Decoding Structure–Optical Property Relationships in TiO₂ Nanocomposites through Hybrid Features Integration

Optical titanium dioxide nanocomposites have a wide range of applications, but ill-defined structure–property relationships pose a challenge to the systematic discovery of functional materials. We designed a TGEML multimodal nanocomposite processing framework to address this challenge. The framework consists of a polymer multimodal featurizer named TGEML-polymer and a nanoparticle-specialized featurizer named TGEML-nano. TGEML-polymer integrates a Transformer-based graph neural network specialized in representing polymer information to capture the chemical semantic sequence and graph information of polymers. TGEML-nano includes processing strategies for nanoparticle dispersion scattering and particle size distribution to provide a comprehensive picture of the nanosystem. We correlated polymer feature vectors with nanosystem parameters to generate hybrid features. This enabled the successful construction of a quantitative “structure–refractive index” mapping. Moreover, the structure–optical property relationships were decoded through a multilevel feature interpretability analysis. In conclusion, TGEML can provide powerful and useful hybrid features for artificial intelligence to accelerate the development and design process of nanocomposites.