Comparison between Multi-Linear- and Radial-Basis-Function-Neural-Network-Based QSPR Models for The Prediction of The Critical Temperature, Critical Pressure and Acentric Factor of Organic Compounds

Abstract

Critical properties and acentric factor are widely used in phase equilibrium calculations but are difficult to evaluate with high accuracy for many organic compounds. Quantitative Structure-Property Relationship (QSPR) models are a powerful tool to establish accurate correlation between molecular properties and chemical structure. QSPR multi-linear (MLR) and radial basis-function-neural-network (RBFNN) models have been developed to predict the critical temperature, critical pressure and acentric factor of a database of 306 organic compounds. RBFNN models provided better data correlation and higher predictive capability (an AAD% of 0.92–2.0% for training and 1.7–4.8% for validation sets) than MLR models (an AAD% of 3.2–8.7% for training and 6.2–12.2% for validation sets). The RMSE of the RBFNN models was 20–30% of the MLR ones. The correlation and predictive performances of the models for critical temperature were higher than those for critical pressure and acentric factor, which was the most difficult property to predict. However, the RBFNN model for the acentric factor resulted in the lowest RMSE with respect to previous literature. The close relationship between the three properties resulted from the selected molecular descriptors, which are mostly related to molecular electronic charge distribution or polar interactions between molecules. QSPR correlations were compared with the most frequently used group-contribution methods over the same database of compounds: although the MLR models provided comparable results, the RBFNN ones resulted in significantly higher performance.

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

“…Therefore, it is necessary to obtain baseline models using our data set for a fair comparison. We compare our models with the group contribution methods developed by Joback and by Nannoolal et al, and a radial basis function neural network with Morgan fingerprints based on the models from the literature . The radial basis function layer is imported from the referenced GitHub repository and implemented using the GATv2 framework.…”

“…In addition, several existing models in the literature have problems such as minor data leakage caused by using a test set to determine early stopping, 30,31 unclear data set selection criteria, 24,25 and lack of information to enable retraining. 30 Therefore, it is necessary to obtain baseline models using our data set for a fair comparison. We compare our models with the group contribution methods developed by Joback 78 and by Nannoolal et al, 22 and a radial basis function neural network with Morgan fingerprints based on the models from the literature.…”

“…Several ML models − have been developed to predict various subsets of critical properties and acentric factor, but a single ML model that can predict all four properties has not yet been reported. All ML and GC models in the literature have been exclusively tested on random test-training splits, which assumes that chemical property prediction is an interpolation problem.…”

Performance Prediction and Working Fluid Active Design of Organic Rankine Cycle Based on Molecular Structure

“…The molecular descriptors were selected through the enhanced repla method. The results showed that the average absolute deviations were 3.42% and Banchero et al [24] used a radial basis function neural network and multiple linear sion to establish QSPR model of critical temperature, critical pressure, and acentr of organic compounds. The results showed a strong relationship between the pr and descriptors characterizing electron charge distribution in the molecule.…”

Quantitative structure–property relationship for the critical temperature of saturated monobasic ketones, aldehydes, and ethers with molecular descriptors

“…The QSPR approach has been applied to forecast many physicochemical properties, such as normal boiling point [3–5], thermal conductivity [6], retention time [7], critical micelle concentration (CMC) [8], standard enthalpies [9], toxicity degree [10], lower flammability limits [11], piC50 [12] and logP [13], and so forth. In the past 30 years, the relationship between the critical temperature and molecular descriptors of compounds has aroused widespread interest among chemists [14–24].…”