An Integrated Framework Combining Graph Neural Networks, Physicochemical Modeling, and Optimization for Fragrance Reproduction

Reproducing olfactory experiences remains a major challenge, with implications for heritage science, sensory analysis, and the fragrance industry. This work presents an integrated perfume-engineering framework that combines chemical engineering principles with computational modeling to reproduce a target scent. The methodology quantifies the olfactory profile of a reference fragrance, predicts the evaporation behavior of its constituent compounds, and combines first-principles thermodynamic modeling with machine learning and optimization to identify molecular substitutes and propose a cost-efficient, sustainable reproduction mixture. The framework includes quantitative characterization of the target scent, estimation of gas-phase contributions through evaporation dynamics, generation of candidate molecules, and solution of an optimization problem to determine mixture composition. The approach is tested on a perfume formulation with known composition reported in the literature. Results show that 1,2-dimethoxybenzene can act as an effective and cost-effective substitute for vanillin while closely matching the target olfactory profile. Overall, the study demonstrates the potential of integrated computational methods for rational fragrance design.