An accurate model for prediction of autoignition temperature of pure compounds

https://doi.org/10.1016/j.jhazmat.2011.02.014Get rights and content

Abstract

Accurate prediction of pure compounds autoignition temperature (AIT) is of great importance. In this study, the Artificial Neural Network-Group Contribution (ANN-GC) method is applied to evaluate the AIT of pure compounds. 1025 pure compounds from various chemical families are investigated to propose a comprehensive and predictive model. The obtained results show the squared correlation coefficient of 0.984, root mean square error of 15.44 K, and average percent error of 1.6% for the experimental values.

Introduction

If the temperature of a flammable gas–air mixture is uniformly raised, it eventually reaches to a value, at which combustion occurs. For the range of flammable mixtures, there is a mixture composition which has the lowest ignition temperature. This minimum temperature is called autoignition temperature (AIT) or spontaneous ignition temperature (SIT). The AIT is defined as the lowest temperature, at which a substance will produce hot-flame ignition in air at atmospheric pressure without the aid of an external energy source such as a spark or flame [1].

At the AIT, the rate of heat evolved by exothermic oxidation reaction overbalances the rate at which heat is lost to the surroundings and causes ignition. The AIT is dependent not only on the chemical and physical properties of the substance but also on the method and instrument employed for its determination such as the volume and the material of the used vessel, test pressure, and oxygen concentration [1].

Due to importance of the AIT of fuels, time-consuming, and laborious processes of experimental determination of the AIT, theoretical computation of the AIT is of great interest. There have been several models so far presented for this purpose. For instance, Suzuki [2] developed some equations to estimate the AIT using a number of several molecular-based parameters and physical properties. The model showed average absolute deviation of 4.5% and squared correlation coefficient of 0.9. The correlation was developed using a dataset for 250 chemical compounds. Furthermore, the correlation was unable to estimate the AIT of 23 compounds with enough accuracy. In the case that these compounds are added to the calculations, the squared correlation coefficient decreases to 0.85 and the average absolute deviation increased to 5.4%. Another attempt has been done by Tetteh et al. [3], who tried to modify the Suzuki's correlation by presenting an Artificial Neural Network instead of the correlation. Therefore, they used the parameters previously applied by Suzuki as the inputs of their models. They reported the error of 30 °C using the same correlation used by Suzuki. Mitchel and Jurs [4] developed quantitative structure–property relationship (QSPR) for estimation of the AIT. They applied a dataset including 327 heterogeneous organic compounds to develop their model. Moreover, they stated that their attempt to model all the compounds together was unsuccessful. Hence, these researchers presented several models each of them appropriate for calculation of the AIT of a portion of the compounds in the dataset. Another QSPR approach was developed by Kim et al. [5] using a dataset including 200 compounds. Their model showed that the square of the correlation coefficient (R2) for the AIT of the 157-member training set was 0.920, and the root mean square error was 25.876. The squared correlation coefficient and root mean square error of their model for a 43-member prediction set were 0.910 and 28.968, respectively. Albahri and George [6] and Albahri [7] presented two very similar models for estimation of AIT. It seems the first one was a bit better. The model showed squared correlation coefficient of 0.98 and average error of 2.8% over a data set including 490 compounds.

Recently, some other models have been presented for estimation of AIT property. It seems that these models are not more general and accurate than the model presented by Albahri and George [6]. For instance, we can refer to the two models proposed by Pan et al. [8], [9], [10]. For the first one, they used a small data set including 192 compounds to develop a QSPR-based support vector machine type method. Although the square correlation coefficient of the model obtained was 0.984, the considerable difference between the leave-one-out parameter and squared correlation parameter showed its poor predictive power (QLOO). Later, they presented an Artificial Neural Network-Group Contribution model using a dataset including 116 hydrocarbons. The average absolute error and root mean square error of the model were 21.6 °C and 31.1, respectively. Finally, using a data set containing 446 organic compounds, Pan et al. [10] developed a QSPR-based support vector machine type mode to estimate the AIT. The squared correlation coefficient and root mean square of the model were 0.9 and 36.86 °C. A recent model presented by Chen et al. [11]. They used the classic group contribution method to calculate the AIT. The squared correlation coefficient of the model was 0.85.

One of the main problems that have been mentioned in majority of the presented models is that the AIT values are hard to be correlated with chemical structure of compounds. This issue can be easily inferred from the low accuracy of the previously presented models. Another issue is related to the number of compounds that have been used in developing the previous models. Even though, there are experimental data for much more than 500 compounds available in the literature, but none of the aforementioned researchers applied such wide ranges of data for developing their models.

The main aim of this study is to present a new comprehensive model for calculation/estimation of AIT using a large dataset containing more than 1000 compounds.

Section snippets

Data set preparation

DIPPR 801 [12] database has been found especial applications in developing new methods for prediction of physical properties because it contains a large number of pure compounds as well as their physical properties. In order to provide a data set for calculation of the AIT, 1025 pure compounds from 78 different chemical groups were investigated and their related AITs were considered for the study. These compounds are presented as supplementary material.

Development of a new group contribution

Having defined the dataset, the chemical

Results and discussion

An optimized Feed Forward Artificial Neural Network has been obtained using the aforementioned procedure for prediction of the AIT of 1025 pure compounds. For this purpose, several 3FFANNs modules have been generated assuming numbers 1 through 50 for n (number of neurons in hidden layer) using the previously described procedure. The most accurate results, without overfitting are observed for n = 10. It should be noted that this value is not the global value, because the optimization method used

Conclusion

A group contribution-based model was presented for prediction of the autoignition temperature (AIT) of pure compounds. The model is the result of a combination of group contributions and Feed Forward Artificial Neural Networks. The parameters of the model are the numbers of occurrences of 146 functional groups in each investigated molecule. It should be noted that majority of these 146 functional groups are not simultaneously available in a particular molecule, so computation of these

Acknowledgement

The author gratefully thanks Mr. Ali Eslamimanesh from CEP/TEP, Mines ParisTech, for fruitful comments on the manuscript.

References (43)

  • A.H. Mohammadi et al.

    Determining phase diagrams of tetrahydrofuran + methane, carbon dioxide or nitrogen clathrate hydrates using an artificial neural network algorithm

    Chem. Eng. Sci.

    (2010)
  • A.H. Mohammadi et al.

    Use of an artificial neural network algorithm to predict hydrate dissociation conditions for hydrogen + water and hydrogen + tetra-n-butyl ammonium bromide + water systems

    Chem. Eng. Sci.

    (2010)
  • F. Gharagheizi

    A new group contribution-based method for estimation of lower flammability limit of pure compounds

    J. Hazard Mater.

    (2009)
  • S.A. Kalogirou

    Artificial neural networks in renewable energy systems applications: a review

    Renew. Sust. Energy Rev.

    (2001)
  • ASTM International

    ASTM Standard Test Method E659-78

    (2000)
  • T. Suzuki

    Quantitative structure–property relationships for auto-ignition temperatures of organic compounds

    Fire Mater.

    (1994)
  • B.E. Mitchel et al.

    Prediction of autoignition temperatures of organic compounds from molecular structure

    J. Chem. Inf. Comput. Sci.

    (1997)
  • Y.K. Kim et al.

    Prediction of autoignition temperatures (AITs) for hydrocarbons and compounds containing heteroatoms by the quantitative structure–property relationship

    J. Chem. Soc. Perk Trans.

    (2002)
  • T.A. Albahri et al.

    Artificial neural network investigation of the structural group contribution method for predicting pure components auto ignition temperature

    Ind. Eng. Chem. Res.

    (2003)
  • Project 801, Evaluated Process Design Data, Public Release Documentation, Design Institute for Physical Properties (DIPPR)

    (2006)
  • F. Gharagheizi et al.

    Prediction of Henry's law constant of organic compounds in water from a new group-contribution-based model

    Ind. Eng. Chem. Res.

    (2010)
  • Cited by (46)

    • Numerical study on self-ignition temperature of biomass gasified gas for the application of MILD combustion

      2022, Fuel Processing Technology
      Citation Excerpt :

      However, these Tsi data reported by different publications differ greatly due to different operation and experimental errors. For example, the Tsi of CH4 vary from 810 to 903 K [26]; the Tsi of H2 vary from 500 to 551 K [27,28]; the Tsi of CO vary from 609 to 869 K [27,29]. BGG is a mixture of combustible components of CH4, CO and H2 and non-combustible components of CO2 and N2.

    • Experimental analysis and data-driven machine learning modelling of the minimum ignition temperature (MIT) of aluminium dust

      2022, Fuel
      Citation Excerpt :

      The developed model was tested on 727 compounds with a 1.36 per cent AARE. To forecast the auto-ignition temperature (AIT) of 1025 pure chemicals, Farhad Gharagheizi [19] trained a machine learning model based on the Artificial Neural Network-Group Contribution (ANN-GC) technique. Compounds from several chemical families were used to build a robust and accurate predictive model.

    • An improved method for predicting autoignition temperatures based on first principles

      2022, Fuel
      Citation Excerpt :

      Data must be evaluated not only for methodological consistency, but also must be vetted to ensure it is experimental. Of the 21 publications referenced in Table 1, at least 11 sourced their regression data from AIChE’s DIPPR 801 Database using database snapshots from various years without considering whether the data were experimental or predicted [9,11,15,18,20,21,23,24,26–28]. While the DIPPR 801 Database recommends reliable AIT values, many of the recommended values in the database are predicted and therefore are unsuitable for training prediction methods.

    View all citing articles on Scopus
    View full text