Ground-Motion Prediction Model Based on Neural Networks to Extract Site Properties from Observational Records

Abstract

Choosing the method for inputting site conditions is critical in reducing the uncertainty of empirical ground-motion models (GMMs). We apply a neural network (NN) to construct a GMM of peak ground acceleration that extracts site properties from ground-motion data instead of referring to ground condition variables given for each site. A key structure of the model is one-hot representations of the site ID, that is, specifying the collection site of each ground-motion record by preparing input variables corresponding to all observation sites. This representation makes the best use of the flexibility of NN to obtain site-specific properties while avoiding overfitting at sites where a small number of strong motions have been recorded. The proposed model exhibits accurate and robust estimations among several compared models in different aspects, including data-poor sites and strong motions from large earthquakes. This model is expected to derive a single-station sigma that evaluates the residual uncertainty under the specification of estimation sites. The proposed NN structure of one-hot representations would serve as a standard ingredient for constructing site-specific GMMs in general regions.