Property Prediction of Light Oil Fractions

Machine learning method for predicting physicochemical properties of light oil fractions

This repository contains the GauL-HDAD-derived algorithm for representing chemical mixtures as a numeric vector and neural networks for mixture property prediction.

Installation

This package is written in python and can be run from command prompt. This requires the installation of several python packages, which are listed below. It is recommended to use Anaconda for managing packages (https://docs.anaconda.com/anaconda/install/).

Python 3.8 or higher
NumPy conda install numpy
RDKit conda install -c conda-forge rdkit
Scikit-learn conda install -c conda-forge scikit-learn
joblib conda install -c anaconda joblib
TensorFlow 2 pip install tensorflow
matplotlib conda install -c conda-forge matplotlib

The package can then be installed by cloning this repository.

Data

The data that is used in the original work is available in the folder Data. Five files are included:

labeled_library.pickle A library with all C₄ to C₁₂ PIONA molecules and their experimental data (if available)
mei_input.pickle The input values (lumped naphtha samples) from Mei et al.
mei_output.pickle The output values (boiling points and mixture properties) from Mei et al.
pyl_input.pickle The input values (lumped naphtha samples) from Pyl et al.
pyl_output.pickle The output values (boiling points and mixture properties) from Pyl et al.

Changing the input and output sources is possible in the file input.py.

Training

Pretrained Model

This repository contains a folder named pretrained. In that folder, pretrained pure compound property models are available and the gaussian mixture models for the molecular representations are premade using all molecules in Data/labeled_library.pickle. Using the pretrained model speeds up training, since only the boiling points and the desired mixture property has to be trained. Training mixture properties using the pretrained model is possible via following command: python train.py pretrained <property> Replace with the mixture property that you want to predict.

Currently available properties:

Using data from Pyl et al.:
Specific Gravity: sg >>> python train.py pretrained sg
Using data from Mei et al.:
Liquid Density: d20 or density >>> python train.py pretrained d20 or python train.py pretrained density
Dynamic Viscosity: mu or viscosity >>> python train.py pretrained mu or python train.py pretrained viscosity
Surface Tension: st or surface tension >>> python train.py pretrained st or python train.py pretrained "surface tension"

All results will be found in the pretrained folder.

New Model

It is also possible to train all models yourself. Due to the large number of hydrocarbons, the creation of Gaussian mixture models will take several hours.
You can train the models simply using python train.py <folder> <property>
Replace with the name of the folder where you want to store your results.
Replace with the desired property, as stated above.

How to cite?

When using this prediction model for your own publication, please cite the original papers:

Learning Molecular Representations for Thermochemistry Prediction of Cyclic Hydrocarbons and Oxygenates
Dobbelaere, M.R.; Plehiers, P.P.; Van de Vijver, R.; Stevens, C.V.; Van Geem, K.M.
J. Phys. Chem. A 2021, 125, 23, 5166–5179

Machine Learning for Physicochemical Property Prediction of Complex Hydrocarbon Mixtures
Dobbelaere, M.R.; Ureel, Y.; Vermeire, F.H.; Stevens, C.V.; Van Geem, K.M.
Submitted to Industrial and Engineering Chemistry Research, 2022

mrodobbe / naphtha-mixtures