PaperlayerView exact revision on arXiv

Source-linked AI summary

TimesNet-Gen: Deep Learning-based Site Specific Strong Motion Generation

Baris Yilmaz, Bevan Deniz Cilgin, Erdem Akagündüz, Salih Tileylioglu

arXiv:2512.04694v3cs.LGcs.AI

TL;DR

Earthquakes have historically led to significant loss of life and extensive economic damage. TimesNet-Gen uses station-restricted, Dirichlet-based latent resampling to generate site-specific time-domain ground motions. The frozen AFAD-trained encoder forms station-specific latent banks for NGA-West2 without regional fine-tuning, and TimesNet-Gen outperforms the spectrogram-based conditional VAE in capturing sharp resonance peaks and non-stationary temporal patterns.

  • Problem

    Earthquakes have historically led to significant loss of life and extensive economic damage.

    Source: §I · p. 1

  • Method

    TimesNet-Gen uses station-restricted, Dirichlet-based latent resampling to generate site-specific time-domain ground motions.

    Source: §I · pp. 1–2, §IV · p. 10

  • Results

    The frozen AFAD-trained encoder forms station-specific latent banks for NGA-West2 without regional fine-tuning, and TimesNet-Gen outperforms the spectrogram-based conditional VAE in capturing sharp resonance peaks and non-stationary temporal patterns.

    Source: §IV · pp. 10, 12

  • Takeaways & Limitations

    Generated records remain consistent with empirical coupling between site resonance and peak intensity under the reported log-HVSR and joint f0–PGA analyses.

    Source: §IV · p. 12

    The method requires empirical recordings at the target station and does not explicitly condition on magnitude or rupture distance, limiting it to probabilistic rather than scenario-controlled generation.

    Source: §IV · p. 12

Abstract

from arXiv · show

Effective earthquake risk reduction relies on accurate site-specific evaluations, which require models capable of representing the influence of local site conditions on ground motion characteristics. We address strong ground motion generation from time-domain accelerometer records and introduce the TimesNet-Gen, a deep generative framework. In this framework, site-specific generation is directly achieved through a station-restricted, Dirichlet-based latent space resampling strategy, without relying on explicit conditioning inputs or dimensionality reduction. Pre-trained on the AFAD dataset via self-supervised learning, the frozen model demonstrates robust cross-regional generalization by successfully generating station-specific NGA-West2 records without any fine-tuning. Model performance is evaluated by comparing the distributions of generated and real records in the log-HVSR space, alongside the joint analysis of peak ground acceleration and fundamental site frequency. As a baseline, we construct a spectrogram-based conditional variational autoencoder (CVAE) explicitly formulated for station-specific latent space modeling. The results show strong station-wise alignment, consistent cross-regional ground motion synthesis, and a favorable comparison with a spectrogram-based conditional variational autoencoder baseline, demonstrating that the model empirically maintains the essential physical coupling between frequency content and peak amplitude. Our codes are available via https://github.com/brsylmz23/TimesNet-Gen.

I. INTRODUCTION

Site-specific ground-motion generation remains difficult because recordings contain complex temporal and spectral patterns that existing models may not capture. TimesNet-Gen addresses this gap with station-restricted latent resampling and evaluates transfer across regional datasets.

I. INTRODUCTION

  • Local geology can modify ground-motion amplitude, frequency content, and duration, while existing simulations and prediction equations may not capture nonlinear, site-dependent behavior across regions.The paper motivates data-driven generation as a way to learn from recordings without restrictive parametric assumptions.Source: §I · p. 1
  • TimesNet-Gen adapts TimesNet for time-domain ground-motion generation using station-restricted, Dirichlet-based latent resampling and coupled aggregation.The paper also evaluates transfer from AFAD to NGA-West2 and assesses fidelity using log-HVSR distributions and joint f0–PGA analysis.Source: §I · pp. 1–2

A. Related Work

  • The related-work discussion frames classical seismic synthesis as a set of approaches with assumptions and computational costs that can limit reproduction of strong-motion complexity.The paper situates its generation approach among empirical, semi-empirical, and physics-based methods.Source: A. Related Work · p. 2, 1) Classical Methods on Seismic Data… · p. 2

1) Classical Methods on Seismic Data Synthesis:

  • Prior seismic-synthesis methods span empirical, semi-empirical, and physics-based approaches, which face trade-offs in computational cost and representation of fine-scale variability and nonstationarity.More realistic nonstationary modeling generally costs more, while faster empirical and semi-empirical methods capture only limited aspects.Source: 1) Classical Methods on Seismic Data… · p. 2

2) Deep Learning-based Seismic Data Generation:

Deep generative models offer data-driven alternatives for seismic synthesis, but evaluating realism and physical consistency remains challenging. The paper adapts TimesNet for time-domain reconstruction and station-specific generation, and assesses generated records through spectral and amplitude-frequency comparisons.

2) Deep Learning-based Seismic Data Generation:

  • Deep generative seismic models reduce reliance on restrictive parametric assumptions, but no broadly accepted, physically comprehensive foundation model has been established.VAEs, GANs, and diffusion models offer different capabilities and challenges, including GAN training instability and diffusion computational cost.Source: 2) Deep Learning-based Seismic Data… · p. 2

3) Evaluation:

  • No established standard evaluates generated seismic data, and the paper combines log-HVSR distribution comparisons with joint f0–PGA analysis to assess spectral envelopes and physical dependence.The evaluation is intended to test whether generated motions preserve relationships between spectral content and amplitude.Source: 3) Evaluation · pp. 2–3

II. METHODOLOGY

  • The benchmark compares TimesNet-Gen with a conditional VAE and presents the architectures, station-conditioning method, preprocessing, and datasets used.The baseline is described as a conditional VAE for reconstruction and conditional generation.Source: §II · p. 3

A. TimesNet-Gen

  • TimesNet-Gen uses TimesNet’s period-aware 2D representations to model intra- and inter-period patterns, then reconstructs time-domain waveforms and generates station-specific records through latent resampling.The model preserves the encoder, adds a linear reconstruction head, and uses a high-dimensional latent representation; seismic signals’ mixed wave components motivate period-aware modeling.Source: A. TimesNet-Gen · p. 3

1) Conditioning and Dirichlet-based Latent Space Sampling

TimesNet-Gen creates station-specific outputs by drawing from the target station’s latent bank and resampling selected representations. It applies the same Dirichlet weights to latent features and their scaling parameters to preserve their physical dependency.

1) Conditioning and Dirichlet-based Latent Space Sampling

  • Station identity is enforced by restricting latent draws to the target station’s bank rather than adding explicit class embeddings at the encoder.The Dirichlet-based resampling strategy also controls generation diversity.Source: 1) Conditioning and Dirichlet-based… · pp. 3–4
  • The model builds station-specific latent banks from self-supervised representations and randomly bootstrap-samples up to five latent vectors for generation.The selected count is k=min(5,N), where N is the number of available vectors for that station.Source: 1) Conditioning and Dirichlet-based… · p. 4
  • Applying identical Dirichlet weights to latent features and per-record scaling parameters maintains spectral-content and amplitude dependency while preserving target-station variability.The paper reports that coupled aggregation mitigates physically inconsistent amplitude scaling and preserves the station’s aleatory variability and amplitude bounds.Source: 1) Conditioning and Dirichlet-based… · p. 4

B. Variational Autoencoders

The baseline is a convolutional variational autoencoder that reconstructs spectrogram inputs through a symmetric deconvolutional decoder. Its station-class latent space is conditioned first with class-specific priors, then refined by clustering.

B. Variational Autoencoders

  • The VAE uses a four-layer convolutional encoder and a symmetric four-layer deconvolutional decoder to reconstruct inputs.The encoder predicts latent mean and standard deviation; the decoder ends with a sigmoid output layer.Source: B. Variational Autoencoders · p. 5

1) Latent Space Conditioning:

  • In the baseline VAE’s first sub-phase, conditioning uses one-hot class priors, assigning each class a distinct prior mean vector formed by tiling one-hot encodings across latent dimensions; prior variance is fixed.Source: 1) Latent Space Conditioning · p. 5
  • The latent dimension is 510, chosen as a multiple of the five classes to keep their priors separable.The KL-divergence term is given in the displayed equation.Source: 1) Latent Space Conditioning · p. 5
  • In phase two, encoded samples are clustered by class, and a sharpened target distribution guides refinement through a KL clustering loss weighted by 100.Initial cluster centers are class-wise sample means; the clustering loss is added to the original VAE objective.Source: 1) Latent Space Conditioning · p. 5

2) Input Preprocessing:

Preprocessing converts each three-channel seismic record into amplitude and phase spectrograms, then reverses that transformation to compare generated signals in the time domain.

2) Input Preprocessing:

  • Short-time Fourier transforms of all three channels produce amplitude and phase spectrograms for each record.The resulting spectrograms form a multi-channel representation intended to retain amplitude and phase information.Source: 2) Input Preprocessing · p. 5
  • Amplitude spectrograms are converted to logarithmic decibels, while wrapped phase is unwrapped to remove time-axis discontinuities.The stated aim is to support learning while preserving signal structure.Source: 2) Input Preprocessing · p. 5
  • For time-domain comparison, the spectrograms are denormalized, amplitude is returned to linear scale, and amplitude and phase are combined before inverse STFT.Source: 2) Input Preprocessing · p. 5

C. Datasets

The study uses AFAD for self-supervised training and primary evaluation, and a geographically distinct NGA-West2 subset to assess cross-regional transfer. Selected stations span differing site properties and fundamental frequencies.

1) AFAD Dataset

  • The AFAD dataset supplies 36,417 strong-motion records from 2012–2018 for self-supervised training and primary evaluation.Source: 1) AFAD Dataset · p. 5
  • Five AFAD stations with differing site properties and fundamental frequencies are selected for generative-fidelity analysis.Fundamental site frequencies are calculated from strong-motion signals using HVSR analysis; the selected stations and frequencies are listed in Table I.Source: 1) AFAD Dataset · p. 6, Table I · p. 6

2) NGA-West2 Dataset (Cross-Regional Evaluation)

D. Evaluation

Evaluation combines spectral analysis with distribution-level metrics to assess site-specific generation and interstation discrimination. It extracts site frequency and horizontal peak acceleration, then compares station-wise distributions.

D. Evaluation

  • The evaluation framework combines spectral analysis and distribution-level metrics to assess generation quality and physical realism.Source: D. Evaluation · p. 6

1) Spectral Analysis

  • HVSR curves are computed from the average of the two horizontal amplitude spectra divided by the vertical spectrum; f0 is the peak frequency within 1–20 Hz.Horizontal PGA is extracted as the maximum magnitude of the combined horizontal acceleration vectors over the record.Source: 1) Spectral Analysis · p. 6

2) Interstation Discrimination

  • Station-wise f0 distributions are compared using Jensen–Shannon divergence converted to a similarity score, then interstation discrimination is assessed against an ideal identity-block matrix using NCC.The similarity scale assigns 1 to perfect similarity and 0 to none.Source: 2) Interstation Discrimination · p. 6

3) Physical Consistency

The evaluation checks whether generated records preserve spectral structure and the relationship between frequency content and peak amplitude. It also examines latent-space coverage and waveform characteristics in AFAD results.

3) Physical Consistency

  • Log-HVSR histograms compare generated and empirical spectral envelopes, with higher SJS indicating closer reproduction.The similarity is based on Jensen-Shannon distance between smoothed, normalized histogram distributions.Source: 3) Physical Consistency · pp. 6–7
  • The second consistency check compares generated and real records in the joint f0–PGA space to assess their frequency–amplitude dependency.The paper describes this relationship as strongly coupled in empirical seismic events.Source: 3) Physical Consistency · p. 7

III. EXPERIMENTS AND RESULTS

  • The experiments assess AFAD latent-space characteristics, spectral consistency, and interstation discrimination before testing transferability on NGA-West2.The latent-space projection for station 1716 shows generated samples covering the empirical domain without isolated, non-physical artifacts, although t-SNE is not strictly metric-preserving.Source: §III · p. 7, 1) Latent Space Representation · p. 7, Figure 2 · p. 7

1) Latent Space Representation

  • Generated E-W waveforms show realistic durations and amplitudes, with Fourier amplitude spectra consistent with empirical data.Figure 3 compares real and generated samples.Source: 1) Latent Space Representation · p. 7

2) Spectral Consistency

TimesNet-Gen reproduces AFAD site-frequency distributions and resonance peaks, while distinguishing stations with similar f0 values more successfully than the VAE baseline. Cross-region tests report latent-bank coverage and alignment of NGA-West2 spectral responses.

2) Spectral Consistency

  • For average HVSR curves, TimesNet-Gen captures distinct peak behaviors and sharp resonance characteristics, whereas the VAE deviates significantly from actual peaks and struggles to reproduce sharp spectral features.Sharp resonance characteristics are reported at stations 0205 and 4628.Source: 2) Spectral Consistency · p. 7
  • TimesNet-Gen distinguishes stations with comparable f0 values, whereas the VAE tends to mix them.The examples include station pairs 0205/4628 and 1716/2020, whose overall site responses can differ despite similar f0 values.Source: 2) Spectral Consistency · pp. 7–8

B. Cross-Regional Evaluation (NGA-West2 Dataset)

1) Latent Space Transferability

  • Across selected NGA-West2 stations, generated HVSR curves reproduce the overall shape and predominant resonance peaks of empirical curves.The paper connects this alignment to learned fundamental site-response characteristics and natural variability relevant to probabilistic seismic hazard applications.Source: 1) Latent Space Transferability · p. 8

3) Joint Distribution of Site Resonance and Peak Ground Acceleration

The cross-regional evaluation tests whether generated records preserve the relationship between site resonance and peak shaking, and whether stations remain distinguishable without regional fine-tuning. The reported results show qualitative f0–PGA overlap and cross-regional station alignment.

3) Joint Distribution of Site Resonance and Peak Ground Acceleration

  • Generated and empirical records show significant qualitative overlap in the f0–horizontal PGA distribution across selected NGA-West2 stations, preserving their reported physical coupling.The passage characterizes the aligned records as remaining within physically plausible bounds for target environments.Source: 3) Joint Distribution of Site… · pp. 8, 10

4) Cross-Regional Interstation Discrimination

  • TimesNet-Gen distinguishes Southern California stations with comparable fundamental frequencies, achieving an alignment score of 0.82 against an ideal identity-block matrix without regional fine-tuning.The analysis uses empirical and generated distribution JSD similarity scores.Source: 4) Cross-Regional Interstation… · p. 10

IV. CONCLUSIONS

TimesNet-Gen generates site-specific strong-motion records in the time domain through station-restricted latent resampling. Evaluations report empirical resonance–peak-intensity consistency, cross-regional transfer, and stronger capture of sharp resonance peaks and non-stationary patterns than a spectrogram-based conditional VAE.

IV. CONCLUSIONS

  • The frozen AFAD-trained encoder maps NGA-West2 records into station-specific latent banks without regional fine-tuning or weight updates.The evaluation covers a geographically and tectonically distinct dataset.Source: §IV · pp. 10, 12
  • Dirichlet-based sampling couples latent embeddings with instance-level scaling, mitigating physically inconsistent amplitude scaling while preserving target-station variability.The resampling uses the target station’s empirical latent bank.Source: §IV · p. 10
  • Synthesized records remain consistent with empirical coupling between site resonance and peak intensity, while outperforming the spectrogram-based conditional VAE on sharp resonance peaks and non-stationary temporal patterns.The evaluations include f0 distributions and average HVSR curves, as well as joint f0–PGA distributions.Source: Figure 7 · p. 11, §IV · p. 12, Figure 8 · p. 12
  • TimesNet-Gen can support single-station pipelines, probabilistic catalogs checked against intensity measures and HVSR, and data augmentation at sparse stations.The paper also describes it as a pre-design exploratory synthesis pool and a training-data enrichment tool.Source: §IV · p. 12
  • The method requires empirical recordings at the target station and lacks explicit magnitude or rupture-distance conditioning, limiting it to probabilistic rather than scenario-controlled generation.The authors note this constrains direct use at uninstrumented sites.Source: §IV · p. 12
  • Future work will add physical constraints such as magnitude, rupture distance, and Vs30, while positioning the framework as complementary to—not a replacement for—physics-based simulations.The authors propose broader benchmarking against simulations under shared scenarios once explicit conditioning is achieved.Source: §IV · p. 13
Loading 2512.04694v3…