UNIAXIAL COMPRESSIVE STRENGTH PREDICTION OF JET GROUTING COLUMNS USING SUPPORT VECTOR MACHINES Joaquim Tinoco and António Gomes Correia Departamento de Engenharia Civil/C-TAC Universidade do Minho Campus de Azurém, Guimarães, Portugal E-mail: {jabtinoco|agc}@civil.uminho.pt Paulo Cortez Centro Algoritmi, Departamento de Sistemas de Informação Universidade do Minho Campus de Azurém, Guimarães, Portugal E-mail: pcortez@dsi.uminho.pt KEYWORDS Soft soils, soil-cement mixtures, soil improvement, jet grouting, uniaxial compressive strength, regression, data mining, support vector machines, sensitivity analysis ABSTRACT Uniaxial compressive strength (UCS) is the mechanical properties currently used in geotechnical works design, namely in jet grouting (JG) treatments. However, when working with this soil improvement technology, due to its inherent geological complexity and high number of variables involved, such design is a hard, perhaps very hard task. To help in such task, a support vector machine (SVM), which is a data mining algorithm particularly adequate to explore high number of complex data, was trained to estimate UCS of JG samples extracted from real JG columns. In the present paper, the performance reached by SVM algorithm in UCS estimation is shown and discussed. Furthermore, the relation between mixture porosity and volumetric content of cement and the JG system were identified as key parameters by performing a 1-D sensitivity analysis. In addition, the effect and the interaction between the key variables in UCS estimation was measured and analyzed. INTRODUCTION Jet grouting (JG) technology is one of the most used soft-soil improvements methods (Falcão et al. 2000). According to JG technology, a high speed and pressure of grout (with or without other fluids) is injected into the subsoil, which cut and mixes the soil. At the end an improved mass of soil, often termed as Soilcrete is obtained. According to the number of fluids injected, three systems are conventionally in use: single, double and triple fluid system. Due to the heterogeneity of the soils, the constructive process of JG technology and nature of treatment fluid injected (normally water cement grout) there are many variables involved in treatment process (Nikbakhtan et al. 2010). Such conditions make the design of JG technology a complex geotechnical task. Nowadays, such design is almost performed based on empirical methods (Lee et al. 2005; Narendra et al. 1996), mainly in the initial project stages and in small scale geotechnical works where information is scarce. Therefore, and since these empirical methods are often too conservative and have a very limited applicability, the quality and the economy of the treatment can be compromised. Hence, and bearing in mind the high versatility of JG technology and its role in important geotechnical works, it is very important to develop rational models to estimate the effects of the different variables involved in JG process. On the other hand, in the last few years some powerful tool, incorporating advanced statistic analysis, has been developed and are able to automatically extract important rules from vast and complex data. Such tools, usually known as data mining (DM) techniques, has been successfully applied in several scientific areas namely in Civil Engineering domain (Lai and Serra 1994; Rezania and Javadi 2007). One of the most interesting DM algorithms is the Support Vector Machines (SVM), which was used in the present work and has the particularity to be applied in both classification and regression problems. SVM is especially usefully to explore data with nonlinear relationships between several inputs and the target variable and had been successfully applied to solve geotechnical problems (Goh and Goh 2007; Tinoco et al. 2011b). The main criticism of "black box" DM techniques, such as SVM or artificial neural networks is the lack of explanatory power, i.e. the data-driven models are difficult to interpret by humans (Goh and Goh 2007). However, to overcome such drawback a sensitivity analysis (SA) procedure can be applied (Cortez and Embrechts 2011). The performance reached by SVM algorithm trained with data collected directly from JG columns (JGS), with different JG parameters and soilcrete characteristics are shown and discussed in the present paper. Moreover, the key variables in UCS estimation are identified by applying a 1-D SA. Furthermore, the influence of the key variables in UCS estimation are quantified and discussed. In addition, and keep in mind a more realistic interpretation of the results a 2- D SA was performed to the first two key variables. SUPPORT VECTOR MACHINES Support Vector Machines are very specific class of algorithms, which is characterized by use of kernels, absence of local minima, sparseness of the solution and capacity control obtained by acting on the margin, or on number of support vectors. When compared with other types of base learners, such as the famous multilayer perceptron, SVM represents a significant enhancement in functionality. The supremacy of SVM lies in their use of non-linear kernel functions that implicitly map inputs into high dimensional feature spaces. In this feature spaces linear operations may be possible that tray to find the best linear separating hyperplane (  = ω 0 + ∑ ω i ∅ i (x) m i=1 ), related to a set of support vector points, in the feature space. Thus, although