Quantitative Structure–Property Relationship Predictions of Critical Properties and Acentric Factors for Pure Compounds

Abstract

Knowledge of critical constants and phase boundary pressure properties is essential to understanding thermodynamic behavior of substances and is often required in practical process design applications. Where critically evaluated data are unavailable, a quantitative structure−property relationship (QSPR) regression method can be used to relate molecular properties (descriptors) to properties of interest.The relationship is trained and tested using existing critically evaluated data and is dynamic; as new data become available, the relationship can be updated to reflect changes. In this work, we use support vector regression (SVR) to develop estimation methods for critical properties and acentric factors based on critically evaluated data for over 900 pure compounds. From threedimensional geometry and connectivity information, we calculate over 500 descriptors for each compound. A matrix of descriptor values defines the input vectors for SVR, whereas critically evaluated data for critical temperature, the ratio of critical temperature to critical pressure, and saturation reduced pressure form the targets. We determine optimal SVR parameters by minimizing the sum of absolute deviations between the SVR outputs and the target values. We use a genetic algorithm to find the Pareto front points that optimize the output fit while reducing the number of input vectors (descriptors). We use a single Pareto front point to make a final evaluation in SVR. To define uncertainties of predicted values, we use uncertainty propagation calculations based on a Monte Carlo method that employs Latin hypercube sampling.

Low-GWP refrigerants for medium and high-pressure applications

“…The PubChem database does not provide these data for the vast majority of the compounds, so they had to be estimated using novel methods developed within this study (Kazakov et al ., 2012, Carande et al ., 2014, Carande et al ., 2016); the development and application of these estimation methods constituted a major effort. These methods were based solely on the molecular structure.…”

Machine Learning Enabled Quickly Predicting of Detonation Properties of N‐Containing Molecules for Discovering New Energetic Materials

“…Recently, with the development of artificial intelligence and machine learning, neural networks have been employed to efficiently predict molecular properties based on corresponding structures. [ 13–16 ] Atomization energies, [ 17,18 ] bandgap of inorganic solids, [ 19 ] bond energies, [ 20 ] dielectric breakdown strength in polymers, [ 21 ] critical point properties of molecular liquids [ 22 ] and exciton dynamics in photosynthetic complexes [ 23 ] were predicted by various neural networks with small errors. Recently, some researchers used machine learning as a tool to predict the properties of energetic materials and screen energetic molecules.…”

Flexible heuristic algorithm for automatic molecule fragmentation: application to the UNIFAC group contribution model

“…The idea behind this method is to divide the value of a property of the complete molecule into its contributions based on the chemical groups or other molecular subunit. Group contribution models have been successfully applied to a wide variety of properties including density [1, 2], critical properties [3–5], enthalpy of vaporization [6], normal boiling points [7, 8], water–octanol partition coefficients [9–11], infinite dilution activity coefficients [12] and many more. Also, from Gibbs excess energy models [13–15] and equations of states [16–19] they provide an approach that allows widening their application range to molecules composed of the same chemical groups relatively easily.…”