Generation of broadband spectra from physics-based simulations using stochastic LSTM network

This study aims to develop a model that predicts high-frequency response spectra and damage-related ground motion parameters using low-frequency physics-based simulations (PBS) for horizontal and vertical components. The model is a stochastic dropout-based long short-term memory (LSTM) network, whic...

Full description

Saved in:
Bibliographic Details
Published in:Engineering applications of artificial intelligence 2023-11, Vol.126, p.106801, Article 106801
Main Authors: Sreenath, Vemula, Sreejaya, K.P., Raghukanth, S.T.G.
Format: Article
Language:English
Subjects:
Broadband spectrum
Ground motion models
LSTM network
Physics-based simulation
Spectrum compatible time histories
Stochastic neural network
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:This study aims to develop a model that predicts high-frequency response spectra and damage-related ground motion parameters using low-frequency physics-based simulations (PBS) for horizontal and vertical components. The model is a stochastic dropout-based long short-term memory (LSTM) network, which accounts for spectra interdependencies and high-frequency spectra’s stochastic nature. Adding anelastic distance as an input term significantly improved the model’s performance. Models are developed for different cutoff periods (validity of PBS low-frequency spectra): 0.75 s, 1 s, 1.5 s, and 2 s, using the 2015 Nepal main and aftershocks and combining them with global near-source strong-motion data. The models are validated using near- and far-field predictions and a good agreement with the recorded data is observed. The mean squared error, mean absolute error, and coefficient of determination for the 0.75 s horizontal model is 0.174, 0.3171, and 0.9295, respectively. Additionally, the study generates spectra and time histories for the 2001 Mw 7.6 Bhuj earthquake, which had no recordings. The obtained spectral values agree well with global stable-continental region models, and the time histories could capture characteristics such as duration, amplitude, and arrival times. •Developed a stochastic network to predict high-frequency spectra from low-frequency.•LSTM network captures frequency dependencies.•Distance and PSA(T = T*) terms have high weightage in the prediction.•Developed broadband time histories to the 2001 Bhuj earthquake.
ISSN:0952-1976
DOI:10.1016/j.engappai.2023.106801