Molecular composition reconstruction of naphtha fractions through data-driven modeling and interpretable optimization

Molecular composition reconstruction models for petroleum fractions is important for the application in the process upgrading of modern refineries. In the traditional data-driven reconstruction framework, the underdetermined mapping from low-dimensional bulk properties to high-dimensional molecular compositions and the limited interpretability remain prominent challenges. In this study, a novel strategy is proposed for molecular composition reconstruction of naphtha fractions by integrating data-driven modeling with interpretable optimization. A feedforward neural network (FNN) is pre-trained to map molecular composition to bulk properties. For each sample to be reconstructed, the parameterized FNN is then embedded into an inverse optimization problem to reconstruct the molecular composition. Furthermore, a sample-specific reference composition is constructed from the database to guide the reconstruction, along with carbon number distribution and structural distribution profiles, ensuring chemical plausibility and similarity to realistic petroleum compositions. The reconstructed compositions demonstrate high accuracy. Model interpretability analysis using SHapley Additive exPlanations (SHAP) demonstrates that selecting key bulk properties facilitates the establishment of a robust and unbiased mapping between composition and properties. Furthermore, a broader set of bulk properties are predicted by a supplementary extreme gradient boosting model using limited key properties and its reconstructed composition as inputs.