Semantic Device Graphs for Perovskite Solar Cell Design

Materials science faces two persistent challenges: the multiscale nature of functional devices, where performance emerges from the complex interplay of components across different length scales, and the prevalence of incomplete characterization data that precludes conventional featurization approaches. These challenges are exemplified in perovskite solar cells, where device optimization requires consideration of multiple interacting layers while much of the materials data exists only as text descriptions. While machine learning has accelerated the discovery of isolated material properties, translating promising materials into functional devices remains a significant bottleneck. Here, we introduce \textit{semantic device graphs}: a physics-inspired representation that captures the multi-scale architecture of perovskite solar cells while leveraging large language models to generate meaningful embeddings for incomplete material descriptions. Our approach achieves a 10\% improvement in performance prediction compared to state-of-the-art methods (CrabNet), enabling holistic device optimization rather than isolated material screening. The framework generates physically meaningful device fingerprints that reveal patterns in high-performing architectures, providing insights for future device optimization. This work demonstrates how combining physics-informed architectural choices with language models can address fundamental materials science challenges of multiscale modeling and incomplete information, serving as a stepping stone toward more holistic materials discovery approaches.