Leveraging spatial charge descriptor in deep learning models: Toward highly accurate prediction of vapor-liquid equilibrium

Background

Vapor-liquid equilibrium (VLE) data is essential for separation processes, but traditional models such as UNIFAC require extensive experimental parameters. To improve the efficiency of VLE modeling and reduce the dependence on experiments, we developed machine learning models using COSMO-based σ-profiles and MACCS keys.

Methods

Two deep learning models, MLP-COSMO and MLP-MACCS, are developed. MLP-COSMO, based on σ-profiles, allows high-precision phase equilibrium predictions using only molecular structures, eliminating the need for experimental interaction parameters. The pressure range spans 0.947 to 817 kPa and the temperature range spans 199.93 to 548.15 K.

Significant findings

By utilizing σ-profiles, our developed model MLP-COSMO, which surpasses COSMO-SAC (2010) in accuracy and achieves a level comparable to UNIFAC, with specific results of R²-y = 0.9926, R²-P = 0.9889, AAD − y(%) = 1.04% and AARD − P(%) = 2.88%, evaluated on a test dataset excluded from training process. This study successfully demonstrated that high-precision VLE predictions can be achieved using only a molecular structure, effectively addressing the challenge of missing experimental parameters. Furthermore, the results indicate that the spatial charge descriptor (σ-profile), encapsulating molecular polarity information, is considered to be more suitable as input data for machine learning models in VLE prediction than structural fingerprints of MACCS.