Deep learning-based stochastic ground motion modeling using generative adversarial and convolutional neural networks

This paper proposes a probabilistic framework for generating three-dimensional (3D) synthetic ground motions using deep learning techniques—specifically, generative adversarial networks (GAN) and convolutional neural networks (CNN). Deep learning methods have been shown to surpass classical model classes in performance when provided with large datasets, and the ever-increasing number of ground motion records provides an opportunity to design generative models to produce artificial ground motions that outperform classical models. In addition, these methods can directly extract features and patterns from ground motion data without loss of generality, enabling prediction and generation of synthetic ground motions. The proposed framework consists of two distinct deep learning modules. The first generates normalized 3D synthetic ground motions given source and site characteristics. For this purpose, a conditional Wasserstein GAN comprising a generator and a critic in an adversarial setup is designed in which they engage in a simultaneous competitive process. Through learning from the dataset of real ground motions, the generator attempts to generate artificial ground motions that are more convincing to the critic, whereas the critic seeks to improve its ability to identify the realness or artificialness of the motions and provide the generator with feedback. The second module produces peak ground accelerations (PGA) for the three spatial components of the generated normalized ground motion, given the normalized motion and the said characteristics. For this purpose, a CNN is designed with “inception” layers, each of which concurrently applies multiple convolution filters of varying sizes to the input and concatenates their outputs, enabling the network to efficiently capture features at various scales. The learning performance of both modules is improved by realistic data augmentation techniques that increase training data size and are specifically designed for 3D ground motion records, including random rotations and cropping. The proposed framework is trained and validated using the dataset of over 200,000 records of the KiK-net database. The site and source characteristics utilized in the application of the study comprise the moment magnitude, distance, fault mechanism, and shear wave velocity. The signal generation module is validated through a novel procedure based on the diversity of the generated signals and its comparison with that of the real ground motions, which here demonstrates the absence of overfit and mode collapse. The amplitude prediction module is validated using classical metrics, such as the correlation coefficient between real and predicted PGAs, which, at 0.97 for the test data, demonstrates a satisfactory prediction quality and absence of overfit. Finally, the framework as a whole is validated in time and frequency domains both qualitatively by comparing time-moving averages, pseudo-spectral ordinates, and Fourier amplitude spectra and quantitatively by comparing the distribution of intensity measures of the generated synthetic ground motions with that of the real ground motions using Jensen-Shannon (JS) divergence. The results of JS divergence generally lie below 0.3 with an average of 0.18, which demonstrate a strong similarity between the generated and real distributions.