{"id":"c379ec6c-b416-465e-b530-ecb8e61cb219","arxiv_id":"2412.18099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SENSE, an attention-based multistation model with per-station embeddings, improves earthquake intensity prediction over ISMP and TEAM on Taiwan data, but not on Japan and with no public code.","lead":"This paper introduces SENSE, a deep-learning system that predicts earthquake shaking intensity by combining signals from many seismic stations at once. It reports better warning accuracy than two existing systems on Taiwan data, with caveats about data access, warning lead time, and deployment flexibility.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The input waveform window and alarm threshold are never specified, so the F1/leading-time trade-off in Tables III–VI is uninterpretable and the central early-warning claim is not yet reproducible.","rationale":"The reader's weakest assumption focuses on fixed-station embeddings and generalization to newly installed stations. That is a legitimate deployment limitation, and it is discussed in the paper only obliquely. However, it does not threaten the validity of the Taiwan benchmark itself: even with known stations, the reported F1 and leading-time numbers must be produced by some well-defined input window and alarm threshold. The paper omits both. The model cannot be using the full 120-second record, because then the peak PGA would be directly visible in the input and the learning task would be trivial; therefore some prefix window is used, but its length and trigger alignment are never stated. Similarly, the continuous model's alarm rule is described only as 'if the probability of a predefined PGA exceeds a threshold' with no threshold value. Without these specifications, Tables III–VI cannot be reproduced, and the comparison with TEAM and ISMP could be driven by operating-point choices rather than by the model architecture. This is a more direct threat to the central claim than the fixed-station issue, because it calls into question the benchmark evidence itself. The paper does deserve credit for event-based temporal splitting, for reporting precision/recall in addition to F1, and for including ablation study (Table VII) showing the contribution of locality-specific encodings. Those strengths mean the work is likely conditionally acceptable once the missing protocol details are supplied, so I do not recommend changing the reader's CONDITIONAL verdict.","tokens_in":15026,"tokens_out":9255,"duration_ms":101441,"concrete_test":"Ask the authors for (i) the exact input window length and P-trigger alignment used for Taiwan and Japan, and (ii) the alarm probability thresholds used for each PGA level. Then re-run the Table V comparison with a causal sliding-window protocol (e.g., 1, 3, 5, 10 s after trigger) and a threshold sweep for both TEAM and SENSE. If the reported SENSE F1 margin over TEAM disappears or shrinks when thresholds are selected on validation data, or if the leading times change materially with window length, the central claim needs to be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence for SENSE is the Taiwan comparison in Table V, but the paper never defines the input waveform window. The model takes w ∈ R^{3×T} (Section III-A); the Japan dataset is described as '15 seconds of pre-trigger data for a total length of 120 seconds,' yet there is no statement of how much of each record is fed to the convolution module or how the window is aligned to the P trigger. Since the task is to predict PGA from a prefix available before strong motion, the reported precision, recall, and the leading times in Table VI depend entirely on this window. If a large fraction of the strong-motion record is inside the input, the F1 scores are inflated and the 'early warning' claim is not supported; if the window is short, the leading-time numbers could be optimistic. Additionally, the continuous model requires an alarm probability threshold (Section III-B), but the threshold(s) used for Tables III–V are never given. Both window length and alarm threshold are free parameters; without them the comparison against TEAM and ISMP is not reproducible. The fixed-station embedding limitation identified by the reader is real, but it is a generalization issue; the window/threshold ambiguity affects the validity of the headline benchmark itself.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SENSE, a deep learning framework for earthquake early warning intensity prediction that takes multistation waveform statistics and geographical information as input. The architecture combines a convolutional front-end, sinusoidal positional encoding, learnable per-station early and late locality-specific embeddings, and a self-attention-based feature blending module (Transformer or Conformer), with either a discrete classification head or a continuous mixture-density-network head. The model is evaluated on Taiwan and Japan strong-motion datasets using event-based chronological splits, and compared against two baselines (ISMP and TEAM) on the Taiwan dataset. The authors report that SENSE outperforms both baselines at all five PGA thresholds on Taiwan, and also provide ablation studies of the model components. The central claim is that SENSE delivers competitive or better accuracy than state-of-the-art methods by exploiting multistation information.","tokens_in":15346,"tokens_out":3511,"duration_ms":36052,"significance":"If the reported results are reproducible, SENSE would represent a substantial improvement over two published deep-learning baselines for intensity-based earthquake early warning, with F1 gains on the Taiwan test set of roughly 0.3 or more at several thresholds. The paper's strengths include the event-based train/validation/test split, the use of two national-scale datasets, the explicit comparison to TEAM and ISMP, and the systematic ablation of the proposed components. The architecture is clearly described with reference to specific equations, and the training schedule is reported. However, the current evaluation omits several operationally critical details (input waveform window, alarm probability threshold, leading-time definition), and the comparison is confined to a single dataset without uncertainty quantification. These gaps currently prevent the reader from assessing whether the headline accuracy gains are real or an artifact of the evaluation protocol.","major_comments":[{"comment":"The evaluation protocol is underspecified in a way that is load-bearing for the central claim. The input waveform length T in w ∈ R^{3×T} (Section III-A) is never stated, and the Japan dataset description (\"15 seconds of pre-trigger data for a total length of 120 seconds\") does not say how much of each record is fed to the convolution module or how the input window is aligned to the P trigger. Since the task is early warning, the precision/recall values in Tables III–V and the leading times in Table VI depend critically on whether a prefix of the record or the full strong-motion record is used. Additionally, the continuous model in Eqs. (10)–(13) requires a probability threshold to issue an alarm, but the threshold used to compute the reported precision/recall/F1 is never given, and the quantity \"Leading Time\" in Table VI is never defined. These three missing details make the central comparison in Table V impossible to reproduce or interpret.","section":"Sections III-A, IV-A, IV-B2, Tables III–VI"},{"comment":"The comparison against ISMP and TEAM is reported only for Taiwan (Table V), while the Japan experiments in Table III have no baseline results. The abstract's claim that SENSE is \"competitive or even better\" than state-of-the-art methods therefore rests on a single national dataset. Furthermore, no confidence intervals, error bars, or significance tests are reported for any comparison, and the test set is a single chronological split. Given the very large F1 gaps in Table V (e.g., 0.692 vs. 0.306 at 0.81%g), it is important to know whether these gaps are stable across random seeds or multiple splits; I would like to see repeated runs with variance estimates, or a significance test, and ideally Japan baseline numbers.","section":"Section IV-D"},{"comment":"The early and late locality-specific embeddings are per-station learned parameters indexed by station identity. For a station not present in the training set, these embeddings are undefined, so the model cannot produce a prediction at that station. This contradicts one of the stated motivations, namely the ability to warn distant areas, unless those areas already contain a station that was in training. The paper should either explain how embeddings are obtained for unseen stations or explicitly scope the method to a fixed station configuration.","section":"Section III-A, Eqs. (4) and (6)"}],"minor_comments":[{"comment":"There is a typo in the caption: \"dataseet\" should be \"dataset.\"","section":"Table V caption"},{"comment":"The text states that the continuous objective with the Transformer is a \"better and more stable choice,\" but in Table IV the discrete Conformer result at 14%g (F1 = 0.510) exceeds the continuous Transformer result at the same threshold (F1 = 0.469). The claim should be qualified as applying to most but not all thresholds.","section":"Section IV-C"},{"comment":"In the ablation study, configuration (D) on the Taiwan dataset at 8.1%g (F1 = 0.529) beats the full SENSE model (F1 = 0.503). The text says SENSE achieved the best results \"in most cases,\" which is accurate, but this counterexample could be acknowledged for completeness.","section":"Section VII (Ablation), Table VII"},{"comment":"The three-stage training schedule is described, but the hyperparameter values (e.g., learning rate, batch size, optimizer) are not reported, which makes reproduction difficult. This is a presentation issue but should be addressed in a revision.","section":"Section IV-B1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central empirical claim is promising, but the missing evaluation-protocol details (input window, threshold, leading-time definition) are precisely the kind of thing that would prevent a reader from trusting the headline numbers, and they are fixable in revision. The fixed-station embedding issue is also real and should be addressed either by a practical solution or by a clearer scope statement. I do not see grounds for rejection, because the architecture and dataset descriptions are otherwise detailed and the Taiwan comparison, if confirmed, would be a meaningful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the architecture is a real but modest step past TEAM, and the Taiwan numbers are impressive. But the paper never says how much of each waveform is fed to the model or what alarm probability threshold was used, so the headline benchmark is not reproducible and the early-warning claim is not yet validated.\n\nWhat the paper does well: the event-based chronological split is correct, the ablation study is genuinely informative, and the comparison on Taiwan is honest in spirit — they acknowledge that TEAM and SENSE consider all stations while ISMP only scores stations with records, which makes the comparison harder for the multistation methods. The per-station early and late locality-specific embeddings are a plausible way to absorb site effects, and the learnable weighting between waveform and geographic encodings is a reasonable addition. The model-size and execution-time table is useful context.\n\nNow the soft spots, in order of severity. First, the stress-test concern is on target: the input waveform window w ∈ R^{3×T} is never specified, and the Japan data note of \"15 seconds pre-trigger, 120 seconds total\" does not tell the reader what portion the convolution module sees. Since the task is to predict PGA from a prefix before strong motion, every precision/recall/F1 and every leading-time number in Tables III–VI depends on this window. If most of the strong-motion record is inside the input, the F1 scores are inflated. Relatedly, the continuous model issues an alarm when the probability of exceeding a PGA exceeds a threshold, but the threshold used for the tables is never given. Both are free parameters, and without them the central comparison against TEAM and ISMP is uninterpretable.\n\nSecond, the fixed-station embeddings are a genuine generalization problem: a newly installed station would have no learned embedding, and the paper never discusses how predictions would be made for stations absent from training. That limits the claimed ability to warn distant areas. Third, there are no baselines on Japan, no error bars or significance tests, and no code or data released. Those are common in this literature but they matter because the Taiwan gap is huge — SENSE F1 0.692 vs TEAM 0.306 at 0.81%g — and without uncertainty or an independent dataset, I can't tell how much is architecture versus station-memorization or threshold tuning.\n\nThe paper is worth a serious referee, but only with a request for the missing experimental details. I would not desk-reject it. The core idea is sensible, the ablation is careful, and the authors are clearly capable. For a reader in the EEW-ML niche, the paper is useful once the window and threshold are stated. I'd bring it to a reading group as a case study in reproducibility gaps.\n\nRecommendation: send to peer review with major revision. Require the input window and alarm threshold, a Japan baseline, and a discussion of how the model handles unseen stations.","headline":"Genuine but incremental extension of TEAM; strong Taiwan result is undercut by missing input-window and alarm-threshold specifications.","tokens_in":15807,"tokens_out":1925,"would_cite":true,"duration_ms":21577,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SENSE, an attention-based model fed waveforms and coordinates from all stations in a region, predicts earthquake intensity at each station more accurately than the compared single-station and multistation baselines.","keywords":["earthquake early warning","intensity prediction","peak ground acceleration","multistation modeling","self-attention","locality-specific embeddings","Transformer","mixture density network"],"falsifier":"Train SENSE on early Taiwan events, hold out one station from training entirely, and evaluate F1 at that station on later events with its embedding initialized as the mean of the trained stations; if the held-out station's F1 collapses to the no-embedding ablation level, the model is memorizing stations rather than generalizing across the network.","tokens_in":14842,"feed_emoji":"🌏","tokens_out":9286,"duration_ms":84477,"temperature":0.7,"pith_summary":"SENSE is an attempt to show that earthquake early warning should use the whole station network, not just the closest sensor. The model takes statistics from a set of stations, lets attention layers exchange information between them, and outputs an intensity prediction for every station, including ones that have not yet recorded the event. On the Taiwan dataset the continuous-objective Transformer variant reports F1 scores of 0.692, 0.630, 0.503, 0.469, and 0.333 at the five PGA thresholds, compared with 0.306, 0.261, 0.165, 0.075, and 0.008 for the TEAM baseline. The paper argues that this shows multistation attention with per-station embeddings can give earlier and more reliable warnings to distant areas.","feed_headline":"Attention across seismic stations beats single-station quake alerts","feed_subtitle":"A multistation model predicts shaking at stations that have not yet felt the quake, using the whole network's waveforms.","key_machinery":"The load-bearing component is the pair of locality-specific embedding tables: for each station $n$ the model learns an early embedding $\\mathbf{l}_n^e$ added to the fused waveform-plus-coordinate representation and a late embedding $\\mathbf{l}_n^l$ added after the attention layers (Eqs. 4 and 6). These vectors let the network store station-dependent biases that the raw waveforms and coordinates do not carry. Between them sits the feature blending module, several stacked self-attention layers (Transformer or Conformer), which lets information from stations that have already felt the quake flow to stations that have not. A learnable scalar $\\alpha_n$ per station balances CNN-extracted waveform features against sinusoidal positional encoding of longitude, latitude, and instrument height. For continuous predictions, the output head is a mixture density network whose Gaussian mixture is integrated to give the probability that peak ground acceleration at a station exceeds a warning threshold.","core_discovery":"The central claim is that an encoder–decoder network which blends information across stations with stacked self-attention, and which injects a learned embedding for each station both before and after the attention layers, can predict peak-ground-acceleration intensity levels better than a single-station CNN (ISMP) and a multistation Transformer baseline (TEAM). The reported Taiwan results put SENSE ahead of TEAM at every PGA threshold, with the largest gap at 25%g where F1 rises from 0.008 to 0.333. The paper attributes this to the locality-specific embeddings and to the learnable weighting between waveform and geographic information, and it identifies the continuous Gaussian-mixture objective as the more stable configuration across both the Japan and Taiwan datasets.","pith_inferences":["The paper does not test how SENSE would predict for a station installed after training, since each station's identity is a hard-coded embedding; a natural extension is to replace the lookup table with an embedding generated from station coordinates and site conditions.","Both evaluation networks are dense national arrays (707 stations in Japan, 250 in Taiwan), so the benefit on sparse regional networks is unknown; reducing station density in the input and measuring the F1 drop would test how much of the gain depends on having many nearby stations.","The paper reports average execution and leading times but not an end-to-end trigger-to-alarm latency analysis; a pipeline-level evaluation would show whether the accuracy gain survives the time budget of a real early-warning system."],"forward_implications":["A single forward pass over a regional network can emit intensity estimates for every station as soon as the first stations report P-wave statistics, so warnings can target stations that have not yet felt the shaking.","The continuous Gaussian-mixture objective is the recommended configuration: it yields higher F1 scores than the discrete classification objective on both datasets and produces an exceedance probability that maps directly to alarm thresholds.","The ablation study indicates that both the early and late locality-specific embeddings contribute to the gain, with early embeddings helping most at high PGA levels and late embeddings at low PGA levels, so per-station bias modeling is a substantive part of the method.","On the Taiwan comparison, SENSE's F1 scores exceed TEAM's by more than a factor of two at every threshold, which the paper takes as evidence that multistation attention with per-station embeddings uses network-wide data more effectively than TEAM's coordinate-cross-attention design."],"supporting_citations":[{"why":"Supplies the ISMP single-station CNN baseline whose PGA-threshold predictions SENSE is compared against.","marker":"[13]"},{"why":"Supplies the TEAM multistation Transformer baseline, the main model SENSE is designed to beat.","marker":"[16]"},{"why":"Provides the Japan dataset used for training and evaluation.","marker":"[21]"},{"why":"Supplies the self-attention architecture and positional encoding that the feature blending module is built on.","marker":"[20]"},{"why":"Provides the PGA threshold set used for the Taiwan evaluation.","marker":"[45]"},{"why":"Supplies the mixture density network used for the continuous PGA-probability objective.","marker":"[48]"}],"fun_headline_variants":["Attention across stations predicts quake shaking before it hits","SENSE uses station attention to give earlier quake intensity forecasts","Multistation attention network outperforms single-station quake alerts","Attention over seismic network predicts shaking at unalerted sites","Earthquake early warning improved by multistation attention"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model depends on per-station identity embeddings learned during training, so it has no defined way to predict for a station that was not in the training set; a real network that gains or moves stations is exactly the case where this assumption would break the claimed distant-warning ability.","fun_headline_variants_meta":{"raw":{"variants":["Attention across stations predicts quake shaking before it hits","SENSE uses station attention to give earlier quake intensity forecasts","Multistation attention network outperforms single-station quake alerts","Attention over seismic network predicts shaking at unalerted sites","Earthquake early warning improved by multistation attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001349,"raw_usage":{"total_tokens":5476,"prompt_tokens":939,"completion_tokens":4537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":4455}},"tokens_in":555,"tokens_out":4537,"duration_ms":28155,"temperature":1.0,"reasoning_tokens":4455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:01:49.794348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SENSE on early Taiwan events, hold out one station from training entirely, and evaluate F1 at that station on later events with its embedding initialized as the mean of the trained stations; if the held-out station's F1 collapses to the no-embedding ablation level, the model is memorizing stations rather than generalizing across the network.","supporting_citations":[{"cited_title":"Neural network-based strong motion prediction for on-site earthquake early warning","cited_arxiv_id":null,"evidence_quote":"Supplies the ISMP single-station CNN baseline whose PGA-threshold predictions SENSE is compared against."},{"cited_title":"The transformer earthquake alerting model: A new versatile approach to earthquake early warning","cited_arxiv_id":null,"evidence_quote":"Supplies the TEAM multistation Transformer baseline, the main model SENSE is designed to beat."},{"cited_title":"Recent progress of seismic observation networks in japan—hi-net, f-net, k-net and kik- net—","cited_arxiv_id":null,"evidence_quote":"Provides the Japan dataset used for training and evaluation."},{"cited_title":"End- to-end PGA estimation for earthquake early warning using transformer networks","cited_arxiv_id":null,"evidence_quote":"Provides the PGA threshold set used for the Taiwan evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mixture density network used for the continuous PGA-probability objective."}],"review_version":1}