Quantitative Structure−Property Relationship (QSPR) for the Adsorption of Organic Compounds onto Activated Carbon Cloth: Comparison between Multiple Linear Regression and Neural Network

• Published in: Environmental science & technology Vol. 33; no. 23; pp. 4226 - 4231
• Main Authors: Brasquet, C, Bourges, B, Le Cloirec, P
• Format: Journal Article
• Language: English
• Published: Washington, DC American Chemical Society 01.12.1999
• Subjects: ACTIVATED CARBON; ADSORBENTS; ADSORPTION; Applied sciences; Carbon; Chemistry; ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATION; Exact sciences and technology; General and physical chemistry; General purification processes; MOLECULAR STRUCTURE; Neural networks; Organic chemistry; ORGANIC COMPOUNDS; Pollution; Regression analysis; Solid-liquid interface; STRUCTURE-ACTIVITY RELATIONSHIPS; Surface physical chemistry; WASTE PROCESSING; Wastewaters; Water treatment and pollution; Statistical analysis; Molecular structure; Filtering cloth; Liquid solid adsorption; Linear regression; Organic adsorbate; Adsorption isotherm; Adsorption capacity; Inorganic adsorbent; Molecular connectivity; Neural network; Multiple regression; Activated carbon; Physicochemical purification; Property structure relationship; Waste water purification; Organic compounds
• ISSN: 0013-936X; 1520-5851
• DOI: 10.1021/es981358m

The adsorption of 55 organic compounds is carried out onto a recently discovered adsorbent, activated carbon cloth. Isotherms are modeled using the Freundlich classical model, and the large database generated allows qualitative assumptions about the adsorption mechanism. However, to confirm these assumptions, a quantitative structure−property relationship methodology is used to assess the correlation between an adsorbability parameter (expressed using the Freundlich parameter K) and topological indices related to the compounds molecular structure (molecular connectivity indices, MCI). This correlation is set up by mean of two different statistical tools, multiple linear regression (MLR) and neural network (NN). A principal component analysis is carried out to generate new and uncorrelated variables. It enables the relations between the MCI to be analyzed, but the multiple linear regression assessed using the principal components (PCs) has a poor statistical quality and introduces high order PCs, too inaccurate for an explanation of the adsorption mechanism. The correlations are thus set up using the original variables (MCI), and both statistical tools, multiple linear regression and neural network, are compared from a descriptive and predictive point of view. To compare the predictive ability of both methods, a test database of 10 organic compounds is used. Results show the good descriptive ability of NN compared with that of MLR, with more than 68% variance explained by NN, whereas MLR allows only 44% variance explanation. However, the predictive ability of NN seems to be low, especially when the structure of the test compounds is not well described in the training database. The good descriptive ability of NN is then exploited to carry out a variable analysis using the Garson weight partitioning method and to give information about the adsorption process. This study shows that flat molecules seem to be better adsorbed onto activated carbon fibers than bulky molecules, because of an adsorption which is located between the micrographitic planes of fibers. The adsorption process occurs via an electron donor−acceptor interaction between the surface of the activated carbon fiber (donor) and the solute (acceptor). Consequently, the aromatic compounds with electron-withdrawing substituents seem to be favored. Furthermore, the lower the solute affinity for the aqueous media, the greater seems to be the adsorption.