10:45 AM - 11:00 AM
[SSS11-12] Ground motion prediction models using artificial neural network
Recently, machine learning (ML) methods and artificial intelligence (AI) became very popular. Especially, AI can be useful for ground motion prediction models (GMPMs). In contrast with the classical approach for GMPM, with some methods of AI such physical aspects can be considered which is impossible to take into account. As a result, we can obtain models with almost all parameters that have an influence on strong ground motion and at the same time we can reduce both epistemic and aleatory uncertainty. In the study GMPMs using AI for different ground motion parameters such as PGA, PGV, and Sa are tested, tried them for different number of network parameters and obtained the best models. Artificial neural network (ANN) was selected for study because it can provide more accurate prediction of ground motion intensity measures for all distances and magnitudes. Also, in contrast with classical regression analysis ANN has a capability of adaptively learning from experience and extracting various discriminators in pattern recognition. For GMPM development fault type and local soil effect were also considered. In this method local soil conditions were considered based on Vs30. ANN was used for two different datasets obtained from Georgia (Institute of Earth Sciences and National seismic Monitoring Center, Ilia State University) and Japan (NIED). As mentioned above, predictor variables in our case were magnitude, distance, faulting mechanism and Vs30.
A hybrid technique combining genetic algorithm and Levenberg–Marquardt technique was used for training the model. After training and testing selection of the best model was based on several criteria: Tried several activation functions (between input and hidden layers, and between hidden and output layers). Fast convergence (lower number of iterations), lower values of residuals.
We have used Tansig for hidden layer. By using it for the hidden layer, the model benefits from non-linear representations that enhance learning capabilities and convergence rates. For output Purelin function was used. The choice of this function for the output layer enables the network to output continuous values effectively, making it well-suited for tasks that require regression, such as predicting ground motion parameters. The Tansig-Purelin combination was likely the best fit for our data. It balances fast convergence with low error rates, which suggests it effectively captures both the nonlinearity in the ground motion data and the need for stable output values in the final layer. For the best model fast convergence (lower number of iterations) and lower values of residuals were prioritized.
Finally, comparison of the obtained model with the existing classical models was done to find out the benefit of the ANN. For comparison in case of Georgia, it was compared with Jorjiashvili et al. (2022). In case of Japan, it was compared with Si and Midorikawa (1999) that use for the National Seismic Hazard Maps for Japan. Quite good agreement with a slight difference of obtained models with existing classical models was observed for Georgia as well as for Japan which indicates the sophistication of the ANN method compared to the classical method.
A hybrid technique combining genetic algorithm and Levenberg–Marquardt technique was used for training the model. After training and testing selection of the best model was based on several criteria: Tried several activation functions (between input and hidden layers, and between hidden and output layers). Fast convergence (lower number of iterations), lower values of residuals.
We have used Tansig for hidden layer. By using it for the hidden layer, the model benefits from non-linear representations that enhance learning capabilities and convergence rates. For output Purelin function was used. The choice of this function for the output layer enables the network to output continuous values effectively, making it well-suited for tasks that require regression, such as predicting ground motion parameters. The Tansig-Purelin combination was likely the best fit for our data. It balances fast convergence with low error rates, which suggests it effectively captures both the nonlinearity in the ground motion data and the need for stable output values in the final layer. For the best model fast convergence (lower number of iterations) and lower values of residuals were prioritized.
Finally, comparison of the obtained model with the existing classical models was done to find out the benefit of the ANN. For comparison in case of Georgia, it was compared with Jorjiashvili et al. (2022). In case of Japan, it was compared with Si and Midorikawa (1999) that use for the National Seismic Hazard Maps for Japan. Quite good agreement with a slight difference of obtained models with existing classical models was observed for Georgia as well as for Japan which indicates the sophistication of the ANN method compared to the classical method.