acceptodds
Under review as a conference paper at ICLR 2027

SeisTailFlow: Tail-Aware Conditional Flow Modeling for Reliable Seismic Phase Picking

Abstract

Seismic phase picking is fundamental to automated earthquake monitoring, yet existing deep learning pickers still exhibit heavy-tailed error distributions, where a small fraction of large arrival-time errors can compromise the reliability of downstream phase association and event localization. We propose SeisTailFlow, a tail-aware conditional flow model for reliable phase picking. Rather than treating large prediction errors merely as residual uncertainty, SeisTailFlow uses out-of-sample residuals to identify waveform conditions associated with elevated error risk and explicitly models their conditional residual structure alongside dominant arrival-time patterns. Technically, SeisTailFlow builds parallel Main and Tail Flows and integrates residual-based tail supervision, background-referenced spectral enhancement, extreme-value regularization, and adaptive vector-field routing to jointly capture dominant conditional structure and high-error-related residual behavior. This design preserves accurate modeling of dominant arrival-time patterns while providing additional capacity for scarce high-error regimes and large arrival-time deviations. Experiments on synthetic data and the STEAD dataset show that SeisTailFlow improves overall picking accuracy and probabilistic prediction quality while substantially reducing tail errors and large-error events, with particularly robust performance under low-SNR conditions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Seismic Phase Picking, Conditional Flow Matching, Tail Modeling, Generalized Pareto Distribution, Spectral Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QBUiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the reliability of seismic phase picking, focusing on the small fraction of large arrival-time errors. It proposes SeisTailFlow, a tail-aware conditional flow model that uses cross-fitted, out-of-sample residuals to identify high-error training cases. The framework combines shared waveform encoding, background-referenced spectral enhancement, Main and Tail conditional flows, GPD exceedance modeling, and adaptive vector-field fusion. Arrival time is represented as a waveform-dependent location plus a stochastic residual. The fused flow generates arrival-time samples, whose mean provides the point prediction and whose distribution provides predictive intervals. The paper reports improvements on synthetic waveforms and STEAD in overall accuracy, tail errors, low-SNR performance, and probabilistic prediction quality, and includes ablations and stratified analyses.

Critical strengths
  1. The paper addresses large arrival-time errors that may affect downstream phase association and localization. The use of tail-risk and probabilistic metrics provides a more comprehensive evaluation than MAE alone.

  2. Event-level data splitting, training-only normalization, and out-of-fold tail-label construction help reduce data leakage. The controlled synthetic experiments further support the analysis of different factors affecting performance.

Critical weaknesses
  1. Tail metrics raise consistency concerns. The P99, ES@99, and [email protected] values in Table 1 appear difficult to reconcile with Eqs. (53)-(55); the full model's Tail MAE also differs between Table 1 and Tables 6/7. How are these discrepancies explained, and are the same evaluation thresholds used across these tables?

  2. Tail Flow training is insufficiently specified. How are the final model's target states z_1^main and z_1^tail constructed, including any normalization and gradient detachment? How are high-error samples used in Tail Flow matching?

  3. Unclear ablation configurations limit interpretation of the results. How do "w/o Tail Flow" and "Main Flow only" differ in retained modules, auxiliary losses, and generation steps?

Questions for authors
  1. Please provide per-run P99, ES@99, and [email protected] for Table 1, together with the quantile conventions, units, percentage conversion, and aggregation procedure. Are all three computed from the same absolute errors within each run?

  2. What numerical evaluation threshold and selection rule are used for Tail MAE in Table 1 and Tables 6/7? Do these tables use the same threshold, underlying test predictions, and aggregation procedure? Please reconcile the full model's values of 0.0268 s and 0.0152 s.

  3. For the final dual-flow model, what are the exact formulas for the Main and Tail target states, including any normalization? Does Tail Flow matching select or reweight high-error cases? Is the location prediction detached when constructing these targets?

  4. Which modules, auxiliary losses, and generation steps are retained in 'w/o Tail Flow' versus 'Main Flow only'?

5 comments
AuthorsOpenReview ↗

We sincerely appreiciate the valuable comments from the reviewer.

W1 seeQ1 & Q2.

W2 see Q3.

W3 see Q4.

Q1 To ensure a fair comparison between SeisTailFlow and the baselines under the same computational time budget, Table 1 uses the SeisTailFlow configuration with a fixed fusion weight of α=0.1\alpha=0.1. In the verified run, P99 and ES@99 are both computed from the same set of absolute arrival-time errors measured in seconds: P99 is the empirical 99th percentile calculated using linear interpolation, and ES@99 is the mean absolute error exceeding that percentile. [email protected] is defined as the proportion of samples with errors exceeding 0.3 s, multiplied by 100. Our experimental results are based on multi-seed evaluations: each metric is first calculated within each run, and its mean and standard deviation are then aggregated across runs, rather than being calculated directly in a single step. Our reported tail metrics are consistent with Eqs. (53)–(55). The ablation experiments use adaptive routing and therefore correspond to a different model configuration from Table 1.

Q2 Tail MAE is the mean absolute error above a predefined evaluation threshold uevalu_{\mathrm{eval}}. This metric uses the training-set OOF tail-screening threshold:

\mathrm{MAE}_{S}=\frac{1}{|S|}\sum_{i\in S}e_{i}^{\mathrm{adaptive}}.$$ The adaptive model can predict samples identified as difficult by Seis Main more accurately; therefore, its mean error can be 0.0152 s, below the screening threshold. The same quantile threshold is used across all models. The ablation results were obtained by retraining the models. Due to randomness in the training process, training durations varied slightly; training in the ablation experiments was longer and achieved somewhat better convergence. **Q3** We do not separately normalize the Main and Tail Flow outputs before combining them. Instead, we adaptively fuse their vector fields and normalize the fused result. This is also one of our innovations, and this treatment can improve predictive accuracy. $$v_{\mathrm{fused}}(z_{t},t\mid\mathbf{X})=\rho_{t}\\,v_{\mathrm{main}}(z_{t},t\mid\mathbf{h}_{\mathrm{main}})+(1-\rho_{t})\\,v_{\mathrm{tail}}(z_{t},t\mid\mathbf{h}_{\mathrm{tail}}).$$ Both the Tail and Main Flows are trained on all samples, with additional weights assigned to tail and non-tail samples to strengthen their respective training, as follows: $$L_{\mathrm{FM}}=\underbrace{\frac{1}{|B|}\sum_{i\in B}\ell_{i}^{\mathrm{fused}}}_{\text{All samples}}+\underbrace{\frac{\beta}{\max(|B_{0}|,1)}\sum_{i\in B_{0}}\ell_{i}^{\mathrm{main}}}_{\text{Non-tail samples}}+\underbrace{\frac{1-\beta}{\max(|B_{1}|,1)}\sum_{i\in B_{1}}\ell_{i}^{\mathrm{tail}}}_{\text{Tail samples}}.$$ As shown by the formula, the quantile properties of the Tail and Main Flows are preserved. The location prediction is not detached when constructing the target states; therefore, the prediction targets remain connected to the location prediction. **Q4** We apologize for the ambiguity caused by our naming and will provide additional clarification in the manuscript. **Main Flow only** is a single-flow model that retains only the shared waveform encoder, Main branch, and location prediction head. It uses only time-domain conditioning representations and is trained with the Main FM loss and location NLL. During generation, only the Main vector field is integrated. **w/o Tail Flow** is a single-flow ablation model that retains the spectral-enhancement module: spectral features are fused with the Main conditioning representation and supplied to a single Main vector field, while the Tail vector field, routing module, and associated BCE and GPD training terms are removed. Training retains the Main FM loss and location NLL, and FM gradients can update the spectral encoder through the enhanced conditioning. During generation, only the Main vector field is integrated, but it uses both time-domain and spectral information. **We have thoroughly revised the manuscript in response to the reviewer's insightful comments and uploaded the updated version.** We have added a table note to Table 8 (P32) and highlighted it in color. **We sincerely appreciate the reviewer for your great efforts, valuable time and insightful comments for our paper.**
AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

We sincerely appreciate your insightful comments. We have carefully addressed your concerns and submitted our rebuttal. We would be grateful if you could take a moment to review our responses and share any further comments. Thank you so much for your valuable time and consideration.

AuthorsOpenReview ↗

Second Follow-up on Our Rebuttal

I am following up once more on the rebuttal we submitted. We greatly appreciate the reviewer. Please let us know if any further information would be helpful. Thank you very much for your kind attention.

Reviewer_fWoLOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SeisTailFlow presents a framework for predicting P-wave arrival times in the context of seismic phase picking. For this, the authors use conditional flow models. Given a seismic waveform input, the input is transformed into a shared waveform representation. Afterwards, there are two modules in SeisTailFlow: Main Flow and Tail Flow. Before both networks are trained jointly, an auxiliary predictor (SeisMain) uses K-fold cross-fitting to identify high errors. Main Flow learns the distribution of P-wave arrival-time residuals using conditional flow matching. Tail Flow learns the complementary tail residuals associated with high errors. Then, through adaptive routing, the vector field outputs of the two flows are combined. This is then applied to two datasets: 1. Synthetically generated data 2. Stanford Earthquake Dataset In the presented cases, SeisTailFlow consistently outperforms the selected baselines.

Critical strengths
  • Includes two datasets, synthetic and real-world. The low-SNR case is impressive. Well selected training/test splits are used for an extensive analysis of results. The improvement is demonstrated in several different cases and compared to a variety of baselines. The framework results in a probabilistic prediction - the uncertainity could be used in additional downstream modeling.
  • Tail distribution modeling is difficult and very relevant to many extreme-event domains. The approach of using out-of-fold predictions to identify high-error samples is clean and non-invasive.
Critical weaknesses
  • The framework is very dependent on SeisMain and its prediction capability as the high-error samples used for training are identified based on its predictions. However, no results from SeisMain are reported. This makes it difficult to assess how much additional improvement SeisTailFlow provides compared to the auxiliary predictor.
  • We have both Main Flow and Tail Flow in addition to Seis Main. Main Flow alone already achieves very good results, and while Tail Flow improves the tail errors, the improvement in overall accuracy is relatively small, e.g. MainFlow MAE on STEAD ~0.404, Full SeisTailFlow MAE ~0.399. This adds a lot of complexity, but it is not clear whether the improvements justify the additional computational cost.
Questions for authors
  • How does SeisTailFlow compare directly to Seis Main?
  • What is the (additonal) computational cost of the proposed framework?
4 comments
AuthorsOpenReview ↗

We sincerely appreiciate the valuable comments from the reviewer.

W1 Seis Main is not specifically designed for tail modeling, but its prediction errors can be used to identify tail samples. We conducted numerical experiments to compare it directly with SeisTailFlow; see Q1.

W2 We emphasize that the core value of SeisTailFlow lies in improving predictions for a small number of extremely important samples, as these rare samples are often critical to earthquake occurrence. Main Flow can predict most routine samples but cannot handle these rare, high-error cases at all. Because these samples constitute only a small fraction of the dataset, their improvements can be diluted by the large number of routine samples when calculating overall MAE. Therefore, a modest reduction in overall MAE does not fully reflect the value of reducing tail errors. We particularly focus on high-quantile metrics such as P99, which more directly measure performance at the high-error end. SeisTailFlow improves both overall accuracy and tail metrics while specifically addressing rare, important samples. Details of the computational costs are provided in Q2.

Q1 Seis Main generates out-of-sample residuals through five-fold cross-fitting within the training set. Seis TIS uses the 95th percentile of these absolute residuals to determine and freeze the tail/non-tail labels, providing high-error risk supervision for SeisTailFlow. All samples participate in fused vector-field matching, with additional Tail-FM supervision for tail samples and Main-FM supervision for non-tail samples.

To compare their predictive performance directly, we evaluate Seis Main as a standalone arrival-time predictor against SeisTailFlow on the same test set, using the same training/test split and evaluation protocol:

Metric ↓Seis MainSeisTailFlowRelative reduction
MAE (s)0.01770.006961.0%
P95 (s)0.11860.016086.5%
P99 (s)0.12610.019784.4%
ES@99 (s)0.12970.022882.4%

These results show that SeisTailFlow improves overall predictive accuracy over the auxiliary predictor, with larger gains in high-error metrics, demonstrating the full framework's ability to reduce tail errors.

Q2 All computational costs were compared on the same server equipped with an NVIDIA GeForce RTX 5080 GPU. Parameter counts include all parameters in the final prediction model and are reported in millions; final-model training time is measured as the actual wall-clock duration of the corresponding training stage. Inference uniformly uses FP32, a batch size of 16, and 64 trajectories per record with 16 Euler integration steps per trajectory. Average inference time per record is obtained by dividing the total inference time by the number of processed records. SeisTailFlow controls the additional overhead by reusing the shared waveform encoder and conditioning representations: the encoder is evaluated once per input, while the two conditional vector fields are evaluated separately and fused at each integration step, without repeatedly running two complete waveform encoding networks. The computational costs are as follows:

MetricMain Flow onlySeisTailFlow
Parameters0.391M0.448M
Final-model training time1.182h1.430h
Inference time10.2ms13.6ms

The additional computational cost is relatively small, and the modules we introduce, such as spectral learning, are lightweight. From the perspective of model complexity, while maintaining strong overall performance, we target a small number of extremely important samples. These samples correspond to moments when earthquakes actually occur and early warning is needed, making them particularly important.

We sincerely appreciate the reviewer for your great efforts, valuable comments and kind help for our paper.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

We sincerely appreciate your insightful comments. We have carefully addressed your concerns and submitted our rebuttal. We would be grateful if you could take a moment to review our responses and share any further comments. Thank you so much for your valuable time and consideration.

Thank you very much for taking the time in addressing my concerns. I will increase my score!

AuthorsOpenReview ↗

Thank you so much for your support.

Thank you very much for your positive feedback and willingness to increase the score. We sincerely appreciate your time and consideration. We would also be grateful if you could update the score when convenient. Thank you so much for your support!

Reviewer_CcgROpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SeisTailFlow addresses the problem of heavy-tailed arrival-time errors in seismic phase picking. It proposes a conditional flow-matching framework with parallel Main and Tail pathways. The Tail Flow uses out-of-fold residuals to construct targeted tail supervision, background noise-referenced spectral features to support low-SNR cases, and a generalized Pareto distribution (GPD) objective to alleviate the scarcity of high-error observations. Experiments on controlled synthetic waveforms and the STEAD dataset report improvements in overall accuracy, tail errors, low-SNR robustness, and probabilistic prediction quality. The topic is practically relevant, but some evaluation details need further validation.

Critical strengths

(1) The proposed tail-aware dual-flow framework integrates residual-based supervision, spectral enhancement, and adaptive routing to improve phase-picking reliability. (2) The experimental results demonstrate promising improvements in overall accuracy and tail-error reduction.

Critical weaknesses

(1) The experiments are primarily conducted on synthetic data and the STEAD dataset, which may not adequately capture the diversity of seismic station responses, regional geological conditions, instrumentation, and data distributions. Further validation using additional seismic datasets (e.g., the DiTing dataset), continuous seismic waveform records, and seismic data from different regions is needed to demonstrate the generalizability of the proposed method. (2) Although the paper presents extensive experimental results, some details require further clarification. For example, the authors should specify how the Tail MAE evaluation threshold is selected and whether a common threshold is applied across all methods. Additionally, the discrepancy in Tail MAE values between Tables 1 and 6 should be explained.

Questions for authors
  1. How exactly are training samples assigned to the Main Flow and Tail Flow matching objectives? Is the Tail Flow trained only on samples selected by Seis TIS, on all samples with reweighting? Please give the exact objective and sampling strategy.
  2. The proposed framework introduces two flow branches for prediction. Could the authors provide a comparison of computational cost with existing methods to better demonstrate its practical applicability?
  3. Can the method be evaluated on an additional independent real seismic dataset or under a cross-region/cross-dataset shift?
  4. How are the GPD shape and scale parameters constrained to ensure valid and numerically stable extreme-error modeling during training?
  5. The authors point out that large arrival-time errors can compromise the reliability of downstream event localization. Could the authors conduct multi-station earthquake location experiments to assess whether the proposed method improves location accuracy?
7 comments
AuthorsOpenReview ↗

This is part one of our reply (Responses to W1, W2, Q1, Q2). We sincerely appreiciate the valuable comments from the reviewer.

W1 To demonstrate the generalizability of the proposed method, we evaluated it under a cross-dataset shift (see Q3).

W2 Tail MAE is the mean absolute error above a predefined evaluation threshold uevalu_{\mathrm{eval}}. This metric uses the training-set OOF tail-screening threshold:

\mathrm{MAE}_{S}=\frac{1}{|S|}\sum_{i\in S}e_{i}^{\mathrm{adaptive}}.$$ The adaptive model can predict samples identified as difficult by Seis Main more accurately; therefore, its mean error can be 0.0152s, below the screening threshold. The same quantile threshold is used across models. Our ablation results were obtained by retraining the models, with longer training in the ablation experiments allowing better convergence. **Q1** Both the Main and Tail Flows are trained on all samples, as follows: $$L_{\mathrm{FM}}=\underbrace{\frac{1}{|B|}\sum_{i\in B}\ell_{i}^{\mathrm{fused}}}_{\text{All samples}}+\underbrace{\frac{\beta}{\max(|B_{0}|,1)}\sum_{i\in B_{0}}\ell_{i}^{\mathrm{main}}}_{\text{Non-tail samples}}+\underbrace{\frac{1-\beta}{\max(|B_{1}|,1)}\sum_{i\in B_{1}}\ell_{i}^{\mathrm{tail}}}_{\text{Tail samples}}.$$ Here, $\ell_i^s = |v_{s,i} - u_{s,i}|^2$ denotes the velocity-matching error along the path, and $\beta = 0.5$. Seis TIS uses the 95th percentile of the training residuals to determine and freeze the non-tail and tail labels, thereby forming $B_0$ and $B_1$. Additional supervision is applied to non-tail and tail samples, respectively, to strengthen the training of the corresponding flows. For an empty subset, the corresponding loss term is set to zero. Training records are randomly shuffled with a batch size of 16, without weighted sampling, FM oversampling, or a fixed tail proportion. A separate batch of 32 tail records is used exclusively for GPD training. **Q2** All experiments were conducted on the same server equipped with an NVIDIA GeForce RTX 5080 GPU. Training and inference times for all models were measured on the same hardware to ensure a fair comparison of computational costs. The results are shown below: **Model size and computational efficiency** | Method | Parameters (M) | Final-model training time (h) | Inference time (ms/record) | Throughput (records/s) | Timing settings | | :---: | :---: | :---: | :---: | :---: | :--- | | PhaseNet | 0.354 | 0.78 | 2.7 | 370.4 | FP32; batch size 16; single forward pass | | EQTransformer | 0.013 | 0.92 | 3.4 | 294.1 | FP32; batch size 16; single forward pass | | Armstrong | 1.063 | 2.35 | 8.1 | 123.5 | FP32; batch size 16; probabilistic prediction budget | | Vanilla CFM | 0.391 | 1.18 | 10.8 | 92.6 | FP32; batch size 16; 64 trajectories × 16 Euler steps | | CW-Flow | 0.440 | 1.30 | 12.1 | 82.6 | FP32; batch size 16; 64 trajectories × 16 Euler steps | | Main Flow only | 0.391 | 1.182 | 10.2 | 98.0 | FP32; batch size 16; 64 trajectories × 16 Euler steps | | SeisTailFlow | 0.448 | 1.43 | 9.6 |146.5 | FP32; batch size 16; 64 trajectories × 16 Euler steps; encoder evaluated once and both vector fields evaluated at each step | SeisTailFlow has a moderate parameter count and training time, comparable to those of existing models, while achieving higher predictive accuracy. **Full-pipeline computational cost** The pipeline costs below are estimated from historical execution logs and comparable configurations. | Cost component | Wall-clock time (h) | GPU-hours | Estimation basis | | :---: | :---: | :---: | :--- | | Seis TIS / 5-fold OOF construction | 0.92 | 1.55 | Estimated from historical parallel five-fold execution schedules and comparable configurations | | Final SeisTailFlow training | 1.43 | 1.43 | Estimated from historical 30-epoch dual-flow runs at the same scale, including additional overhead for dedicated GPD batches | | Full pipeline (sum of stage costs) | 2.35 | 2.98 | Shared TIS construction counted once; wall-clock time: 0.92 + 1.43; GPU-hours: 1.55 + 1.43 |
AuthorsOpenReview ↗

This is part two of our reply (Respones to Q3, Q4, Q5). Thank you very much for your understanding.

Q3 Access to the DiTing dataset requires an offline user approval process. We will report the results after obtaining the data. We selected another representative dataset, INSTANCE.

INSTANCE is a seismic dataset released by the Italian National Institute of Geophysics and Volcanology (INGV), containing records from Italy and surrounding regions between January 2005 and January 2020. It covers 54,008 earthquakes and approximately 1.16 million three-component event waveforms, each sampled at 100 Hz with a duration of 120 s.

MethodMAE (s)P95 (s)P99 (s)ES@99 (s)[email protected] (%)CRPS (s)
Armstrong21.640±0.18021.640 \pm 0.18052.380±0.74052.380 \pm 0.74057.420±1.35057.420 \pm 1.35060.130±1.82060.130 \pm 1.82090.15±1.2190.15 \pm 1.2120.970±0.22020.970 \pm 0.220
Vanilla CFM25.420±0.37025.420 \pm 0.37056.210±1.14056.210 \pm 1.14061.080±1.85061.080 \pm 1.85065.340±2.21065.340 \pm 2.21094.30±0.8294.30 \pm 0.8224.160±0.35024.160 \pm 0.350
CW-Flow20.380±0.19020.380 \pm 0.19050.280±0.62050.280 \pm 0.62054.160±1.21054.160 \pm 1.21056.720±1.48056.720 \pm 1.48087.10±1.5487.10 \pm 1.5419.020±0.24019.020 \pm 0.240
EQTransformer20.842±0.15120.842 \pm 0.15150.011±0.23250.011 \pm 0.23252.300±0.83552.300 \pm 0.83555.836±1.51955.836 \pm 1.51989.72±1.2989.72 \pm 1.2920.018±0.19620.018 \pm 0.196
PhaseNet20.526±0.21320.526 \pm 0.21351.220±1.51551.220 \pm 1.51558.093±2.82858.093 \pm 2.82861.687±2.98361.687 \pm 2.98386.88±1.6586.88 \pm 1.6519.106±0.31119.106 \pm 0.311
SeisTailFlow18.952±0.157\mathbf{18.952 \pm 0.157}47.974±0.429\mathbf{47.974 \pm 0.429}52.095±0.926\mathbf{52.095 \pm 0.926}53.197±1.105\mathbf{53.197 \pm 1.105}85.01±2.22\mathbf{85.01 \pm 2.22}18.051±0.269\mathbf{18.051 \pm 0.269}

The results show that SeisTailFlow performs well on INSTANCE, with improvements, particularly ES@99.

Q4 We constrain the GPD shape and scale parameters through smooth transformations to ensure distributional validity and improve numerical stability during training. First, let r=∣y−y^∣r=|y-\widehat{y}| denote the absolute arrival-time error. For samples exceeding the tail threshold uu, we define the exceedance as: e=r−u,r>u.e=r-u,\qquad r > u. Here, ee represents the portion of the prediction error above the threshold. For example, a threshold of 0.30.3 s and an absolute error of 0.50.5 s yield an exceedance of 0.20.2 s. Strict selection using r>ur > u gives e>0e > 0; nonnegative exceedances also include the distribution boundary e=0e=0.

For the shape parameter, which controls tail heaviness, we transform the unconstrained network output aξa_{\xi} as: ξ=0.45,sigmoid⁡(aξ).\xi=0.45\\,\operatorname{sigmoid}(a_{\xi}). Thus, 0<ξ<0.450 < \xi < 0.45. Since the GPD has finite variance when ξ<0.5\xi < 0.5, this constraint excludes the infinite-variance parameter region.

For the scale parameter, which controls the scale of the exceedance distribution, we use: σ=exp⁡[4tanh⁡(aσ4)].\sigma=\exp\left[4\tanh\left(\frac{a_{\sigma}}{4}\right)\right]. This gives e−4<σ<e4e^{-4} < \sigma < e^4, or approximately 0.0183<σ<54.600.0183 < \sigma < 54.60, ensuring a positive scale while preventing it from approaching zero or growing without bound. These bounds apply on the scale used to model the exceedances; if the exceedances are standardized, the bounds cannot be interpreted directly in seconds. For the distributional support, the GPD requires: 1+ξeσ>0.1+\frac{\xi e}{\sigma} > 0. Because ξ>0\xi > 0, σ>0\sigma > 0, and e≥0e\ge0, we necessarily have: 1+ξeσ≥1>0.1+\frac{\xi e}{\sigma}\ge1 > 0.

All nonnegative exceedances satisfy the support condition, avoiding invalid log-likelihood evaluations caused by support violations. Together, these constraints improve numerical stability in extreme-error modeling by ensuring finite variance, a positive and bounded scale, and valid distributional support. We have revised the manuscript and uploaded the updated version. More details see Section 3.4 (P10) of our revised paper

Q5 We added a small-scale multi-station earthquake localization experiment comparing locations obtained from manual picks and model-predicted arrival times for the same 12 events in a compact region of California, each recorded by at least six physical stations. All methods use a fixed P-wave velocity model (VP=6.0V_P=6.0 km/s) and a multi-start least-squares procedure to jointly estimate hypocenter location, including depth, and origin time. Using catalog locations as references, we report median and P90 horizontal errors and median 3D error.

MethodMedian horizontal error (km) ↓\downarrowP90 horizontal error (km) ↓\downarrowMedian 3D error (km) ↓\downarrow
Ground-truth P-wave picks (reference)33.3781.3734.23
PhaseNet40.7987.5441.05
EQTransformer35.5088.3741.80
Armstrong45.6598.2048.35
Vanilla CFM41.8090.0044.70
CW-Flow43.6096.1046.20
Main Flow only42.2795.7545.45
SeisTailFlow34.1982.8139.67

The results show that SeisTailFlow performs best, providing evidence that reduced picking errors can translate into improved downstream localization. More details see Section 3.3 (P9-P10) of our revised paper.

We sincerely appreciate the reviewer.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

We sincerely appreciate your insightful comments. We have carefully addressed your concerns and submitted our rebuttal. We would be grateful if you could take a moment to review our responses and share any further comments. Thank you so much for your valuable time and consideration.

Comments on the paper

2 comments

Dear Reviewers,

The authors have replied to your initial comments. Please have a look and engage in the discussion.

Thanks!

Your AC

AuthorsOpenReview ↗

Summary and some illustrations. Thank you so much for your great efforts and kind help.

Dear AC,

We sincerely appreciate you and the three reviewers for reviewing our paper. We have addressed each reviewer’s comments directly and supplemented the experiments and methodological explanations. Two reviewers explicitly stated that their concerns had been resolved and raised their scores; the other reviewer acknowledged that our response addressed the performance concerns but maintained the original score because of the lack of evaluation on real continuous waveforms. To address further this concern, we have now added continuous-waveform experiments.

The main revisions and additional evidence include:

  • Real continuous-waveform evaluation. We used complete continuous recordings of the 2019 Ridgecrest earthquake sequence from the SCEDC archive, covering three stations for 72 hours each, totaling 216 valid station-hours. Input windows were generated automatically using a previously frozen STA/LTA triggering rule, without selection based on known events or manual arrival annotations; reference arrivals were used only for evaluation after inference. This extends the experiments from pre-segmented event windows to automatic triggering and picking on real continuous recordings, directly addressing the first reviewer’s main concern.

  • Additional datasets and regional validation. Beyond STEAD, we added a zero-shot evaluation of frozen models on the Italian INSTANCE dataset. After the National Earthquake Science Data Center approved our application for DiTing access, we also immediately completed supplementary DiTing experiments using event-disjoint training, development, and test splits, with 30,000 records in total, further examining performance across different data sources and acquisition settings.

  • Methodological and downstream validation. We added multi-station earthquake localization experiments, a direct comparison with Seis Main, and computational cost analyses. We also clarified the dual-flow target states, gradient propagation, tail-sample supervision, GPD parameter constraints, and differences between ablation configurations to help reviewers assess the methodological contributions and experimental settings.

We believe that these experiments on large-scale benchmark datasets can validate our method, and truly hope they will be recognized by the reviewer. 

Once again, we sincerely thank the AC and the reviewers for the patience in reviewing, great efforts, and valuable help for our paper.

The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.