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

Introduction:

Refineries often need molecular-level information about petroleum fractions, but directly measuring every molecular detail is expensive and time-consuming. This work reconstructs the molecular composition of naphtha from a limited set of bulk properties, such as density and boiling point information. A feedforward neural network first learns how molecular composition maps to measurable properties. The trained model is then embedded inside an inverse optimization problem, which searches for a realistic composition that would produce the observed properties. To keep the answer chemically plausible, the method uses database-derived reference compositions, carbon number patterns, compound-type distributions, and entropy-based regularization. Interpretable analysis with SHAP helps explain why selecting key bulk properties improves the reconstruction.