Ground-motion prediction by ANN using machine learning for the Tohoku region, Japan

The purpose of this study is developing new ground motion prediction method that is applicable to mega earthquakes with small prediction errors, comparted with ordinary ground motion prediction equations (GMPEs). We construct the strong-motion database, 38871 records from 342 stations and 652 earthquakes, from NIED K-NET data in Tohoku region, Japan. Using this database, artificial neural network (ANN) by using deep learning method are evaluated. The input variables are: Mw, hypocentral coordinate, station coordinate, station Vs30, and the outputs are PGA, PGV and acceleration response spectra. Several techniques of deep learning method are applied in order to find the appropriate evaluation procedure, and “normalization”, “dropout”, “pre-training”, “ReLU (Rectified Linear Unit)” are found to be efficient to construct reliable ANN avoiding over-training. Finally, the ANN having two hidden layers of 500 units is obtained. We compared the ANN's estimates with those by ordinary GMPEs, whose parameters are Mw, source-to- site distance, Vs30. The prediction standard errors of the evaluated ANN are about 0.16 (in log 10 scale), approximately independent of periods, which are clearly smaller than those of the ordinary GMPE: 0.25-0.3. Also, the ANN can evaluate site-specific spectral peaks, large spectral amplitudes due to deep intra-slab earthquakes, which are difficult to estimate by ordinary GMPEs without adding special parameters. In conclusion, ANN with deep learning method can be used to ground motion prediction with smaller prediction errors compared with ordinary GMPE, while ANN's applicability to extended faults and the stability are still to be investigated.