SITE AMPLIFICATION PREDICTION MODEL OF SHALLOW BEDROCK SITES - BASED ON MACHINE LEARNING MODELS

Aug 3, 2026

Upload by Jiven

Prediction of site amplification is essential for accurate site-specific seismic hazard assessment. Traditional empirical and simulation-conditioned models rely on simplified proxies such as time-averaged shear-wave velocity (VS30) and site period (TG), yielding only approximate estimates. This study builds site amplification prediction models using random forest (RF) and deep neural networks (DNN) trained on simulation outputs from shallow bedrock profiles. Using response-spectrum matrix inputs and shear-wave velocity profiles, the ML models achieve higher accuracy for both linear and nonlinear amplification, producing reliable binned means and standard deviations, with DNN performing best.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages39

Reading Time1 h 38 min

AccessFree to download

Abstract

1. Introduction

Seismic site amplification concept

Limitations of regression-based models

Prior machine learning studies

Related neural network and feature choices

1 SITE AMPLIFICATION PREDICTION MODEL OF SHALLOW BEDROCK SITES 2 BASED ON MACHINE LEARNING MODELS 3 Yong-Gook Leea, Sang-Jin Kimb, ZeinepAchmetc, Oh-Sung Kwond, Duhee Parka *, and Luigi 4 Di Sarnoe 5 6 ABSTRACT 7 Prediction of the site amplification is of primary importance for a site-specific seismic hazard 8 assessment. A large suite of both empirical and simulation-based site amplification models has 9 been proposed. Because they are conditioned on a few simplified site proxies including time- 10 averaged shear wave velocity up to a depth of 30 m ( VS30) and site period ( TG), they only 11 provide approximate estimates of the site amplification. In this study, site amplification 12 prediction models are developed using two machine learning algorithms, which are random 13 forest (RF) and deep neural network (DNN) . A comprehensive database of site response 14 analysis outputs obtained from simulations performed on shallow bedrock profiles is used. 15 Instead of simplified site proxies and ground motion intensity measures, matrix data which 16 include the response spectrum of the input ground motion and shear wave velocity profile. Both 17 machine learning based models provide exceptional prediction accuracies of both the linear 18 and nonlinear amplifications compared with the regression-based model, producing accurate 19 predictions of both binned mean and standard deviation of the site amplification. Among two 20 machine learning techniques, DNN-based model is revealed to produce better predictions. 21 Keywords: Machine learning, random forest, deep neural network, site amplification, site 22 response analysis. a * b c d e Department of Civil and Environmental Engineering, Hanyang University, Seoul, Korea Corresponding author. Email: [dpark@hanyang.ac.kr](dpark@hanyang.ac.kr) (D. Park). Infrastructure Division, Infrastructure Engineering Group, Infrastructure Geotechnical Engineering Team, Hyundai Engineering and Construction Group, Seoul, Korea. Department of Civil Engineering, National Technical University of Athens, Zografou, Greece. Department of Civil & Mineral Engineering, University of Toronto, Toronto, Canada. Department of Civil Engineering and Industrial Design, University of Liverpool, Liverpool, UK 23 1. INTRODUCTION 24 The vertically propagating shear waves are generally amplified as they radiate upwards from 25 the bedrock through soil layers, which have relatively lower stiffness and density. This 26 phenomenon is referred to as the seismic site amplification. The prediction of site-specific 27 seismic amplification is critical for estimation of the design ground motion and seismic design 28 of various types of structures and facilities. 29 Regression-based site amplification models that are linked to ground motion models (GMMs) 30 have been developed from both recorded ground motions and numerical simulation outputs. 31 The models are based on site proxies, which include the time-averaged shear wave velocity of 32 top 30 m ( VS30), depth at which shear wave velocity ( VS) reaches 1 km/s or greater (Z 1), and 33 natural site period ( TG) . They are also conditioned on motion proxies including peak ground 34 acceleration (PGA) and spectral acceleration (SA) at selected periods. Although widely used 35 because of their ease of use, the regression-based site amplification models inevitably contain 36 large levels of uncertainty. 37 A number of studies proposed to use machine learning (ML) algorithms instead of regression 38 equations to develop site amplification models [1-7] . Kamatchi et al. [3] developed an artificial 39 neural network (ANN)-based methodology to predict site-specific acceleration response using 40 the outputs from one-dimensional (1D) equivalent linear (EQL) site response analyses 41 performed for a selected site in Delhi, India. The input parameters of the ANN model were the 42 moment magnitu

FAQs

Why is site amplification prediction important in seismic hazard assessment?

Site-specific seismic amplification controls the design ground motion and supports seismic design of structures and facilities, making prediction critical for hazard evaluation.

What is the main limitation of regression-based site amplification models?

They depend on simplified site proxies such as VS30 and TG and motion proxies like PGA and spectral acceleration, which introduce substantial uncertainty and only approximate amplification estimates.

How do the proposed machine learning models improve prediction accuracy?

They use simulation-derived databases and richer input representations, including the response spectrum of the input ground motion and the shear-wave velocity profile. RF and DNN outperform regression-based approaches for both linear and nonlinear amplification, with DNN showing better results.

Aug 3, 2026

Upload by Jiven