Journal of Engineering Geology

---

(Paper pages 649-676) Engineering characteristics of alluvium and cemented materials of the slopes around the Milad Tower, and the results of slopes stability analyses under static condition is presented in this paper. Also in the paper, the feasibility of developing and using artificial neural networks (ANNS) for slope stability prediction is investigated. According to the geometry of slopes and strength and deformation properties of alluviums, factor of safety is calculated in 2D and 3D by PLAXIS7.2 and PLAXIS 3D Tunnel codes, respectively, and the results are also compared. In addition, stability of slopes is investigated through the use of MLP artificial neural networks (ANNs), which developed in MATLAB environment. The database used for development of the model comprises a series of 252 factor of safety for different slopes conditions (2D, 3D, flatted and 18 inclined from horizon at top of cut). The optimal ANN architecture (hidden nodes, transfer functions and training) is obtained by a trial-and-error approach in accordance to error indexes and real data. The input data for slope stability estimation consist of values of geotechnical and geometrical input parameters. As an output, the network estimates the factor of safety (FoS). The results indicate that the ANN model is able to accurately predict the FoS of the slopes.

---

One of the most important issues in the Reverse Analysis is analyzing the density resulting from the compaction of in fine soils. The conventional methods in d etermination of soil density are: sand cone, rubber balloon and nuclear density gauge. Trained neural network, as a suitable alternative for conventional methods based on models analyzed by those methods, is not only as accurate but it is also easier to calculate and implement. In the present article, a model based on multilayer perceptron of neural network is presented for prediction of the behavior of fine soils density in Sarabi Dam. The paper presents the implementation process and density of the soil layers. The input variables include 4 geotechnics and 4 implementation parameters. The geotechnic parameters consist of: optimum moisture content, maximum specific gravity, liquid and plasticity limit implementation parameters consist of: the number of cross rollers, thickness of the layers and density and moisture of the soil obtained from the site. The model is based on multilayer neural network, using the error back propagation approach and it is capable of calculating the density. As a result, the maximum specific gravity laboratory, using the aforementioned geotechnic and implementa-tion parameters, is presented. The method compates the maximum specific gravity laboratory accurately at almost 100 percent.

---

Prediction of location of future earthquakes with event probability is useful in reduction of earthquake hazard. Determination of predicted locations has attracted more attention to design, seismic rehabilitation and reliability of structures in these sites. Many theories were proposed in the prediction of time of occurrence of earthquake. There is not a method for prediction time of future earthquakes. Many studies have been done in the prediction of magnitude of earthquakes, but there are not any investigations on prediction of earthquake hazard zonation. In this study, the locations that have probability of the event of future earthquake have been predicted by artificial neural networks in Qum and Semnan. Neural networks used in this study can extract to complicate properties of patterns by receipting the interval patterns. Furthermore, the map of earthquake hazard zonation has been drawn. Properties of occurred earthquake were collected since 1903. The most probable event of earthquake in Qum has been predicted 31.6% in center, and 28.9% in north of Semnan

---

This paper presented the feasibility of developing and using artificial neural networks (ANNs) for modeling the monotonic large scale triaxial tests over angular, rounded rockfill and materials contained various percentages of fines as a construction material in some dams in Iran. The deviator stress/excess pore water pressure versus axial strain behaviors were firstly simulated by employing the ANNs. Reasonable agreements between the simulation results and the tests results were observed, indicating that the ANN is capable of capturing the behavior of gravely materials. The database used for development of the models comprises a series of 52 rows of pattern of strain-controlled triaxial tests for different conditions. A feed forward model using multi-layer perceptron (MLP), for predicting undrained behavior of gravely soils was developed in MATLAB environment and the optimal ANN architecture (hidden nodes, transfer functions and training) is obtained by a trial-and-error approach in accordance to error indexes and real data. The results indicate that the ANNs models are able to accurately predict the behavior of gravely soil in CU monotonic condition. Then, the ability of ANNs to prediction of the maximum internal friction angle, maximum and residual deviator stresses and the excess pore water pressures at the corresponding strain level were investigated. Meanwhile, the artificial neural network generalization capability was also used to check the effects of items not tested, such as density and percentage smaller of 0.2 mm.

---

Landslides are natural hazards that make a lot of economical and life losses every year. Landslide hazard zonation maps can help to reduce these damages. Taleghan watershed is one the susceptible basin to landslide that has been studied. In this paper, landslide hazard zonation of the study area is performed at a scale of 1:50,000. To achieve this aim, layers information such as landslides distribution, slope, aspect, geology (lithology), distance from the faults and distance from rivers using artificial neural network-based Radial Basis Function (RBF) and perceptron neural network (MLP), has been studied. Principal of RBF method is similar to perceptron neural network (MLP), which its ability somewhat has been identified up to now and there are several structural differences between these two neural networks. The final results showed that the maps obtained from both methods are acceptable but the MLP method has a higher accuracy than the RBF method.

---

This paper presents a feed-forward back-propagation neural network model to predict the retained tensile strength and design chart to estimate the strength reduction factors of nonwoven geotextiles due to the installation process. A database of 34 full-scale field tests was utilized to train, validate and test the developed neural network and regression model. The results show that the predicted retained tensile strength using the trained neural network is in good agreement with the results of the test. The predictions obtained from the neural network are much better than the regression model as the maximum percentage of error for training data is less than 0.87% and 18.92%, for neural network and regression model, respectively. Based on the developed neural network, a design chart has been established. As a whole, installation damage reduction factors of the geotextile increases in the aftermath of the compaction process under lower as-received grab tensile strength, higher imposed stress over the geotextiles, larger particle size of the backfill, higher relative density of the backfill and weaker subgrades.