betaFeedback

WorksComputer Science

Are Seismic Foundation Models More Than Glorified Feature Extractors?

Abstract

Earthquake sequences are stochastic on two separated timescales: the seismic waves of any individual earthquake propagate over seconds to minutes, while the earthquakes themselves arrive as a sparse, self-exciting occurrence process, with inter-event intervals from seconds to months. Respecting both scales in one training objective is the open challenge. Foundation models (FMs) for seismic waveforms instead import their learning targets from audio and vision, where every target is computed from a single fixed-length window. Such a window holds one realization of the wave propagation (an earthquake trace), never the occurrence process that decides when earthquakes happen. We develop nine model designs sharing one Mamba encoder-decoder, trained on a fixed corpus of 3M windows, varying the self-supervised objective, quantization bottleneck and mixer; the two key comparisons are architecture-matched, varying the objective alone. Each checkpoint is evaluated frozen on three downstream tasks. Varying the target redistributes downstream skill rather than raising it, and no checkpoint wins everywhere. Adding masking pressure to our strongest classifier re-specializes it rather than making it a generalist: it gains 6.9 points of few-shot P-wave picking precision and loses 9.6 points of few-shot classification accuracy. By the standard imported from language and vision, these models are feature extractors rather than FMs. Controls on capacity, architecture, and domain proximity do not explain why skill moves rather than rises. Against task-specific baselines the advantage is regime-dependent: our checkpoints beat them few-shot on classification and ground-motion prediction, and trail them on full data. We argue the binding constraint is the target, and that settling the question requires a learning target defined across events, the timescale that makes seismicity stochastic in the first place.

Introduction

Seismicity is a canonical stochastic dynamical system at two distinct levels. Within a recorded earthquake trace, ground motion is a continuous, non-stationary process governed by the wave equation [5], with randomness from source and path heterogeneity, site response, and ambient noise [4]. Across the catalog (the record of past events with their times, locations and magnitudes), earthquakes arrive as a sparse, self-exciting point process: inter-event intervals span seconds during an aftershock sequence to months during interseismic quiescence [33], and magnitudes follow the Gutenberg–Richter law [12], so the events that matter most (large magnitude events) are by construction the rarest. The modeling tradition for the second level, like ETAS [24] and its neural descendants [8, 21, 37], is explicitly a stochastic-dynamics tradition. Characterizing a sequence means saying when the next event will come and how large it will be: quantities that live in the catalog, not in any one trace. Machine learning has entered at the first level, the waveforms themselves. These are continuous ground-motion recordings from networks of seismometers irregularly placed at the surface, cut into fixed-length windows of three orthogonal components (vertical and two horizontal), typically sampled at 100 or 200 Hz. Recording never stops, but only the sparse windows around catalogued events carry labels, and the curated benchmarks used for pretraining are smaller still: on the order of a million traces [1, 23]. Recent foundation models (FMs) for seismic waveforms [6, 7, 18, 19, 20, 28, 29, 34] import self-supervised recipes from language, vision, and audio domains [2, 9, 13]. These recipes differ in architecture and objective, but agree on something more basic and rarely stated: the learning target is always a functional of a single fixed-length window. Masked prediction [20], compressed reconstruction [18], contrastive alignment [29], and diffusion denoising [7] all ask the model to recover or discriminate one realisation of the observation process, on a fixed clock; none references the occurrence dynamics, which bounds the abilities of the FM. When the physics is local to a short window, a window-level target is enough, and phase picking is the clear case, except when signal of interest is the pattern of occurrence (§ 4).

We ask what that choice costs, under conditions tight enough to attribute the answer: corpus, heads and splits fixed throughout, and, for the two comparisons we build on, architecture fixed too with the objective varied alone. The finding is negative, and we argue it is diagnostic. Varying the window-level target redistributes downstream skill rather than raising it. Controls on capacity, architecture, and domain proximity do not explain the ceiling. That points at the target without saying which property of it is at fault, so we distinguish throughout what we demonstrate from what we motivate.

Setup

We pretrain nine models on one shared backbone: a SEANet-style [32] convolutional encoder–decoder whose every stage is paired with a block of Mamba selective state-space layers [11], operating on five-minute, three-component 100 Hz traces. Mamba is chosen over attention for linear-time scaling at these lengths; SEANet is an audio codec, so the backbone is itself one of the imports we examine. It is held fixed, to isolate the target rather than the architecture. Details on the backbone are in Appendix § 6.§ 1. These are small models by FM standards, 2.8–6.0M parameters, hence the 57M scaling control (§ 3, with the recipe in Appendix § 6.§ 5). Pretraining draws on CREW [1], ≈1.6M regional earthquake waveforms, paired batch-for-batch with an ambient-noise corpus we assembled from 462 PDSN networks (≈1.5M training windows), so the model reconstructs one pure seismic noise window per earthquake window. Appendix § 6.§ 3 gives full detail about the corpora.

The nine checkpoints span three axes, not one. Objective: reconstruction (Rec) composed with a magnitude head (Mag), an adversarial discriminator (Disc), and span masking (Mask). Bottleneck: a continuous latent, or residual vector quantisation (RVQ) updated by backpropagation or EMA. Mixer: SSM-only, or SSM plus a gated-MLP branch (+MLP). Two sub-comparisons are architecture-matched, and they are the ones the argument rests on: the four Rec-family rows share one architecture and differ only in loss composition, and +MaskRaw and +MaskCode match RVQ-EMA+MLP in state-dict shape, adding only a training-time objective that is discarded at inference. See Appendix § 6.§ 2 for full objectives detail. A seismic FM could in principle serve the whole monitoring pipeline [35] and the study of how seismicity evolves [17, 27]. We attempt three of these tasks, listed below.

Each checkpoint is frozen and evaluated with lightweight heads on three tasks (see Appendix § 6.§ 4 for more details about the downstream tasks): foreshock/aftershock (earthquakes before or after the largest event of a sequence) classification at station NRCA around the 2016 M6.5 Norcia mainshock [17]; regression of five peak ground-motion measures for early warning over a 39-station Central-Italy network [3, 15]; and detection of P/S arrival times (picking the arrival of fast primary and slower secondary waves) [36] on STEAD [23]. Each is scored full-data and few-shot (≈100 labels) against the published task-specific model trained directly on it [3, 17, 36]. The encoder sees none of the three downstream corpora during pretraining, so every downstream task is a cross-corpus generalisation test: what we report is specialisation under distribution shift, the regime FMs are meant to serve.

Results

No target wins everywhere: objective composition dominates Table 1, and dominates differently per task. RVQ-EMA+MLP is our best classifier in both regimes (96.7, 85.5) yet is beaten on ground motion and S-picking by Rec+Disc+Mag+Mask (PGA 0.161 vs. 0.306 full-data), which in turn trails it by 9.5 points full-data. The effect is large: adding adversarial, magnitude and masking terms to reconstruction raises few-shot P-precision from 18.8 to 83.5, and larger than any architectural change we tested. Baselines are likewise task-dependent: our best checkpoint clears the task-specific CNN [17] few-shot (85.5 vs. 82.4) but trails it full-data (96.7 vs. 99.2), so pretraining helps where labels are scarcest and not otherwise. Ground motion replicates this on the head that admits a comparison: against the GCN baseline [3] every checkpoint wins on all five targets few-shot, whereas full-data only Rec+Disc+Mag+Mask does, the rest on two of five at best, and four of them on none. On picking the two heads are close, splitting by wave: the baseline [36] slightly ahead on P (0.4 points full-data, 3.3 few-shot), ours slightly ahead on S (≤ 0.4, seven of nine full-data and our best few-shot). There is no best checkpoint here, only a best checkpoint for a stated task and label budget.

ObjectiveFEQ/AEQ (%) ↑ fullFEQ/AEQ (%) ↑ fewPGA GCN ↓ fullPGA GCN ↓ fewPGA linear ↓ fullPGA linear ↓ fewP-prec (%) ↑ fullP-prec (%) ↑ fewS-prec (%) ↑ fullS-prec (%) ↑ few
Task-specific baseline99.282.40.200.331——96.586.893.887.7
Rec87.250.60.3440.3280.3380.32693.018.893.566.1
Rec+Mag77.863.80.1980.2990.2130.28889.545.392.574.4
Rec+Disc+Mag93.985.40.2300.3060.1970.28593.968.893.982.9
Rec+Disc+Mag+Mask87.282.90.1610.2680.1620.26395.8<u>83.5</u>94.288.1
RVQ93.975.60.2480.3080.1990.28094.970.193.981.3
RVQ-EMA95.670.30.2430.3230.1930.29395.172.994.083.7
RVQ-EMA+MLP<u>96.7</u>85.50.3060.3300.1970.30595.574.193.984.1
RVQ-EMA+MaskRaw93.675.90.2210.3150.1740.281<u>96.1</u>81.094.088.0
RVQ-EMA+MaskCode96.381.20.2690.3260.1870.29695.879.394.182.5

Table 1. Frozen-encoder downstream evaluation, full-data and few-shot (≈100 labelled examples). FEQ/AEQ: foreshock/aftershock accuracy; PGA: peak-ground-acceleration MSE, reported for both ground-motion heads we train (a GCN over a station-distance graph and an independent-station linear head); P-/S-prec: picking precision at 0.5 s tolerance. Baselines are the external task-specific models; the ground-motion baseline is itself a GCN, so it is matched by our GCN column and has no linear-head counterpart in the literature (dashes). Bold marks the best value in a column, <u>underline</u> our best checkpoint; no ordering survives across tasks.

As the controlled test, we add to our strongest classifier (RVQ-EMA+MLP) the masking ingredient that helps within the reconstruction family, on the hypothesis that more self-supervised signal is generically useful. It does not produce a generalist: +MaskRaw gains 6.9 points of few-shot P-precision (74.1 → 81.0) and our best full-data P-precision (96.1), while losing 9.6 points of few-shot classification (85.5 → 75.9). +MaskCode, whose masking touches only a side branch discarded at inference, trades less both ways, as expected if a target shapes the representation to the extent it applies pressure upstream of the encoder. Adding self-supervised signal therefore moves specialisation rather than dissolving it: specialisation is a property of what the target asks the model to represent, not of how much signal it receives.

Three controls, scored on classification, rule out the readings that do not implicate the target. All three are our own runs from a larger set of pretraining variations not reported here. Capacity: width-scaling the encoder ~11× (encoder+decoder 4.96 → 56.8M) recovers only part of a corpus-induced deficit, so the ceiling is not a small-model artefact. Architecture: bidirectional Mamba mixers, a better inductive bias in principle here, change nothing measurable at roughly double the mixer cost. Domain proximity: adding 1.2M domain-matched local traces leaves both metrics tied, so the limit is not a wrong-distribution corpus. Throughout, the full-data split is insensitive to almost everything we vary; few-shot is both the discriminating and the fragile metric.

Analysis and open problems

Capacity, architecture, and domain proximity are ruled out above, leaving the target. Two properties of it are in play: we demonstrate the first and motivate the second, and mark the difference. We then discuss four problems, all about the target rather than the architecture.

Demonstrated: a point target on a stochastic process is misspecified. In a further experiment (Appendix § 6.6) we asked the model to forecast 12 s of ground motion under an L1/L2L_1/L_2 loss. It output a flat line, 5–10× below the true spectrum where the energy is. This is the optimum, not a failure to reach it: squared error is minimised by the conditional mean E[Y∣X]\mathbb{E}[Y \mid X], and 12 s ahead the phase of the horizon waveform is nearly independent of the context, so that mean is ≈0\approx0. A perfectly trained model is therefore flat, and the loss cannot separate it from one that has learned the physics. The repair is to target something predictable: multi-scale RMS-envelope agreement, which is what early warning needs anyway. A point target silently asks for a conditional mean, so it teaches only where that mean is predictable [31]. Downstream the ordering then inverts: the collapsed arm yields features ∼4.5\sim4.5 points better than the envelope arm whose forecasts look right. What a target buys is what it forces the encoder to retain, not the sample it emits.

Motivated but untested: the single-window target of § 1. A target computed from a single window can only describe what is inside it. The occurrence dynamics are not: the inter-event interval, the triggering relation to a mainshock months earlier, and the clustering statistics operational forecasting relies on [24] are all discarded before the loss is evaluated. This is the gap between modelling one sample path and modelling the law that generates it, and where our work and the neural-SDE literature [16] diverge. We cannot yet show this is the binding constraint: every objective we varied and every task we scored is window-level, so the redistribution is equally consistent with ordinary task conflict at fixed capacity. Separating the two needs an occurrence-level target (discussed below in Problem 1). Figure A1 illustrates the three structural differences behind this paragraph and the problems below.

Comparison of audio and time-series foundation models with seismology

Figure A1. The three structural differences between the signals audio and time-series FMs are built for and the sequences seismology cares about. (1) Audio and forecasting corpora sit on a fixed sampling clock, whereas earthquake occurrence is event-driven with inter-event gaps from seconds to months, the reason a window-level target cannot reference the occurrence process. (2) Audio carries roughly stationary energy throughout; seismic traces are a noise floor punctuated by rare, high-energy transients: the asymmetry that makes an L1/L2L_1/L_2 objective mis-weighted for a heavy-tailed law. (3) Audio codecs can design around a known audible band and discard the rest; seismology has no universally safe band, since picking, magnitude, and engineering applications each depend on a different part of the spectrum, so any compression target pre-selects which downstream tasks survive it.

(P1) Two-timescale learning targets. An objective could couple a waveform-level encoder to a marked temporal point process over the catalogue, so that one target sees both the within-trace dynamics and the between-event occurrence dynamics. The neural point-process literature [8, 21, 37] already models occurrence. What is missing is a formulation in which the waveform representation and the event history condition each other. Such a target would let waveform content sharpen occurrence forecasts, including the prospective version of the foreshock question we can currently only ask after the fact.

(P2) Tail-aware generative targets. Diffusion and score-based models [14, 31] are already viable on seismic-adjacent signals [7], but have been applied to signal processing rather than event-centric inference. Given the conditional-mean collapse above, the question is not whether the target should be distributional: it is how to weight the objective so that rare events contribute gradient signal in proportion to how much they matter, not their frequency.

(P3) Evaluation with a stated cost structure. No rule exists for weighing a 9.6-point classification loss against a 6.9-point picking gain, so which checkpoint to deploy is underspecified within this study, let alone across methods. A benchmark for learned stochastic simulators needs proper scoring rules [10], calibration targets, and an explicit cost structure for cross-task trade-offs. With those, a result like ours reads as a Pareto frontier rather than a null finding. Seismic FMs are new enough that this benchmark can still be designed rather than inherited.

(P4) Targets that use the known physics. Wave propagation has a known governing operator, so a target need not rediscover it: defining the objective on the residual to a forward model, or in a neural-operator basis, would spend capacity only on the stochastic part, that is, what the physics leaves undetermined. Both components already exist (neural operators [25, 26] and physics-residual training [30]) as ways to solve the physics, not as pretraining targets.

Conclusion

We asked whether seismic FMs are more than feature extractors. To answer, we developed nine checkpoints sharing one backbone and one corpus, and we evaluated them with a frozen encoder on three tasks under distribution shift. These are competent pretrained feature extractors for seismology rather than FMs in the sense imported from language and vision. Varying the objective redistributes downstream skill rather than raising it, and no checkpoint generalises without giving something up; pretraining helps where labels are scarce (few-shot training). Controls on capacity, architecture and domain proximity do not explain why skill moves rather than rises, so what remains is the target. We demonstrate one of its failures, the collapse of a point target to a near-zero conditional mean, and can so far only motivate the other, that a single-window target cannot reference the occurrence process. What stochastic theory contributes is therefore not an architecture but a criterion for when a learning target is well posed.

Full experimental detail

Backbone

A SEANet-style encoder–decoder: an SConv1d + ResnetBlock front-end followed by three strided downsampling stages (strides 3, 4, 5), each paired with a block of Mamba selective state-space layers [11], and a symmetric decoder. Mamba is chosen over attention for linear-time scaling at these sequence lengths (3×30,0003 \times30{,}000 samples per window, reduced 60×60\times along time to ∼500\sim500 latent frames). Where present, the RVQ bottleneck is 8×1024×1288 \times1024 \times128; the +MLP setting enables a gated-MLP branch with dintermediate=2dmodeld_{\mathrm{intermediate}} = 2d_{\mathrm{model}}. Downstream heads always consume the continuous pre-quantisation representation, never the discrete codes.

Objectives

Rec combines L1/L2L_1/L_2 losses on the raw waveform (weight 5 each) and on the one-sided amplitude spectrum (1/31/3, 1/1001/100). Disc is a hinge adversarial loss (weight 1/31/3) with an L1L_1 feature-matching term over intermediate discriminator features, the latter carrying a large effective weight (100/3100/3) that matters below. Mag is an auxiliary magnitude-regression head (MSE on the predicted mean plus a small penalty on the log-variance). Mask applies span masking to the raw input (ratio 0.4, span 40 samples), with the waveform term restricted to masked positions; +MaskCode instead masks a separate copy of the encoder output (ratio 0.5, span 10 latent frames) consumed only by an auxiliary cross-entropy head predicting the first three RVQ levels (0.83M parameters, discarded at inference).

Corpora

CREW [1]: ≈1.6\approx1.6M five-minute, three-component regional traces (11–20∘20^\circ). Ambient noise: 1,499,918 training windows (≈125,000\approx125{,}000 hours) from 462 FDSN networks, plus 169,832 validation and 162,846 test windows. Waveforms are peak-normalised jointly across the three components, preserving inter-channel amplitude ratios.

Downstream protocol

Foreshock/aftershock [17]: two months pre- and post-mainshock windows (25 s) at NRCA station, 70/10/20 split full-data and 2% few-shot. Ground motion [3, 15]: 10 s post-origin waveforms, 915 events ×\times 39 INGV stations, predicting PGA, PGV, and PSA0.3/1.0/3.0_{0.3/1.0/3.0} through both a 2-layer GCN head over a station-distance graph and an independent-station linear head; 65% full-data, 8% few-shot. Phase detection [36]: per-sample 3-way (noise/P/S) labelling on the SeisBench STEAD partition, 143,523 training traces, a prediction counted correct within 50 samples (0.5 s) of the labelled arrival; 0.1% few-shot. All numbers in Table 1 use a frozen encoder.

Scale, architecture and corpus controls

The three controls quoted in § 3, drawn from a larger set of pretraining variations, are reported here only in brief.

Width scaling raises the base channel width 16 →\to 56 (encoder+decoder 4.96 →\to 56.8M; with the RVQ codebooks, 6.0 →\to 57.9M, the convention used in § 3), holding depth, downsampling ratios, window length, batch size and learning rate fixed; this architecture scales by width, not depth, since per-stage Mamba depth 56 makes the backward pass effectively hang. Its few-shot figures are quoted against a same-corpus 6M control rather than against Table 1, because the scaling runs use a pooled CREW + INSTANCE [22] + STEAD corpus and a shorter 6000-sample window. Bidirectional Mamba replaces every mixer with a forward-plus-reversed scan fused by a linear projection, at ≈2×\approx2\times mixer cost, on the same CREW-plus-noise corpus as Table 1. Corpus rebalancing compares two mixtures at identical architecture, both drawn from STEAD, INSTANCE, SCEDC and ETHZ with CREW capped at ≈1/10\approx1/10: a proportional mixture, and a domain-first mixture adding 1.2M further locally recorded traces.

The forecasting experiment

Separately from the nine checkpoints in Table 1, this experiment uses a domain-first mixture of ≈3.35\approx3.35M traces (STEAD, INSTANCE, SCEDC and ETHZ, with CREW capped at ≈1/10\approx1/10 of the total) cropped to a common 6000-sample window because the shortest constituent corpus is 60 s.

Each window splits into a 4800-sample context and a genuinely held-out 1200-sample horizon. The horizon is never an encoder input, which distinguishes this from the masking variants. A latent forecast head (20 learnable future queries rolled forward by a 2-layer causal Mamba) predicts a future latent feeding both the shared decoder and a next-token head over the first three RVQ levels. Arm 1 applies L1/L2L_1/L_2 plus a spectra term to the predicted horizon waveform; arm 2 replaces the two time-domain terms with multi-scale RMS-envelope agreement over W={25,50,100,200}W = \{25, 50, 100, 200\} samples (0.25–2 s), keeping the already phase-invariant spectra term. Both are run against a matched control (identical model, data, seed and windowing, forecast terms removed) over 3 seeds. Arm 1 reaches 86.1±1.086.1 \pm1.0 few-shot classification against arm 2’s 81.6±0.981.6 \pm0.9 (the ∼4.5\sim4.5-point gap quoted in § 4) while producing the collapsed forecast described there. As with the scaling control, these figures are comparable to each other and to their own control, not to Table 1: the matched control reaches 78.9 few-shot against Table 1’s 85.5, and that entire gap is the data recipe rather than the objective.

References

  1. [1]Albert Leonardo Aguilar Suarez and Gregory C. Beroza. Curated regional earthquake waveforms (crew) dataset. Seismica, 3(1), 2024. doi: 10.26443/seismica.v3i1.1049.DOI
  2. [2]Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 12449–12460, 2020.arxiv.org/abs/2006.11477
  3. [3]Stefan Bloemheuvel, Jurgen van den Hoogen, Dario Jozinović, Alberto Michelini, and Martin Aztmuller. Graph neural networks for multivariate time series regression with application to seismic data. International Journal of Data Science and Analytics, pages 1–16, 2022. doi: 10.1007/s41060-022-00349-6.DOI
  4. [4]Sylvette Bonnefoy-Claudet, Fabrice Cotton, and Pierre-Yves Bard. The nature of noise wavefield and its applications for site effects studies: A literature review. Earth-Science Reviews, 79(3–4): 205–227, 2006. doi: 10.1016/j.earscirev.2006.07.004.
  5. [5]David M. Boore. Simulation of ground motion using the stochastic method. Pure and Applied Geophysics, 160(3):635–676, 2003. doi: 10.1007/PL00012553.DOI
  6. [6]Yuqi Cai, Ziye Yu, Weitao Wang, Xin Liu, Yunpeng Zhang, Yanru An, Wei Yang, Lu Li, Zhengya Si, Yueyang Xia, and Xiaona Ma. Seismicmx: A cross-task foundation model for single-station seismic waveform processing. Seismological Research Letters, 97(4):2550–2573, 2026. doi: 10.1785/0220250290.DOI
  7. [7]Shijun Cheng, Randy Harsuko, and Tariq Alkhalifah. A generative foundation model for an all-in-one seismic processing framework. Surveys in Geophysics, 46:1173–1215, 2025. doi: 10.1007/s10712-025-09912-9.DOI
  8. [8]Kेलian Dascher-Cousineau, Oleksandr Shchur, Emily E. Brodsky, and Stephan Günnemann. Using deep learning for flexible and scalable earthquake forecasting. Geophysical Research Letters, 50(17):e2023GL103909, 2023.
  9. [9]Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT), 2019.arxiv.org/abs/1810.04805
  10. [10]Tilmann Gneiting and Adrian E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477):359–378, 2007. doi: 10.1198/016214506000001437.DOI
  11. [11]Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023.arxiv.org/abs/2312.00752
  12. [12]B. Gutenberg and C. F. Richter. Frequency of earthquakes in california. Bulletin of the Seismological Society of America, 34(4):185–188, 1944. doi: 10.1785/BSSA0340040185.DOI
  13. [13]Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.arxiv.org/abs/2111.06377
  14. [14]Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, 2020.arxiv.org/abs/2006.11239
  15. [15]Dario Jozinović, Anthony Lomax, Ivan Štajduhar, and Alberto Michelini. Rapid prediction of earthquake ground shaking intensity using raw waveform data and a convolutional neural network. Geophysical Journal International, 222(2):1379–1389, 2020. doi: 10.1093/gji/ggaa233.DOI
  16. [16]Patrick Kidger, James Foster, Xuechen Li, Harald Oberhauser, and Terry Lyons. Neural SDEs as infinite-dimensional GANs. In International Conference on Machine Learning (ICML), 2021.arxiv.org/abs/2102.03657
  17. [17]Laura Laurenti, Gabriele Paoletti, Elisa Tinti, Fabio Galasso, Cristiano Collettini, and Chris Marone. Probing the evolution of fault properties during the seismic cycle with deep learning. Nature Communications, 15:10025, 2024. doi: 10.1038/s41467-024-54153-w.DOI
  18. [18]Laura Laurenti, Christopher W. Johnson, Daniele Trappolini, Elisa Tinti, Fabio Galasso, and Chris Marone. Testing audio compression autoencoders for seismology: Moving toward foundation models. Journal of Geophysical Research: Machine Learning and Computation, 2026. doi: 10.1029/2025JH000787.DOI
  19. [19]Sen Li, Xu Yang, Anye Cao, Changbin Wang, Yaoqi Liu, Yapeng Liu, and Qiang Niu. Seist: A foundational deep-learning model for earthquake monitoring tasks. IEEE Transactions on Geoscience and Remote Sensing, 62:1–15, 2024. doi: 10.1109/TGRS.2024.3371503.DOI
  20. [20]Tianlin Liu, Jannes Münchmeyer, Laura Laurenti, Chris Marone, Maarten V. de Hoop, and Ivan Dokmanić. SeisIm: a foundation model for seismic waveforms. arXiv preprint arXiv:2410.15765, 2024.arxiv.org/abs/2410.15765
  21. [21]Hongyuan Mei and Jason M. Eisner. The neural hawkes process: A neurally self-modulating multivariate point process. In Advances in Neural Information Processing Systems (NeurIPS), 2017.
  22. [22]Alberto Michelini, Spina Cianetti, Sonja Gaviano, Carlo Giunchi, Dario Jozinović, and Valentino Lauciani. Instance – the italian seismic dataset for machine learning. Earth System Science Data, 13:5509–5544, 2021. doi: 10.5194/essd-13-5509-2021.DOI
  23. [23]S. Mostafa Mousavi, Yixiao Sheng, Weiqiang Zhu, and Gregory C. Beroza. Stanford earthquake dataset (stead): A global data set of seismic signals for ai. IEEE Access, 7:179464–179476, 2019. doi: 10.1109/ACCESS.2019.2947848.DOI
  24. [24]Yoshihiko Ogata. Statistical models for earthquake occurrences and residual analysis for point processes. Journal of the American Statistical Association, 83(401):9–27, 1988.DOI
  25. [25]Md Ashiqur Rahman, Manuel A. Florez, Anima Anandkumar, Zachary E. Ross, and Kamyar Azizzadenesheli. Generative adversarial neural operators. arXiv preprint arXiv:2205.03017, 2022.arxiv.org/abs/2205.03017
  26. [26]Md Ashiqur Rahman, Zachary E. Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neural operators. arXiv preprint arXiv:2204.11127, 2022.arxiv.org/abs/2204.11127
  27. [27]Bertrand Rouet-Leduc, Claudia Hulbert, Nicholas Lubbers, Kipton Barros, Colin Humphreys, and Paul A. Johnson. Machine learning predicts laboratory earthquakes. Geophysical Research Letters, 44(18), 2017. doi: 10.1002/2017GL074677.
  28. [28]Omar M. Saad, Yangkang Chen, and Tariq Alkh alifah. U-trans: A foundation model for seismic waveform representation and enhanced downstream earthquake tasks. Scientific Reports, 16:12657, 2026. doi: 10.1038/s41598-026-41454-x.DOI
  29. [29]Xu Si, Xinming Wu, Hanlin Sheng, Jun Zhu, and Zefeng Li. Seisclip: A seismology foundation model pre-trained by multimodal data for multipurpose seismic feature extraction. IEEE Transactions on Geoscience and Remote Sensing, 62:1–13, 2024. doi: 10.1109/TGRS.2024.3354456.DOI
  30. [30]Jonathan D. Smith, Kamyar Azizzadenesheli, and Zachary E. Ross. Eikonet: Solving the eikonal equation with deep neural networks. IEEE Transactions on Geoscience and Remote Sensing, 2021.DOI
  31. [31]Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021.
  32. [32]Marco Tagliasacchi, Yunpeng Li, Karolis Misiunas, and Dominik Roblek. Seanet: A multi-modal speech enhancement network. In Interspeech, 2020.DOI
  33. [33]Tokūji Utsu, Yoshihiko Ogata, and Ritsuko S. Matsu’ura. The centenary of the omori formula for a decay law of aftershock activity. Journal of Physics of the Earth, 43(1):1–33, 1995. doi: 10.4294/jpe1952.43.1.DOI
  34. [34]Xinghao Wang, Feng Liu, Rui Su, Zhihui Wang, Lihua Fang, Lianqing Zhou, Lei Bai, and Wanli Ouyang. SeismoLLM: Advancing seismic monitoring via cross-modal transfer with pretrained large language model. Geophysical Research Letters, 53:e2025GL118505, 2026. doi: 10.1029/2025GL118505.DOI
  35. [35]Jack Woollam, Jannes Münchmeyer, Frederik Tilmann, Andreas Rietbrock, Dietrich Lange, Thomas Bornstein, Tobias Diehl, Carlo Giunchi, Florian Haslinger, Dario Jozinović, Alberto Michelini, Joachim Saul, and Hugo Soto. Seisbench – a toolbox for machine learning in seismology. Seismological Research Letters, 93(3):1695–1709, 2022. doi: 10.1785/0220210324.DOI
  36. [36]Weiqiang Zhu and Gregory C. Beroza. Phasenet: A deep-neural-network-based seismic arrival-time picking method. Geophysical Journal International, 216(1):261–273, 2019. doi: 10.1093/gji/ggy423.DOI
  37. [37]Simiao Zuo, Haoming Jiang, Zichong Li, Tuo Zhao, and Hongyuan Zha. Transformer hawkes process. In International Conference on Machine Learning (ICML), 2020.arxiv.org/abs/2002.09291

Paper details

vlnc:W_37qjptzy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents