pith. machine review for the scientific record. sign in

arxiv: 2512.04694 · v3 · submitted 2025-12-04 · 💻 cs.LG · cs.AI

Recognition: 2 theorem links

· Lean Theorem

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

Baris Yilmaz , Bevan Deniz Cilgin , Erdem Akag\"und\"uz , Salih Tileylioglu
Authors on Pith no claims yet

Pith reviewed 2026-05-17 01:20 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords generative modelsstrong motion generationsite-specific synthesisearthquake ground motionlatent space resamplingcross-regional generalizationdeep learning
0
0 comments X

The pith

A pre-trained generative model produces realistic station-specific earthquake records across regions without fine-tuning or explicit site inputs

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces TimesNet-Gen to generate strong ground motion records that match the characteristics of individual recording stations. It achieves this through a station-restricted resampling strategy in the model's latent space that draws only from records of the target station using a Dirichlet distribution, without feeding in any site condition parameters or reducing dimensions. The model is pre-trained on one large regional dataset and then applied directly to stations in a different region, where it produces outputs whose statistical properties align with real accelerometer data. A sympathetic reader would care because accurate site-specific ground motions are essential for reliable earthquake risk assessment and engineering design in areas where local recordings are sparse.

Core claim

Site-specific strong motion generation is achieved directly through a station-restricted, Dirichlet-based latent space resampling strategy in a deep generative framework, allowing a model pre-trained via self-supervised learning on one regional dataset to synthesize station-specific records from a different region without fine-tuning or explicit conditioning inputs.

What carries the argument

The station-restricted, Dirichlet-based latent space resampling strategy that implicitly encodes local site effects by limiting draws to station-specific records.

If this is right

  • The generated records preserve the essential physical coupling between frequency content and peak amplitude.
  • Strong station-wise alignment occurs in statistical measures without any region-specific retraining.
  • The approach compares favorably to a spectrogram-based conditional variational autoencoder that uses explicit station conditioning.
  • Site-specific synthesis works without requiring site condition data as model inputs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method may allow synthesis for stations that have very few recordings by leveraging latent similarities across the training set.
  • Analogous resampling techniques could transfer to other location-dependent time series such as wind speed or ocean wave records.
  • Testing the model on entirely new tectonic settings would reveal whether the implicit site encoding generalizes beyond the two regions examined.

Load-bearing premise

Restricting latent space resampling to station-specific records via the Dirichlet strategy is enough to capture local site effects without explicit conditioning or dimensionality reduction.

What would settle it

A clear mismatch between the distributions of generated and real records in log-HVSR space, or between their joint peak ground acceleration and fundamental site frequency relations, for multiple stations in the target region would falsify the generalization claim.

Figures

Figures reproduced from arXiv: 2512.04694 by Baris Yilmaz, Bevan Deniz Cilgin, Erdem Akag\"und\"uz, Salih Tileylioglu.

Figure 1
Figure 1. Figure 1: TimesNet-Gen architecture. (seq len+pred len). We apply per-sequence, per-channel stan￾dardization and de-normalize outputs before computing time￾domain losses to preserve the original scale. 1) Conditioning: To be able to generate station (or site)- specific records, we inject station conditioning into the times￾blocks models. By providing one-hot encoded station IDs (s ∈ {0, . . . , S−1}), a conditioning… view at source ↗
Figure 2
Figure 2. Figure 2: TimesNet-Gen sampling via k-sample encoder-feature averaging within a station pool. four consecutive convolutional layers with 3 × 3 kernels and Leaky ReLU activations, followed by a flattening operation and two linear layers that produce the latent mean and standard deviation parameters. Latent variables are sampled using the reparameterization trick and passed to the decoder, which begins with a linear t… view at source ↗
Figure 3
Figure 3. Figure 3: Left: Real samples from the dataset. Right: similar TimesNet-Gen generated. Bottom: Corresponding Fourier amplitude [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average HVSR curves for selected stations. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: f0 distribution Confusion Matrices. Left: TimesNet-Gen (alignment score: 0.93), Right: VAE (alignment score: 0.81). When different HVSR curves show consistent peak frequency and amplitude, they capture stable resonance characteristics of the site, emphasizing the need to represent these peaks accurately. C. f0 Distribution Confusion Matrices Finally, we calculate the f0 Distribution Confusion Matrices for … view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces TimesNet-Gen, a deep generative model for site-specific strong ground motion synthesis from time-domain accelerometer records. It employs self-supervised pre-training on the AFAD dataset followed by a frozen encoder and a station-restricted Dirichlet-based latent space resampling strategy that avoids explicit conditioning or dimensionality reduction. The central claim is robust cross-regional generalization: the model generates station-specific records from the NGA-West2 dataset without fine-tuning, with performance assessed via distributional alignment in log-HVSR space and joint PGA-fundamental frequency statistics, plus favorable comparison to a spectrogram-based conditional VAE baseline.

Significance. If the cross-regional transfer holds, the approach would offer a practical advance in earthquake engineering by enabling site-specific generation from large pre-trained models without region-specific retraining or explicit site parameters. The empirical maintenance of physical coupling between frequency content and amplitude is a positive feature, though the absence of quantitative validation metrics limits immediate impact.

major comments (3)
  1. [Abstract and §5] Abstract and §5 (cross-regional evaluation): the claim that the frozen AFAD-trained encoder produces latents on NGA-West2 that still separate sufficiently by station for Dirichlet resampling to succeed lacks any quantitative check, such as latent-space nearest-neighbor station purity, reconstruction fidelity, or station classification accuracy on the target domain.
  2. [Method] Method section describing the resampling strategy: the manuscript provides no derivation or hyperparameter values for the Dirichlet concentration parameters used in station-restricted resampling, nor any ablation showing that this mechanism (rather than the base TimesNet architecture) is responsible for preserving the PGA-frequency coupling reported in the results.
  3. [Results] Results section on distributional comparisons: alignment in log-HVSR and PGA-f0 joint statistics is asserted but reported without quantitative distances (e.g., Wasserstein or KL divergence), error bars, or statistical significance tests, making it impossible to judge whether the generated distributions are meaningfully closer to real data than the CVAE baseline.
minor comments (2)
  1. [Method] Notation for the latent resampling step is introduced without a clear equation or pseudocode block, forcing the reader to infer the exact sampling procedure from prose.
  2. [Experiments] The baseline CVAE is described as 'explicitly formulated for station-specific latent space modeling' but the conditioning mechanism and loss terms are not compared side-by-side with TimesNet-Gen in a table.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback on our manuscript. We appreciate the opportunity to clarify and strengthen our claims regarding cross-regional generalization and the evaluation of TimesNet-Gen. Below, we provide point-by-point responses to the major comments, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract and §5] Abstract and §5 (cross-regional evaluation): the claim that the frozen AFAD-trained encoder produces latents on NGA-West2 that still separate sufficiently by station for Dirichlet resampling to succeed lacks any quantitative check, such as latent-space nearest-neighbor station purity, reconstruction fidelity, or station classification accuracy on the target domain.

    Authors: We agree that providing quantitative evidence of latent space structure on the target domain would better support the cross-regional generalization claim. In the revised version, we will add an analysis in §5 including latent-space nearest-neighbor station purity metrics and reconstruction fidelity for NGA-West2 records using the frozen encoder. This will demonstrate that station-specific information is preserved without fine-tuning. revision: yes

  2. Referee: [Method] Method section describing the resampling strategy: the manuscript provides no derivation or hyperparameter values for the Dirichlet concentration parameters used in station-restricted resampling, nor any ablation showing that this mechanism (rather than the base TimesNet architecture) is responsible for preserving the PGA-frequency coupling reported in the results.

    Authors: We will include the specific hyperparameter values for the Dirichlet concentration parameters (α = 0.5 for station restriction) and a brief derivation of the resampling strategy in the Method section. Additionally, we will perform and report an ablation study comparing the full model with a version using standard Gaussian resampling to isolate the contribution of the Dirichlet-based mechanism to the observed PGA-fundamental frequency coupling. revision: yes

  3. Referee: [Results] Results section on distributional comparisons: alignment in log-HVSR and PGA-f0 joint statistics is asserted but reported without quantitative distances (e.g., Wasserstein or KL divergence), error bars, or statistical significance tests, making it impossible to judge whether the generated distributions are meaningfully closer to real data than the CVAE baseline.

    Authors: We acknowledge the need for more rigorous quantitative evaluation. In the revised manuscript, we will compute and report Wasserstein distances and KL divergences between the generated and real distributions in log-HVSR space and for the joint PGA-f0 statistics. We will also include error bars from multiple generation runs and perform statistical significance tests (e.g., Kolmogorov-Smirnov tests) to compare against the CVAE baseline. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical training and held-out evaluation

full rationale

The paper trains TimesNet-Gen on AFAD records via self-supervised learning, then evaluates station-specific generation on held-out NGA-West2 records using a Dirichlet-restricted latent resampling strategy without fine-tuning. The central claim of cross-regional generalization is assessed by comparing distributions in log-HVSR and PGA-f0 space against real records and a CVAE baseline. No quantity is defined in terms of itself, no fitted parameter is relabeled as a prediction, and no self-citation or uniqueness theorem is used to force the result. The derivation chain consists of standard model training followed by empirical testing on independent data, remaining self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim depends on the pre-trained model having learned transferable representations and on the resampling strategy implicitly capturing site effects; these rest on standard deep-learning assumptions plus domain-specific data properties.

free parameters (2)
  • Dirichlet concentration parameters for station-restricted resampling
    Control the degree of station-specific restriction in latent space and are determined during pre-training on AFAD.
  • All neural network weights in TimesNet-Gen
    Fitted via self-supervised learning on the AFAD accelerometer dataset.
axioms (2)
  • domain assumption The AFAD pre-training distribution contains sufficient diversity for cross-regional transfer to NGA-West2 without fine-tuning.
    Invoked when claiming robust generalization from the frozen model.
  • domain assumption Log-HVSR space and joint PGA-fundamental frequency distributions are sufficient proxies for physical coupling between frequency content and amplitude.
    Used to evaluate whether generated records preserve essential ground-motion characteristics.

pith-pipeline@v0.9.0 · 5536 in / 1359 out tokens · 31562 ms · 2026-05-17T01:20:11.159881+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 2 internal anchors

  1. [1]

    Two years on: A look at undp’s earth- quake recovery efforts in t ¨urkiye,

    U. N. D. Programme, “Two years on: A look at undp’s earth- quake recovery efforts in t ¨urkiye,” https://www.undp.org/turkiye/news/ two-years-look-undps-earthquake-recovery-efforts-turkiye, 2025, ac- cessed: 2025-09-20

    work page 2025

  2. [2]

    Bayesian-deep-learning estimation of earthquake location from single-station observations,

    S. M. Mousavi and G. C. Beroza, “Bayesian-deep-learning estimation of earthquake location from single-station observations,”IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 11, p. 8211–8224, Nov. 2020. [Online]. Available: http://dx.doi.org/10.1109/ TGRS.2020.2988770 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8

    work page arXiv 2020

  3. [3]

    Deep learning-based average shear wave velocity prediction using accelerometer records,

    B. Yılmaz, M. T ¨urkmen, S. Meral, E. Akag ¨und¨uz, and S. Tileylioglu, “Deep learning-based average shear wave velocity prediction using accelerometer records,” 2024. [Online]. Available: https://arxiv.org/abs/ 2408.14962

    work page arXiv 2024

  4. [4]

    Exploring challenges in deep learning of single-station ground motion records,

    ¨Umit Mert C ¸ a ˘glar, B. Yilmaz, M. T ¨urkmen, E. Akag ¨und¨uz, and S. Tileylioglu, “Exploring challenges in deep learning of single-station ground motion records,” 2025. [Online]. Available: https://arxiv.org/abs/2403.07569

    work page arXiv 2025

  5. [5]

    Data-driven accelerogram synthesis using deep generative models,

    M. A. Florez, M. Caporale, P. Buabthong, Z. E. Ross, D. Asimaki, and M.-A. Meier, “Data-driven accelerogram synthesis using deep generative models,” 2020. [Online]. Available: https://arxiv.org/abs/2011.09038

    work page arXiv 2020

  6. [6]

    TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” 2023. [Online]. Available: https://arxiv.org/abs/2210.02186

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    An introduction to variational autoen- coders,

    D. P. Kingma and M. Welling, “An introduction to variational autoen- coders,”Foundations and Trends in Machine Learning, 2019. [Online]. Available: https://www.nowpublishers.com/article/Details/MAL-056

    work page 2019

  8. [8]

    Deep learning-based epicenter localization using single-station strong motion records,

    M. T ¨urkmen, S. Meral, B. Yilmaz, M. Cikis, E. Akag ¨und¨uz, and S. Tileylioglu, “Deep learning-based epicenter localization using single-station strong motion records,” 2024. [Online]. Available: https://arxiv.org/abs/2405.18451

    work page arXiv 2024

  9. [9]

    A stable algorithm for regression analyses using the random effects model,

    N. A. Abrahamson and R. R. Youngs, “A stable algorithm for regression analyses using the random effects model,”Bulletin of the Seismological Society of America, vol. 82, no. 1, pp. 505–510, 1992

    work page 1992

  10. [10]

    Peak horizontal acceleration and velocity from strong-motion records including records from the 1979 imperial valley, california, earthquake,

    W. B. Joyner and D. M. Boore, “Peak horizontal acceleration and velocity from strong-motion records including records from the 1979 imperial valley, california, earthquake,”Bulletin of the seismological Society of America, vol. 71, no. 6, pp. 2011–2038, 1981

    work page 1979

  11. [11]

    Strong ground motion simulation techniques—a review in world context,

    S. Arora, A. Joshi, P. Kumari, P. Kumar, S. K. Sah, S. Lal, and N. P. Singh, “Strong ground motion simulation techniques—a review in world context,”Arabian Journal of Geosciences, vol. 13, no. 14, p. 673, 2020

    work page 2020

  12. [12]

    Hybrid broadband ground- motion simulations: Combining long-period deterministic synthetics with high-frequency multiple s-to-s backscattering,

    P. M. Mai, W. Imperatori, and K. B. Olsen, “Hybrid broadband ground- motion simulations: Combining long-period deterministic synthetics with high-frequency multiple s-to-s backscattering,”Bulletin of the Seismological Society of America, vol. 100, no. 5A, pp. 2124–2142, 2010

    work page 2010

  13. [13]

    Findings from a decade of ground motion simulation validation research and a path forward,

    S. Rezaeian, J. P. Stewart, N. Luco, and C. A. Goulet, “Findings from a decade of ground motion simulation validation research and a path forward,”Earthquake Spectra, vol. 40, no. 1, pp. 346–378, 2024. [Online]. Available: https://doi.org/10.1177/87552930231212475

    work page doi:10.1177/87552930231212475 2024

  14. [14]

    Seismic labeled data expansion using variational autoencoders,

    K. Li, S. Chen, and G. Hu, “Seismic labeled data expansion using variational autoencoders,”Artificial Intelligence in Geosciences, vol. 1, pp. 24–30, 2020

    work page 2020

  15. [15]

    Learning physics for unveiling hidden earthquake ground motions via conditional generative modeling,

    P. Ren, R. Nakata, M. Lacour, I. Naiman, N. Nakata, J. Song, Z. Bi, O. A. Malik, D. Morozov, O. Azencot, N. B. Erichson, and M. W. Mahoney, “Learning physics for unveiling hidden earthquake ground motions via conditional generative modeling,” 2024. [Online]. Available: https://arxiv.org/abs/2407.15089

    work page arXiv 2024

  16. [16]

    Seismic data augmen- tation based on conditional generative adversarial networks,

    Y . Li, B. Ku, S. Zhang, J.-K. Ahn, and H. Ko, “Seismic data augmen- tation based on conditional generative adversarial networks,”Sensors, vol. 20, no. 23, p. 6850, 2020

    work page 2020

  17. [17]

    Broadband ground-motion synthesis via generative adversarial neural operators: Development and validation,

    Y . Shi, G. Lavrentiadis, D. Asimaki, Z. E. Ross, and K. Azizzadenesheli, “Broadband ground-motion synthesis via generative adversarial neural operators: Development and validation,”Bulletin of the Seismological Society of America, vol. 114, no. 4, pp. 2151–2171, 2024

    work page 2024

  18. [18]

    Site-specific ground-motion waveform generation using a conditional generative adversarial network and generalized inversion technique,

    J. Yamaguchi, Y . Tomozawa, and T. Saka, “Site-specific ground-motion waveform generation using a conditional generative adversarial network and generalized inversion technique,”Bulletin of the Seismological Society of America, vol. 114, no. 4, pp. 2118–2137, 2024

    work page 2024

  19. [19]

    High resolution seismic waveform generation using denoising diffusion,

    A. Bergmeister, K. H. Palgunadi, A. Bosisio, L. Ermert, M. Koroni, N. Perraudin, S. Dirmeier, and M.-A. Meier, “High resolution seismic waveform generation using denoising diffusion,” 2024. [Online]. Available: https://arxiv.org/abs/2410.19343

    work page arXiv 2024

  20. [20]

    Broadband ground motion synthesis by diffusion model with minimal condition,

    J. Jung, J. Lee, C. Jung, H. Kim, B. Jung, and D. Lee, “Broadband ground motion synthesis by diffusion model with minimal condition,”

    work page

  21. [21]

    Available: https://arxiv.org/abs/2412.17333

    [Online]. Available: https://arxiv.org/abs/2412.17333

    work page arXiv

  22. [22]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” inNIPS, 2017, pp. 6626–6637

    work page 2017

  23. [23]

    Rouge: A package for automatic evaluation of summaries,

    C. Lin, “Rouge: A package for automatic evaluation of summaries,” in Proc. ACL Workshop on Text Summarization Branches Out, 2004, p. 10

    work page 2004

  24. [24]

    Stochastic ground motion simulation of the 2016 meinong, taiwan earthquake,

    C.-T. Chen, S.-C. Chang, and K.-L. Wen, “Stochastic ground motion simulation of the 2016 meinong, taiwan earthquake,”Earth, Planets and Space, vol. 69, no. 1, p. 62, 2017

    work page 2016

  25. [25]

    Generative adversarial networks-based ground-motion model for crustal earthquakes in japan considering detailed site conditions,

    Y . Matsumoto, T. Yaoyama, S. Lee, T. Hida, and T. Itoi, “Generative adversarial networks-based ground-motion model for crustal earthquakes in japan considering detailed site conditions,”Bulletin of the Seismolog- ical Society of America, vol. 114, no. 6, pp. 2886–2911, 2024

    work page 2024

  26. [26]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”

    work page

  27. [27]

    Auto-Encoding Variational Bayes

    [Online]. Available: https://arxiv.org/abs/1312.6114

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Unsupervised clustering of seismic signals using deep convolutional autoencoders,

    S. M. Mousavi, W. Zhu, W. Ellsworth, and G. Beroza, “Unsupervised clustering of seismic signals using deep convolutional autoencoders,” IEEE Geoscience and Remote Sensing Letters, vol. 16, no. 11, pp. 1693– 1697, 2019

    work page 2019

  29. [29]

    A method for dynamic characteristics estimation of subsurface using microtremor on the ground surface,

    Y . Nakamura, “A method for dynamic characteristics estimation of subsurface using microtremor on the ground surface,”Railway Technical Research Institute, Quarterly Reports, vol. 30, no. 1, 1989

    work page 1989

  30. [30]

    Site effect evaluation using spectral ratios with only one station,

    J. Lermo and F. J. Ch ´avez-Garc´ıa, “Site effect evaluation using spectral ratios with only one station,”Bulletin of the seismological society of America, vol. 83, no. 5, pp. 1574–1594, 1993

    work page 1993

  31. [31]

    A new set of automated methodologies for estimating site fundamental frequency and its uncer- tainty using horizontal-to-vertical spectral ratio curves,

    M. Yazdi, R. Motamed, and J. G. Anderson, “A new set of automated methodologies for estimating site fundamental frequency and its uncer- tainty using horizontal-to-vertical spectral ratio curves,”Seismological Society of America, vol. 93, no. 3, pp. 1721–1736, 2022

    work page 2022

  32. [32]

    S. L. Kramer and J. P. Stewart,Geotechnical earthquake engineering. CRC Press, 2024. Barıs ¸ Yılmazreceived the B.S. and M.S. de- grees in Electrical and Electronics Engineering from C ¸ ankaya University, Ankara, Turkey, in 2019 and 2022, respectively. From 2020 to 2022, he worked as a Research Assistant in the Department of Electrical and Electronics En...

    work page 2024

  33. [33]

    His research interests include geotechnical earthquake engineering, soil-structure interac- tion, seismology, and deep learning applications in earthquake engineering

    He is currently with the Department of Civil Engineering at Kadir Has University. His research interests include geotechnical earthquake engineering, soil-structure interac- tion, seismology, and deep learning applications in earthquake engineering

    work page