Transformer-Type Architecture for Predicting Physicochemical Properties of Cosmetic Emulsions

Predicting the physicochemical behavior of cosmetic emulsions from formulation variables remains a challenging multiscale problem. In this work, twenty-seven emulsion formulations were prepared and characterized through rheological measurements and optical microscopy. A transformer-based model was trained using formulation descriptors as inputs and experimental curves such as flow sweep, frequency sweep, and droplet size distribution as targets. Transformer architecture achieved strong predictive performance with correlation coefficients exceeding 0.90 for most properties and low normalized mean absolute errors relative to the magnitude of the responses. The largest deviations were observed for the viscous modulus in the inference case, which is likely associated with experimental fluctuations arising from microstructural rearrangements. Compared with conventional regression approaches, the proposed architecture captures experimental behavior without relying on predefined mathematical parametrizations. These results highlight the potential of transformer-based models for data-driven formulation design and predictive modeling of complex cosmetic emulsions.