Ground Motion Model For Data Sparse Regions: Machine Learning Approach

Ground motion models (GMMs) to the recorded ground motion time histories are essential input to the hazard analysis. With recent vast array of strong motion instruments to seismically active regions such as Japan, California, and Mexico, large amounts resulted in abundant recorded data huge metadata. Several global and regional GMMs are developed with these strong motion datasets. However, many active regions (e.g., The Himalayas) are in dearth of recorded strong motion data and metadata to develop predictive models. Despite recent instrumentations by different networks to the Himalayan region, the problem of near-field strong-motion records resulting from sparse instrumentation is the key concern. Traditionally, stochastic models are used in developing GMMs, as developing empirical models with limited data is challenging. Additionally, GMMs developed to other data-rich regions with similar tectonics are used in the hazard estimations. Thus, developing predictive models to these data-poor regions is a key concern which needs to be addressed. In the current work, we address this problem from the data-driven approach such as neural network. Neural networks learn the functional form from the data during training making it suitable for our present problem. Magnitude, epicentre distance, hypocentre depth, and shear wave velocity flag are used as inputs to estimate both the horizontal and vertical response spectra. In this regard, we attempt several approaches in developing the GMM using shallow neural network. Initially we develop model with seven neurons in the hidden layer using the available regional Western Himalayan crustal data and as one expects the model scaled poorly at the near-field. The obtained mean squared error (MSE) mean absolute error (MAE), and coefficient of determination (R2) are 0.6858, 0.6504, and 0.7592, respectively. To address this lack of near-field data, we supplement our regional data with records from global near-field strong motion and in developing GMM. This model has seven neurons in the hidden layer and performed better than the previous model but still had scaling issues at the large magnitude near-field. Further, supplementing data from other regions would influence the predictions. The obtained MSE, MAE, and R2 of the combined database are 0.5690, 0.5830, and 0.8659, respectively. However, the MSE, MAE, and R2 of the Western Himalaya data are 0.8006, 0.7057, and 0.7216, respectively. Finally, we use transfer learning technique: we develop GMM to the global crustal data and global near-field data and use it as a base model to develop GMM with six neurons in the hidden layer using the Western Himalayan data. The obtained MSE, MAE, and R2 of the Western Himalayan database are 0.8688, 0.7282, and 0.6970, respectively. Despite large error compared to previous two models, this model could capture large magnitude near-field effects and distance scaling effects and performed better than the previous two models. We conclude that transfer learning could be used to regions with limited strong motion data in developing GMM.