Quantitative structure property relationship schemes for estimation of autoignition temperatures of organic compounds

Abstract

We have extended a quantitative structure–property relationship (QSPR) scheme to estimate the auto-ignition temperatures (AIT) of organic compounds by employing GA-ANFIS, PSO-ANFIS, DE-ANFIS and GP methods. The average absolute relative deviations (%AARD) are 7.96, 6.29, 8.85 and 8.26, respectively. The range of these values appears to match that of experimental error in AIT measurements, suggesting strong models. For organic compounds, the AIT can be estimated using the above-mentioned methods, from molecular structure. This goal is possible using only 9 theoretical descriptors.

Introduction

Physicochemical data have a broad applicability in chemical industry, engineering and risk assessment calculations. In process design, in order to accurately estimate the consumption of utility, cost and the dimensions of the equipment, correct data should be used for material and energy balances. Therefore; valid and precise values of physicochemical properties are always crucial. However, in the actual industrial processes not all the essential data are available. This issue occurs especially for the properties involved with combustion; among these are flash point, auto-ignition temperature, and flammability limits [1].

The definition of auto-ignition temperature is as follows: the minimum temperature at which a substance ignites when there is no external flame or spark. According to the thermal theory of ignition and classical reaction-rate theory, a mixture that is combustible should be heated to auto-ignition temperature in order to ignite. Hence, when there is a leakage of a combustible liquid, the auto-ignition temperature of a substance is one of the important parameters in assessing the possible aftereffects. This is a significant issue in risk assessment methods, for instance API-581 [2]. We can also use engine knocking to describe auto-ignition in the context of combustion engines performance [3]. Based on the above-mentioned definition of auto-ignition temperature, it depends on the characteristics of the substance as well as the apparatus and the method(s) we use to determine it. The test pressure, physical properties of the vessel, and oxygen concentration are examples as such. Hence, the value of the auto-ignition temperature of a compound may not be reported the same in different literature. These differences can vary within 300 K range [[4], [5], [6], [7], [8]]. One of the cause of these discrepancies is the different experimental methods by which AIT values are measured [3]. However, the AIT values are normally reported without mentioning any detailed information of the implemented experiments. For instance, finding the data reported in SAX's dangerous properties of industrial materials or the hazardous chemical substances data bank is not possible. Even the data quality of AIT used in the famous DIPPR® project, is still flagged as “unevaluated”. Therefore; we cannot include experimental methods as additional descriptive variables, nor can we group the variables by different experimental methods used. Moreover, the flame appears instantly inside the auto-ignition vessel and we detect that by visual inspection, therefore, determining auto-ignition temperature significantly depends on human error [9]. In experimentally measured auto-ignition temperature values, the average error is around ±30 K, as reported in the literature [10]. Another issue is that it is not always convenient to determine AIT experimentally [9]. That being the case, estimating AIT values using mathematical modeling is a cost-efficient method.

One can also use the group contribution method in order to determine auto-ignition temperature, which is very convenient to apply. We can use the chemical structure of compounds in order to calculate the required parameters, i.e. the numbers of each group. Using QSPR method is currently common in predicting the physicochemical properties. Molecular descriptors are theoretical parameters that are derived only using molecular structures. These parameters describe the properties of compounds that are related to their structure and can be related to the properties of interest by QSPR method. The inputs of QSPR are these molecular descriptors which are not always easy to calculate. However, there are some obvious advantages to the QSPR method. To begin, the above-mentioned descriptors have certain physical meanings. It is very useful to find the physicochemical information which contributes significantly to the properties of interest. Furthermore, by comparing the group contribution and QSPR methods, considering the same studied dataset, we can conclude that the number of descriptors used in the latter method (usually <10 [11]) is normally less than those of used in the former method. On top of that, theoretically speaking, we should be able to apply the QSPR model to any organic compound. The reason is that only the molecular structure is used to derive the involved theoretical descriptors. QSPR methods are able to predict different parameters, such as physicochemical properties. Melting point, boiling point, vapor pressure, critical properties, water-octanol coefficients and water solubility are a few examples worth mentioning [[11], [12], [13]].

It should also be noted that the QSPR/QSAR studies use linear and non-linear modeling approaches. Linear approaches include principal component regression (PCR), partial least-squares (PLS) regression and multivariate linear regression (MLR). On the other hand, examples of non-linear approaches are as follows: support vector machines (SVMs), artificial neural networks (ANN), genetic programming (GP) and adaptive neuro fuzzy inference systems (ANFIS). Most of the non-linear approaches are categorized as soft computing techniques, which are used in various fields [[14], [15], [16], [17]]. Several groups use evolutionary or metaheuristic algorithms such as particle swarm optimization (PSO) or genetic algorithms (GA) along with linear techniques such as MLR which is used in the optimization or variable selection, according to Bagheri et al. [18] and Valadi et al. [19]. Suzuki [20] predicted the AIT of 250 pure compounds using a QSPR model. They took six descriptors into account resulting in an average error and a correlation coefficient (R) of 4.5% and 0.95, respectively. Two ANN modeling approaches were done by Tetteh et al. [10] using 233 compounds. The bases of this approach are the back-propagation function (BPF) as well as the radial basis function (RBF). Years later, support vector machines were used by Pan et al. [21] in order to develop another type of QSPR model by considering two databases. There were 50 alkenes and 142 organic compounds in the first and second sets, respectively, which included amines with cyclic, saturated, and unsaturated structures, aromatics, halogenated aliphatics, ethers, alcohols, esters and ketones. They compared their results to those of multivariate linear regression and artificial neural networks models. For the first case, the variables were 6 numbers of atom-type electro-topological-state indices, normally abbreviated as ETSI. They stated that these indices can be used to merge both topological and electronic characteristics of the atoms. Furthermore, the atomic binding environment which is related to the molecule of interest is taken into account as well. They also studied 6 molecular descriptors which were reported by Suzuki [20], besides the ETSI-based approaches. In terms of the predictive power of the models, they claimed that support vector machines worked better than multivariate linear regression and artificial neural networks methods. In 2012, Bagheri et al. [22] attempted to predict the auto ignition temperatures for hazardous materials. In order to model flash point (FP) and the auto ignition temperature, they used artificial neural networks and PSO-MLR approaches. This was done for 85 sulfur organic compounds which were from 3 chemical categories. For auto ignition temperature, they calculated R² values of 0.9889 and 0.9259 using artificial neural networks and PSO-MLR, respectively, which indicate reasonable results [23]. Table 1 illustrates the details of the previously published model and their database which are separated into different classes of materials.

The main purpose of this study was to present an alternative technique to esimate the auto ignition temperature of organic compounds. The basis of this method is QSPR and in order to select the variables, we applied GA-ANFIS, GP, DE-ANFIS and PSO-ANFIS. This requires choosing the best and most consistent descriptors subset that has significant contributions to the overall auto ignition temperature. We can achieve this goal by using only nine descriptors that can be obtained from molecular structure.