainable machine learning framework using XGBoost to decode these mechanisms across three physically distinct growth scales. By integrating physically informed descriptors with SHAP and partial dependence analyses, we uncover a systematic evolution of dominant control factors: from nucleation-growth balance at a short time scale, through thermo-structure regulation at an intermediate time scale, to environmental stability at a long time scale. Our results demonstrate that optimal growth conditions are not universal but shift with the dominant physical process, providing a method-agnostic, interpretable strategy for defect-minimized single-crystal growth. This study demonstrates an interpretable machine learning strategy for discovering growth mechanisms of different time scales and provides guidance for optimizing single crystal growth conditions.