Comparison of Machine Learning Approaches for Prediction of the Equivalent Alkane Carbon Number for Microemulsions

This repository accompanies the research article:

Nicholas R. Furth, Adam E. Imel, and Thomas A. Zawodzinski
Published in The Journal of Physical Chemistry A, 2024.
DOI: 10.1021/acs.jpca.4c00936

Overview

This project explores machine learning (ML) methods for predicting the Equivalent Alkane Carbon Number (EACN) of oils based on molecular structure.
Experimental determination of EACN is traditionally time-consuming and purity-sensitive. We developed models to predict EACN quickly and reliably using:

• Graph Neural Networks (GNNs)
• XGBoost Decision Tree Model

Key highlights:

• Three GNN architectures: MPNN, CGCNN, and GCN
• Geometry optimization studies (None, MMFF94, GAFF)
• Feature-based XGBoost using RDKit descriptors

Models Implemented

Model	Input Type	Special Features
MPNN	SMILES (with/without geometry optimization)	Deep message passing
CGCNN	SMILES (with/without geometry optimization)	Crystal graph convolution
GCN	SMILES (with/without geometry optimization)	Spectral convolution
XGBoost	Molecular descriptors from SMILES	Recursive feature elimination

Dataset

• Size: 183 organic molecules
• Data Source: Published experimental EACN values from literature references.
• SMILES Input: Obtained via PubChem or cheminfo.org.

Repository Structure

/MatDeepLearn/           #Our Implementation of MatDeepLearn for GNNs
/Model Results           #Meta Data for our Model Results
Descriptors.ipynb        #XGBoost Implementation
Error Plots.ipynb        #Plot Generation for Results
SSBSSW.ipynb             #F-Statistic Generation for Results (Functional Groups)
Descriptors.py           #RDKit Descriptor Generation called by the Notebook files
README.md            
requirements.txt  # Python dependencies

Installation

git clone https://github.com/nfurth1/EACN.git
cd EACN
pip install -r requirements.txt

Citation

If you use this work, please cite:

@article{furth2024eacn,
  title={Comparison of Machine Learning Approaches for Prediction of the Equivalent Alkane Carbon Number for Microemulsions Based on Molecular Properties},
  author={Nicholas R. Furth, Adam E. Imel, and Thomas A. Zawodzinski},
  journal={The Journal of Physical Chemistry A},
  volume={128},
  number={32},
  pages={6763-6773},
  year={2024},
  publisher={American Chemical Society},
  doi={10.1021/acs.jpca.4c00936}
}