Identification of liquefiable silty soils using cone penetration test

Earthquakes can induce adverse soil behaviors, such as soil liquefaction, cyclic softening, and ground amplification, which can result in significant deformations or settlements in structures located on the affected soils, especially in alluvial environments. Liquefaction issues are commonly observed in alluvial environments with shallow water levels, such as Adapazarı. Saturated sands, silty sands, and silts are the most susceptible soil types to liquefaction. There are several different methods available for determining liquefiable and non-liquefiable soils. Methods used to assess the liquefaction potential of coarse-grained soils are well-established and provide consistent results. However, obtaining a general approach for determining the liquefaction potential of fine-grained soils is challenging. Liquefaction criteria based on the physical properties of soils present different soil characteristics and their respective threshold values. Therefore, this study proposes new recommendations based on machine learning methods using the existing criteria and the soil data available in the database. Within the scope of the study, a model for the liquefaction potential of soils is presented using seven well-known machine learning approaches. The Adapazarı dataset is evaluated based on expert opinions and subjected to preprocessing before training. The results obtained by algorithms with different parameters during the training process are recorded, allowing algorithms with the best parameters to yield better results. When the algorithms are ranked based on the quality of the training set, decision trees (DT) exhibit the highest accuracy rate of 90%. Random Forest (RF) classifier provides the second-best results. The mentioned DT model yields a 91% accuracy on the randomly selected training set and an 84% accuracy on the test set. Rules previously established based on expert opinions to determine liquefaction are improved using machine learning methods. The DT model in this study offers a good alternative for liquefaction prediction and provides more accurate results compared to previous studies. Additionally, the importance of features in the considered dataset is determined using decision trees. According to the feature ranking, the order of significance is discovered to be D50, depth, fines content (FC), clay percentage, wn, wL. In today's soil investigations, liquefaction analysis is also performed using Cone Penetration Test (CPT) results, which play an important role in obtaining reliable data. The primary use of the CPT test is to determine the soil stratigraphy and pre-evaluate geotechnical properties. The test is applied in soft to medium-stiff soils where the probe can be pushed down without the use of impact or rotation, up to the gravel fraction. In the classification of soils using CPT, the soil behavior index (Ic) stands out. CPT results and various Ic formulas proposed by different researchers have been compared in many aspects with the physical properties of fine-grained soils available in the database, leading to findings that could be decisive in distinguishing liquefiable and non-liquefiable soils. The performance of existing Ic formulas in identifying liquefiable and non-liquefiable soils has been examined, and optimization of the Ic formulas has been carried out for improvement. A comprehensive database obtained from field studies conducted since the 1999 Adapazarı earthquake, focusing on the determination of the boundaries of clay and silt soils using Ic, has been utilized. The CPT data was obtained from experiments conducted with a 200 kN capacity acoustic CPTU machine owned by Sakarya University. Physical property test results from disturbed and undisturbed samples obtained from concurrent drilling with CPT experiments are also available. Thus, the physical property data of soils used in traditional classification methods were combined with CPT data, providing an opportunity to compare and analyze these data. A total of 776 data points in the database were used. Each data point includes information such as liquid limit, plastic limit, clay content, water content, fines content, and average grain diameter. It can be stated that the liquid limit values of soils range from 30 to 50, and with increasing liquid limit values, there is a decrease in shear strength and an increase in friction ratios. Similarly, the fines content of soils varies between 60% and 80%, and as the fines content increases, the shear strength decreases and the friction ratio increases. Different researchers have proposed soil behavior index (Ic) values to determine the classes of fine-grained soils, but a precise differentiation for clay and silt soils cannot be achieved with a specific Ic value. Each chosen Ic value can correspond to both clay and silt soils. The aim is to determine a specific Ic value that provides the best distinction between clay and silt, with the highest percentage of clay soils below this value and the lowest percentage above it. When examining the distributions of clay and silt using different Ic formulas proposed by various researchers, the Ic threshold value that can most accurately separate clay and silt soils was determined by considering the minimum and maximum values obtained by clay and silt for each Ic formula. When examining the results based on the percentage of data points falling below the determined threshold value for the separation of clay and silt, it can be said that with an Ic value of 2.55 achieves the best distinction with a 74.5% difference ratio. Soil behavior index (Ic) values have been analyzed separately for liquefiable and non-liquefiable soils in the database, considering the Ic values proposed by different researchers. At this point, it is necessary to determine the liquefaction potential of the soils. For the determination of liquefaction potential, field observations from previous studies were examined. For areas where field observations were not available, liquefaction criteria based on physical properties, as determined in the literature, were taken into consideration. In the study, soils that satisfy at least 3 or more out of 11 liquefaction criteria were categorized as potentially liquefiable soils, while soils that did not satisfy these criteria were categorized as potentially non-liquefiable soils. Accordingly, 545 soils were identified as having low liquefaction potential, and 233 soils were identified as having high liquefaction potential. The soil behavior index Ic for soils with low and high liquefaction potential was calculated using formulas proposed by different researchers, and an analysis was conducted to determine which formulas could distinguish liquefiable and non-liquefiable soils more effectively. Additionally, various coefficients are used in each formula. The possibility of these coefficients, within certain limits of variation, better determining the liquefiable and non-liquefiable regions was examined. The evaluation results indicate tha one of the Ic formula is the most effective in distinguishing liquefiable and non-liquefiable soils, both with its original coefficients and optimized coefficients. The optimization process resulted in improvements of up to 21.6% compared to the original formulas. It is observed that no single Ic formula can clearly distinguish soils with low and high liquefaction potential, and there are some soils that fall into the common area regardless of the formula used. The evaluation of the liquefaction of a fine-grained sample with known physical properties, numbered 776, has been conducted by different researchers. Based on these evaluations, data regarding the liquefied, non-liquefied, and test zone samples have been determined for each method. Subsequently, the average cone penetration resistance, friction resistance, and pore water pressures have been determined for the liquefied and non-liquefied samples, and these raw data have been normalized to determine the normalized cone resistance and friction ratios for soil classification purposes. The normal distributions of the Ic values for the liquefied, non-liquefied, and test zone samples have been examined. Soil behavior indexes, normalized cone resistance, and friction ratios have been taken into account for soils that do not liquefy, can liquefy, and fall into the test zone. The distinction of non-liquefying soils based on CPT data can be made definitively according to all criteria. As observed from both the normal distribution and the classification chart distributions of non-liquefied and test zone soils largely overlap. This is due to the fact that the soils falling into the test zone exhibit a wide range of both liquid limit and plasticity index. Accordingly, considering the CPT data, for a soil to be considered non-liquefiable, the soil behavior index (Ic) should be greater than 2.98. In addition, the normalized cone resistance should be less than 10 and the normalized friction ratio should be greater than 3.08.