ABSTRACT Inorganic ABX 3 perovskites have exceptional optoelectronic properties, yet efficiently navigating their vast chemical space for high‐performance photodetectors remains challenged by the trade‐off between computational fidelity and screening efficiency. To overcome this bottleneck, we develop a hierarchical physics‐aware graph neural network (PAGNN) framework featuring a novel Dynamic RBF Modulation mechanism. Our model achieves deep coupling of chemical priors with geometric equivariance. Crucially, we implement a three‐stage curriculum learning strategy to bridge the fidelity gap between standard PBE and high‐precision mBJ functionals. For photodetector and optoelectronic applications, we predict band gaps and formation energies as well as hull energies, achieving mean absolute errors of 0.11 eV and 0.06 eV/atom, respectively. Validation Structure‐Unseen Split protocol demonstrates the model's superior structural inference capability. Applying our model to screen promising ABX 3 candidates and guide the continuous compositional engineering of solid solutions, we recover known high‐performance perovskites and identify novel compounds for optoelectronic integration. Ultimately, this work establishes a generalizable, data‐efficient, and interpretable paradigm for accelerated materials discovery.
Lei et al. (Fri,) studied this question.