APPLICATION OF ANN IN STRONG GROUND MOTION PREDICTION

Abstract

Strong Ground Motion (SGM) record at a particular site during an occurrence of earthquake is a result of complex – non linear combination of many factors. For design of engineering structures for a specified region the information about the characteristics of strong ground motion is of paramount importance. Peak Ground Acceleration (PGA) is most frequently used parameter to characterize such ground motions. Ground Motion Prediction Equations (GMPE’s) are commonly used for estimating these loading conditions by using strong ground motion data from previous recorded earthquakes. A very little agreement has been reached in the past 30 years of ground motion estimation relation studies and the scatter could not be reduced to requisite level. This is more because the relations not only depends upon data selection, characterization of source, path or site or the regression technique employed but also on the purpose for which equation is intended to be used. So the process of determining the regression relation much depends upon the appropriate judgment of scholar. The duration of strong ground motion is another parameter of paramount importance during an earthquake occurrence. In last few decades numerous definition of duration has been purposed by number of authors. Most popular is the time interval in which significant contribution to the integral of the square of acceleration (∫a2dt) referred to as the accelerograms intensity takes place. Further, local site conditions also play a very critical role to understand damage assessment during an earthquake, since site effect also amplify or de-amplify the ground motion at recording site due to complex phenomena in various layers of soil in top 30 m of recording site. The uncertainties due to both physical as well as computational aspects in GMPE’s to predict SGM, its duration and the local site effects lead to significant residuals/errors and therefore, expose inability to predict the observed values. The inherent uncertainties and limitations of mathematical models make inroads for newer methods to predict SGM. It is in this context an endeavor has been made to search for methodologies for prediction of SGM, duration and local site effects in the present study. The main objective was to develop ANN's (i.e. Artificial Neural Networks) to predict PGA, duration of SGM and local site condition from the observed SGM at a given site using different combination of inputs i.e. magnitude, hypo-central distance, average shear wave velocity, focal mechanism, average primary velocity, average soil density, average Standard Penetration Test (SPT). Artificial Neural Networks (ANNs) are efficient computing models which have shown their strengths in solving many complex problems in numerous fields. They have the versatility iii to approximate a wide range of complex functional relationships between sets of input and output data. The purpose of this study is to predict strong ground motion parameters using ANN that are of primary significance in earthquake engineering. In this study, sets of Multilayer Perceptron (MLP) neural network model are trained to predict the PGA, duration of strong motion, and site characteristics. The ANN model is intended to be developed for Indian region but due to lack of strong motion data, the database used in the study is taken from Kyoshin Net (K-NET) database of Japan. Later, developed ANN model is indigenised its use in Indian context. NeuroIntelligence (Neural Network Simulator) software has been used to model ANN and the standard back-propagation supervised training scheme is used to train all networks. The strong motion data set consists of records from Japanese earthquakes of magnitude more than 4.5 and hypo-central distance less than 200 km for various sites which are classified on the basis of average shear wave velocity calculated using FEMA (356, 2000). In this study, averaged horizontal components of time histories have been used. Basic information such as earthquake magnitude, hypo-central distance, and focal mechanisms, average values of Standard Penetration Test (SPT) blow count, primary wave velocity, shear wave velocity, and density of soil have been used as seven input variables to train the neural network. Since at most locations in world, only average shear wave velocity is used for site characterization, therefore, an attempt has also been made to train the neural network with magnitude, hypo-central distance, focal mechanism and average shear wave velocity as four input variables and magnitude, hypo-central distance and average shear wave velocity as three input variables using Japanese records. For ANN model developed with seven inputs to predict PGA for all site class, the correlation coefficient ‘R’ has been observed to be almost ~0.8 for all the cases showing a good correlation between observed and predicted data. However, it may be noted that values of ‘R’ are becomes minimum i.e., less than 0.8 in case of 4 inputs showing that either the reported focal mechanisms do not match with the actual physical process or this parameter is least correlated showing its relatively lesser influence on the predicted value. The testing scatter plot of predicted PGA to target PGA for all site classes with 7 inputs reveal that the prediction is good for PGA up to 40 cm/s2. A similar trend is observed with 4 and 3 inputs cases for all the three classes. Based on the observation a conclusion of good prediction in lower range of PGA values can easily be drawn. iv In case of ANN model developed with seven inputs to predict duration for all site class, the values of ‘R’ is almost more than 0.8 for all the cases showing a good correlation between observed and predicted data except for two cases with 4 inputs and 3 inputs for Class A. The scatter diagram reveals that the duration between 10 to 50 seconds can be predicted faithfully from this approach. The duration definition and its formulae used by various workers generally show much more scatter than the one seen in the prediction diagrams using ANN. ANN model has been developed with eighteen inputs to predict average shear wave velocity for class site A which shows a good agreement with value of ‘R’ about ~0.9 with tested dataset. The inputs include magnitude, hypo-central distance, PGA and normalized response spectral ordinates ranging from 0.01 sec to 4 sec. This trained network with eighteen inputs has been tested for few significant earthquakes where average shear wave velocity recorded at stations were taken from NGA Flat File version 7.3. ANN model made a successful attempt to predict V􀴥S30 for those sites having V􀴥s30 between 750 - 1100 m/sec. Finally trained model was use to predict V􀴥S30 for those Indian Himalayan sites where at least 4 past earthquake records are available and classified as Class A site. The prediction of V􀴥S30 is first order estimation and need to be validated in future by borehole data at these sites. An alternate methodology for prediction of Strong Ground Motion has been developed in the present study. The method can be used to predict strong ground motion in terms of PGA and duration for its use in earthquake engineering. Further, the site characteristics can also be estimated using the recorded ground motion which is otherwise not worked out for the sites. The present thesis has concluded the successful use of ANN in earthquake engineering.