MDs-NP: a property prediction model construction procedure for naphtha based on molecular dynamics simulation

Area of Science:

• Petrochemical industry
• Molecular modeling
• Machine learning applications

Background:

• Molecular management is crucial for petrochemical industry goals like carbon neutrality.
• Accurate oil property prediction is essential for molecular reconstruction.
• Naphtha property prediction requires novel approaches.

Purpose of the Study:

• To propose a novel property prediction model construction procedure (MDs-NP) for naphtha.
• To utilize molecular dynamics simulations and machine learning for property prediction.
• To identify effective molecular descriptors for naphtha property modeling.

Main Methods:

• Employed molecular dynamics simulations to calculate 348 sets of naphtha mixture properties.
• Utilized gamma distribution for calculating mole fractions from real analytical data.
• Extracted molecular features using open-source toolkits (RDKit, Mordred) and designed 12 novel naphtha knowledge (NK) descriptors.
• Applied and compared support vector regression, extreme gradient boosting, and artificial neural network algorithms.

Main Results:

• Achieved best density prediction using Mordred and NK descriptors with support vector regression.
• Achieved best viscosity prediction using RDKFp and NK descriptors with artificial neural network.
• Identified T, P_w, and CC(C)C as key NK descriptors enhancing model performance via ablation studies.

Conclusions:

• The MDs-NP procedure effectively predicts naphtha density and viscosity.
• The developed models facilitate rapid molecular reconstruction for data-driven petrochemical processes.
• MDs-NP shows potential for extension to other properties and complex petroleum systems.