A Practical DNN-Based Framework for Estimating High-Frequency Ground-Motion Fourier Amplitude Spectra from Low-Frequency Inputs

ABSTRACT Reliable estimation of high-frequency (HF) seismic content from low-frequency (LF) inputs is crucial for seismic hazard analysis as it extends the usable frequency range of deterministic simulations. Deep neural networks (DNNs) are particularly well suited for this task owing to their capacity for feature extraction and nonlinear mapping. However, when applied to this task, DNN approaches often overlook two intrinsic characteristics of strong-motion datasets: (1) the inherent event–station relational structure, which can lead to the leakage of source or site information under random train-test splits, thereby inflating the apparent generalization performance; and (2) the scarcity of large-magnitude records, which does not align with the demands of earthquake engineering. To address these issues, we define two dedicated tasks—station generalization and magnitude extrapolation—and develop a DNN framework for LF-to-HF estimation of ground-motion Fourier amplitude spectra (FAS), building on Aquib and Mai (2024). Our framework incorporates leakage-free partitioning, dedicated parameter encoders, and performance-oriented design choices. Model performance is quantified using a goodness-of-fit (GOF) metric. Evaluated on the Kiban–Kyoshin network dataset, our model achieves an average GOF of 4.95 for station generalization—an improvement of 1.4 compared with previous methods on unseen stations—with 74.7% of predictions rated as at least “fair” (GOF ≥ 4). For the magnitude extrapolation task, the model attains an average GOF of 4.96, with 67.7% of predictions meeting the “fair” threshold. These results indicate that the model can generalize to new stations and extrapolate to larger magnitudes. Finally, a systematic evaluation reveals a strong dependence on training set size, the upper limit of the input band of FAS, and the magnitude distribution of the training set. These findings elucidate inherent limitations in current DNN-based seismic applications, providing a roadmap for their reliable implementation.

Authors

Institutions

Publication Details

Journal
Bulletin of the Seismological Society of America
Published
2026-11-16
DOI
https://doi.org/10.1785/0120260129
Primary Topic
Seismic Performance and Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Practical DNN-Based Framework for Estimating High-Frequency Ground-Motion Fourier Amplitude Spectra from Low-Frequency Inputs

Zhinan Xie, Haotian Li, Wenyue Zhang, Jianqi Lu
4 links
Bulletin of the Seismological Society of America
Seismic Performance and Analysis
article

A Practical DNN-Based Framework for Estimating High-Frequency Ground-Motion Fourier Amplitude Spectra from Low-Frequency Inputs

Zhinan Xie, Haotian Li, Wenyue Zhang, Jianqi Lu
article en

Abstract

ABSTRACT Reliable estimation of high-frequency (HF) seismic content from low-frequency (LF) inputs is crucial for seismic hazard analysis as it extends the usable frequency range of deterministic simulations. Deep neural networks (DNNs) are particularly well suited for this task owing to their capacity for feature extraction and nonlinear mapping. However, when applied to this task, DNN approaches often overlook two intrinsic characteristics of strong-motion datasets: (1) the inherent event–station relational structure, which can lead to the leakage of source or site information under random train-test splits, thereby inflating the apparent generalization performance; and (2) the scarcity of large-magnitude records, which does not align with the demands of earthquake engineering. To address these issues, we define two dedicated tasks—station generalization and magnitude extrapolation—and develop a DNN framework for LF-to-HF estimation of ground-motion Fourier amplitude spectra (FAS), building on Aquib and Mai (2024). Our framework incorporates leakage-free partitioning, dedicated parameter encoders, and performance-oriented design choices. Model performance is quantified using a goodness-of-fit (GOF) metric. Evaluated on the Kiban–Kyoshin network dataset, our model achieves an average GOF of 4.95 for station generalization—an improvement of 1.4 compared with previous methods on unseen stations—with 74.7% of predictions rated as at least “fair” (GOF ≥ 4). For the magnitude extrapolation task, the model attains an average GOF of 4.96, with 67.7% of predictions meeting the “fair” threshold. These results indicate that the model can generalize to new stations and extrapolate to larger magnitudes. Finally, a systematic evaluation reveals a strong dependence on training set size, the upper limit of the input band of FAS, and the magnitude distribution of the training set. These findings elucidate inherent limitations in current DNN-based seismic applications, providing a roadmap for their reliable implementation.

Bulletin of the Seismological Society of America
Institute of Engineering Mechanics, China Earthquake Administration (CN), Ministry of Emergency Management of the People's Republic of China (CN)
Seismic Performance and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon