Using adaptive neuro-fuzzy inference system for modeling unsaturated soils shear strength

Abstract

Determining shear strength of unsaturated soils is generally more complex, time-consuming, and costly than that of saturated soils. In spite of numerous studies conducted on unsaturated test methods, there is still a critical need for the existing studies to be applied while extending the researches on the subject matter of shear strength determination in unsaturated soils. This is primarily in order to save time and money spent on laboratory tests. For this purpose, several empirical approaches to unsaturated soils shear strength testing have been proposed by different researchers. Being based on shear strength and soil matric suction parameters, these empirical methods can be divided into linear and nonlinear variants. Nonlinear empirical methods are usually based on saturated soil shear strength parameters and soil–water characteristic curve (SWCC). However, it is still a time-consuming and costly practice to determine SWCC parameters in the laboratory. As such, several correlations are proposed to estimate SWCC parameters; some of these correlations are, however, based on fitting parameters, and this is again a relatively difficult task to undertake. Recently, artificial intelligence has led to applications in solving nonlinear problems and cases where an accurate understanding of the problem is required. In this research, aiming at proposing adaptive neuro-fuzzy inference system (ANFIS) models to predict unsaturated soils shear strength, two clustering approaches are followed, namely fuzzy c-mean clustering and subtractive clustering. ANFIS input parameters were taken to be the same parameters commonly chosen in most empirical models utilized for determining shear strength; these include net normal stress ($\sigma _{\mathrm{n}}-u_{\mathrm{a}})$, matric suction ($\sigma _{\mathrm{a}}-u_{\mathrm{w}})$, effective cohesion ($c^\prime$), and angle of frictional resistance ($\varphi '^\prime$), and also other especial parameters which were taken in empirical models such as Lamborn (A micromechanical approach to modeling partly saturated soils, Texas A&M University, College Station, 1986), Vanapalli et al. (Can Geotech J 33:379–392, 1996), Öberg and Sällfors (ASTM Geotech Test J 20:40–48, 1997), Bao et al. (Keynote lecture, proceedings of the 2nd international conference on unsaturated soils (UNSAT 98), Beijing, China, pp 71–98, 1998) and Khalili and Khabbaz (Geotechnique 48:681–687, 1998) models collected from SoilVision database. Continuing with the research, a couple of comparisons were made on five empirical models along with the two ANFIS models. The results indicated that both ANFIS models were of high capability in terms of shear strength prediction using the chosen inputs so that more satisfying and adequate predictions be provided compared to empirical models results.