Prediction of Optimal Salinities for Surfactant Formulations Using a Quantitative Structure鈥揚roperty Relationships Approach

详细信息    查看全文

• 作者：Christophe Muller ; Ana G. Maldonado ; Alexandre Varnek ; Benoit Creton
• 刊名：Energy & Fuels
• 出版年：2015
• 出版时间：July 16, 2015
• 年：2015
• 卷：29
• 期：7
• 页码：4281-4288
• 全文大小：322K
• ISSN：1520-5029

文摘

Each oil reservoir could be characterized by a set of parameters such as temperature, pressure, oil composition, and brine salinity, etc. In the context of the chemical enhanced oil recovery (EOR), the selection of high performance surfactants is a challenging and time-consuming task since this strongly depends on the reservoir鈥檚 conditions. The situation becomes even more complicated if the surfactant formulation is a blend of two or more surfactants. In the present work, we report quantitative structure鈥損roperty relationships (QSPR) correlating surfactants鈥?structures and their composition in a mixture with optimal salinity (S_opt), corresponding to minimal interfacial tension in the reference brine/surfactants/n-dodecane system, at T = 313 K and P = 0.1 MPa. Particular attention was paid to selected families of surfactants: 伪-olefin sulfonate (AOS), internal olefin sulfonate (IOS), alkyl ether sulfate (AES), and alkyl glyceryl ether sulfonate (AGES). The models were built and validated on the database containing S_opt values for 75 surfactants鈥?formulations. Molecular structures of amphiphilic molecules were encoded by functional group count descriptors (FGCD), ISIDA substructural molecular fragment (SMF) descriptors, and CODESSA molecular descriptors (CMD). For mixtures, descriptors were calculated as linear combinations of descriptors of individual compounds weighted by their mass fractions in mixtures. Different machine-learning methods鈥攕upport vector machine (SVM), partial least-squares (PLS) regression, and random subspace (RS)鈥攈ave been used for the modeling. Both global (on the entire database) and local (on individual families) models have been built. Models display reasonable accuracy (about 0.2 log S_opt units) which is comparable with the experimental error of measured S_opt. Our results show that the suggested approach can be successfully used to build predictive models for relatively small data sets of mixtures of chemical compounds.