Predicting Single-Substance Phase Diagrams: A Kernel Approach on Graph Representations of Molecules

Area of Science:

• Computational chemistry
• Materials science
• Chemical engineering

Background:

• Predicting thermodynamic properties of pure substances is crucial for chemical process design.
• Existing methods often lack accuracy or transferability across different phases.
• Accurate property prediction aids in material discovery and process optimization.

Purpose of the Study:

• To develop a novel graph representation for molecules to predict thermodynamic properties.
• To implement a Gaussian process regression (GPR) model utilizing this graph representation.
• To assess the accuracy and reliability of the GPR model for single, double, and triple phase properties.

Main Methods:

• A transferable molecular graph representation was developed as input for a marginalized graph kernel.
• Gaussian process regression (GPR) models were employed, incorporating radial basis function kernels for temperature and pressure.
• The model predicted critical temperature, vapor-liquid equilibrium (VLE) density, and pressure-temperature density for pure substances.

Main Results:

• The GPR model achieved prediction accuracy comparable to experimental measurement precision.
• The model demonstrated reliability through quantifiable posterior uncertainty for each prediction.
• The proposed graph representation and GPR approach outperformed Morgan fingerprints and graph neural networks.

Conclusions:

• The novel molecular graph representation coupled with GPR offers a highly accurate and reliable method for predicting thermodynamic properties.
• This approach provides a robust tool for computational chemistry and materials science.
• The ability to quantify prediction uncertainty enhances the practical applicability of the model.