Predicting Critical Properties and Acentric Factors of Fluids Using Multitask Machine Learning.

Area of Science:

• Computational Chemistry
• Chemical Engineering
• Materials Science

Background:

• Accurate critical properties (critical temperature, pressure, density) and acentric factors are vital for calculating thermo-physical properties of chemical compounds.
• Experimental determination of these properties is costly and time-consuming, hindering efficient chemical process design and analysis.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting critical properties and acentric factors directly from chemical structure (SMILES representation).
• To explore various ML architectures and featurization techniques to optimize prediction accuracy.

Main Methods:

• Utilized directed message passing neural networks (D-MPNN) and graph attention networks for molecular representation learning.
• Incorporated additional atomic and molecular features, multitask training, and pretraining with estimated data.
• Employed a multitask learning scheme to predict multiple properties (critical properties, acentric factor, boiling point, melting point, enthalpy of vaporization, enthalpy of fusion) simultaneously.

Main Results:

• The developed D-MPNN model, augmented with Abraham parameters, achieved state-of-the-art accuracies in predicting critical properties and acentric factors.
• Multitask training improved the model's ability to predict a range of thermo-physical properties.
• Evaluation on random and scaffold splits demonstrated the model's robustness and generalization capability.

Conclusions:

• The ML model provides an efficient and accurate method for predicting critical properties and acentric factors, reducing reliance on experimental data.
• The publicly released dataset of 1144 compounds and source code will facilitate further research in computational property prediction.