Interpretable Multimodal Graph Learning Platform for Rational Design of AIEgens: From Molecular Structure and...

Aggregation-induced emission luminogens (AIEgens) exhibit significant application potential in materials science due to their unique photophysical properties. However, systematically elucidating their structure-property relationships remains challenging due to the high dispersion of data, the complex correlations of features, and the limited interpretability of traditional machine learning models. Herein, we constructed a data-driven and interpretable deep learning model (referred to as GATM) that integrates multisource data from the literature, including molecular structures, photophysical parameters, and solvent environments. By integrating graph neural networks with machine learning algorithms, this multimodal predictive framework successfully deciphers the intricate relationships between molecular structural features, solvent environments, and photophysical properties. The visualization of solvent-solute interaction mechanisms was achieved through multilevel attention capture and feature quantification analysis utilizing the graph attention network (GAT). Furthermore, the GAT also provided deep insights into the influence of key structural features-such as atomic type and hybridization state-on the luminescence mechanisms of AIEgens. The results demonstrate that GATM achieves high predictive accuracy (mean R² > 0.90) for key parameters of AIEgens, including fluorescence lifetime, quantum yield, and maximum absorption/emission wavelengths. Subsequent molecular synthesis experiments further validated the model's predictive accuracy. Furthermore, the synthesized molecules underwent organic pesticide detection and discrimination experiments, achieving a low detection limit (0.4 nM) and 100% discrimination accuracy. This intelligent prediction platform provides a novel paradigm for the rational design of new AIEgens and paves the way for future inverse design research for other functional materials.