{"id":"cc162b43-d916-4607-94ad-d99fe4bdc066","arxiv_id":"2506.06939","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PhaseNet+ is a multitask network that performs phase picking, polarity determination, and origin time prediction in one pass, producing Ridgecrest catalogs comparable to dedicated association methods with up to four times more focal mechanisms.","lead":"This paper introduces PhaseNet+, a single deep learning model that simultaneously detects earthquake wave arrivals, determines first-motion polarity, and predicts origin times to link picks across stations. It aims to replace separate neural networks in earthquake monitoring with one efficient pass that yields larger catalogs and more focal mechanisms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'rival state-of-the-art' claim for PhaseNet+ rests on an origin-time association step that is never specified and never validated against independent association labels; without this, the Ridgecrest event count is uninterpretable.","rationale":"Two candidate concerns were considered. The first is the accuracy of the origin-time proxy at long distances and in dense sequences; the reader's weakest_assumption covers this. The second, which I believe is the more load-bearing, is that the association algorithm itself is not specified. On page 4, the method says only that association simplifies to identifying picks with shared onset times; no clustering, threshold, or multi-event handling is given. On pages 10-11, the Ridgecrest application states that the model 'predicts event origin times for association' and compares the resulting catalog to GaMMA, but never describes how origin-time predictions are turned into associations. Without this, the paper's strongest empirical evidence—the 23,103-event catalog and its comparison to GaMMA—cannot be reproduced or checked. The Discussion explicitly concedes the association approach degrades for near-simultaneous events, yet the Ridgecrest aftershock sequence is precisely such a setting. The paper also lacks any comparison to an independent high-resolution catalog (e.g., Shelly 2020), so the increased event count and reduced spatial scatter are not anchored to ground truth. This does not invalidate the model's phase-picking and polarity contributions, which are well supported by Figures 4-6 and the open-source implementation, but it leaves the central 'end-to-end' association claim conditional on details that are not provided.","tokens_in":16251,"tokens_out":7882,"duration_ms":81931,"concrete_test":"Obtain the association code from the linked repository (https://github.com/AI4EPS/EQNet) and reproduce the 23,103-event PhaseNet+ catalog for the same 7-day Ridgecrest window. If the association step is not present or does not exactly reproduce the catalog, the claim is not reproducible. Additionally, compare the PhaseNet+ catalog against the Shelly (2020) high-resolution Ridgecrest catalog on 100 randomly selected aftershocks using a 2 km / 2 s association criterion; report precision and recall of the origin-time associations. If recall is below an acceptable threshold or a large fraction of PhaseNet+ events are unmatched, the 'rival SOTA' assertion is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 states that phase association 'simplifies to identifying picks with shared onset times', but no algorithm is given for converting per-station origin-time predictions into an associated event catalog: no clustering method, threshold, or handling of multiple events at one station. In Section 3.2 the PhaseNet+ catalog is compared to a GaMMA catalog built from the same picks, but the PhaseNet+ association pipeline is never described, so the comparison cannot be reproduced. The only quantitative validation of origin-time accuracy (Fig. 7) is on isolated, pre-selected test waveforms; it does not test association in continuous data with overlapping aftershocks. The related limitation is explicitly acknowledged in Section 4: origin-time association 'becomes less effective when multiple earthquakes occur nearly simultaneously, e.g., in dense aftershock sequences' — exactly the Ridgecrest setting used to claim that the end-to-end approach 'rival[s] state-of-the-art methods.' Since no ground-truth association labels or a high-quality reference catalog (e.g., Shelly, 2020) are compared, the 73–81% event increase over SCSN and the reduced spatial scatter relative to GaMMA could reflect association artifacts, not genuinely better monitoring. This is load-bearing because the central novelty is the origin-time association constraint; if the association step is an unspecified heuristic, the main claim cannot be adjudicated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents PhaseNet+, a multitask extension of the PhaseNet phase-picking architecture that jointly outputs P/S arrival-time picks, first-motion polarities, and a per-station origin-time estimate intended to support phase association. The authors train PhaseNet+ on the Northern California portion of the CEED dataset and evaluate it on 2023 data from Northern and Southern California, comparing phase picking with PhaseNet, polarity picking with the CNN Ross model, and origin-time accuracy against manually reviewed labels. They then apply the model to one week of the 2019 Ridgecrest sequence, producing an event catalog via origin-time-based association and a focal-mechanism catalog using SKHASH. The reported results show phase-picking parity with PhaseNet, polarity accuracy comparable or superior to the CNN Ross baseline, origin-time errors mostly within 2 s, a PhaseNet+ catalog with 73–81% more events than SCSN and less spatial scatter than a GaMMA catalog built from the same picks, and up to 7,815 focal mechanism solutions. The central claim is that origin-time prediction enables a simplified end-to-end association workflow that rivals dedicated phase association methods.","tokens_in":173,"tokens_out":3640,"duration_ms":77502,"significance":"If the central claim is upheld, PhaseNet+ would be a practically valuable contribution: a single model that simultaneously provides phase picks, polarities, and origin-time constraints could streamline earthquake monitoring pipelines and make focal-mechanism catalogs much more complete. The authors provide code and promise SeisBench integration, and the training/evaluation on a large, publicly available dataset is a strength. The phase-picking comparison against PhaseNet is a clean same-training-set test, and the Ridgecrest application demonstrates a realistic use case. However, the novelty rests on the origin-time association concept, and that concept is underspecified and under-validated in the manuscript as written. Because the association step is not described as an algorithm and is only tested on preselected single-event waveforms, the event-catalog comparisons in Section 3.2 are not yet sufficient to establish that PhaseNet+ rivals state-of-the-art association methods.","major_comments":[{"comment":"The manuscript says phase association 'simplifies to identifying picks with shared onset times,' but no concrete algorithm is given. There is no description of how per-station origin-time predictions are turned into associated events: no clustering method, no threshold on origin-time agreement, no procedure for handling multiple events at one station, and no treatment of false or missing origin-time predictions. This is the core novelty of the paper, and without this specification the PhaseNet+ catalog in §3.2 cannot be reproduced and the comparison with GaMMA is not interpretable. The authors should provide a precise association procedure, including any hyperparameters, or clearly state that the association is performed by an external algorithm whose inputs include the origin-time predictions.","section":"§2.1, phase association paragraph"},{"comment":"The claim that the end-to-end approach 'rival[s] state-of-the-art methods' is supported only by internal comparisons: the PhaseNet+ catalog is compared with SCSN and with a GaMMA catalog built from the same PhaseNet+ picks, but there is no independent ground-truth association reference. The 73–81% increase in event count over SCSN and the reduced spatial scatter relative to GaMMA could be influenced by association artifacts, such as splitting one earthquake into multiple events or merging distinct events with similar origin times. The acknowledged limitation in §4 that origin-time association 'becomes less effective when multiple earthquakes occur nearly simultaneously' is directly relevant to the dense aftershock setting of the Ridgecrest application. A validation against a high-quality reference catalog for the same period, e.g., Shelly (2020), or against at least a set of manually reviewed associations, is needed before the 'rival state-of-the-art' conclusion can be accepted.","section":"§3.2 and Figure 9"},{"comment":"The polarity comparison with 'CNN Ross' is not apples-to-apples. The CNN Ross model is described as pre-trained on Southern California data prior to 2018, whereas PhaseNet+ was trained on the Northern California CEED training set (with a different task formulation and architecture). On the Southern California 2023 test set, the comparison is between a model trained on Northern California and a model trained on older Southern California data; on the Northern California 2023 test set, CNN Ross is evaluated out-of-region. The conclusion that PhaseNet+ achieves 'comparable or superior performance' is plausible, but the asymmetry in training data makes the quantitative comparison difficult to interpret. The authors should either retrain CNN Ross on the same Northern California training data, or at least clearly qualify the comparison as a benchmark against a published, frozen baseline rather than a matched model comparison.","section":"§3.1.2 and Figure 5"},{"comment":"The origin-time accuracy evaluation in Figure 7 is performed on isolated, pre-selected test waveforms, not on continuous data with overlapping or near-simultaneous events. The stated purpose of origin-time prediction is phase association in continuous data, yet no experiment measures how well the origin-time predictions separate distinct events in a continuous stream. The §4 statement that the approach 'becomes less effective when multiple earthquakes occur nearly simultaneously, e.g., in dense aftershock sequences' is in tension with the §3.2 claim that the method 'works as effectively for dense aftershocks as dedicated phase association algorithms.' The authors should add a continuous-data evaluation, even a small annotated segment, that tests association performance against known event times, or explicitly narrow the claim about dense aftershock performance.","section":"§3.1.3 and §4"}],"minor_comments":[{"comment":"The SCSN earthquake catalog is attributed to Yang et al. (2012), but Yang et al. (2012) is the Southern California focal mechanism catalog; the routine SCSN event catalog is more commonly attributed to the Southern California Seismic Network/SCEDC (e.g., Hutton et al., 2010). Please clarify the source of the SCSN event catalog used in Figure 9.","section":"§3.2, paragraph on catalogs"},{"comment":"The code is hosted in the EQNet repository, but the paper announces PhaseNet+ as an extension; please provide a specific tag, branch, or model checkpoint identifier so that the exact PhaseNet+ version trained and evaluated here can be retrieved reproducibly.","section":"Open research statement"},{"comment":"The caption says 'green ticks indicating predicted earthquake origin times' but the histogram in panel (b) is not explicitly linked to the per-station green ticks; a sentence clarifying how the histogram is built from the per-station predictions would improve readability.","section":"Figure 8 caption"},{"comment":"The inversion with SKHASH uses 'SP amplitude ratio data,' but the manuscript does not describe how the S/P amplitude ratios are measured from the PhaseNet+ outputs or whether they come from a separate routine. A brief description or reference for this measurement step would make the focal-mechanism workflow more complete.","section":"§3.2, focal mechanism section"},{"comment":"There are a few minor grammatical issues, such as 'detecting the average of P and S arrival-times as a segmentation task' in §2.1, where 'detecting' should be 'predicting' or 'estimating' for clarity.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the multitask idea is timely. The main technical gap is the unspecified and insufficiently validated origin-time association step; I believe this is fixable with additional methodological detail and at least one independent association validation. The self-citation pattern is heavy but not inappropriate given the authors' direct prior work on PhaseNet, CEED, GaMMA, and EQNet. I would not recommend rejection, because the phase-picking and polarity components are concrete and the underlying approach is promising."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: PhaseNet+ is a sensible incremental step—one U-Net doing picks, polarity, and origin time in a single pass—and the Ridgecrest application shows it can produce large catalogs and focal mechanisms quickly. The central novelty is the origin-time association constraint, and that is exactly where the evidence is thinnest. The stress-test note is right: the association step is never specified, so the headline claim cannot be fully adjudicated.\n\nWhat's actually new: combining phase picking, first-motion polarity, and origin-time prediction in one model, with the origin-time branch serving as an association prior. Prior work did picking+polarity or picking+association separately, so this combination is new. The architecture is a natural extension of PhaseNet, and the training setup on CEED (325K events, 1.1M waveforms) is solid. Phase picking parity with PhaseNet is expected and fine; the polarity results are plausible. The four-fold increase in focal mechanisms is striking, and the promised code release plus SeisBench integration is real, reproducible evidence.\n\nSoft spots, in order. The association algorithm is missing: Section 2.1 says association 'simplifies to identifying picks with shared onset times,' but no clustering method, threshold, or handling of multiple events per station is given. The Ridgecrest PhaseNet+ catalog is compared against a GaMMA catalog built from the same picks, but the PhaseNet+ association pipeline is not described, so the comparison cannot be reproduced. Origin-time accuracy in Fig. 7 is tested on preselected waveforms, not continuous data with overlapping aftershocks. The paper itself acknowledges the method degrades in dense aftershock sequences—precisely the Ridgecrest setting. Without independent association labels or a high-quality reference catalog like Shelly (2020), the 73–81% event increase over SCSN could include association artifacts. The polarity baseline CNN Ross is pretrained on older Southern California data and not retrained on the same training set, so that comparison isn't apples-to-apples. The S/P amplitude ratio workflow used for the focal mechanisms is undocumented. None of these are fatal; they are gaps in the evidence.\n\nThe architecture is sound and the paper is clearly written. It deserves peer review, though the review should insist on the association details, independent validation of the association step, and a documented S/P ratio procedure. I'd bring it to a reading group if seismologists are in the room, and I'd cite it as an example of multitask monitoring.","headline":"Useful incremental multitask model with a genuinely new combination, but the origin-time association claim is under-specified and not independently validated.","tokens_in":17054,"tokens_out":1693,"would_cite":true,"duration_ms":19736,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PhaseNet+ simultaneously picks phases, reads polarities, and estimates origin times from raw waveforms.","keywords":["multitask deep learning","seismic phase picking","first-motion polarity","phase association","origin time prediction","Ridgecrest earthquake sequence","focal mechanisms","PhaseNet+"],"falsifier":"Apply PhaseNet+ to a sequence with a high-precision reference catalog that includes near-simultaneous earthquakes, and check whether predicted origin times separate events whose true origin times differ by less than about 2 s; if a substantial fraction of associations merge or split at that threshold, the origin-time short-cut fails. A companion check is to measure origin-time prediction error as a function of epicentral distance beyond about 100 km, since the paper already reports growing errors for travel times exceeding 15 s, and to test whether association accuracy degrades exactly in that regime.","tokens_in":16040,"feed_emoji":"🌍","tokens_out":11280,"duration_ms":97256,"temperature":0.7,"pith_summary":"This paper claims that one deep network can replace the usual chain of separate seismology tools by simultaneously picking P and S arrival times, reading first-motion polarity, and predicting event origin times directly from continuous waveforms. The authors extend PhaseNet, a U-Net phase picker, into PhaseNet+ by adding two branches: one that classifies polarity on the vertical component and one that detects events and regresses origin time. The origin-time prediction is meant to simplify phase association: instead of searching over travel-time moveouts, the network groups picks that share a common origin time. If this works, a single forward pass over a waveform archive yields the measurements needed to build earthquake catalogs with locations and focal mechanisms, avoiding information loss and repeated scanning. The paper reports that on the 2019 Ridgecrest sequence the approach rivals dedicated association methods and yields up to 7,815 focal mechanism solutions versus 1,875 in the routine catalog.","feed_headline":"One network does picking, polarity, and association in a single pass","feed_subtitle":"On the 2019 Ridgecrest sequence it rivaled expert catalogs and quadrupled focal mechanism solutions.","key_machinery":"The central object is the multitask U-Net architecture of PhaseNet+: a shared encoder feeding three branches, namely a phase arrival-time segmentation branch that labels P, S, and noise; a polarity branch that outputs a per-sample score in $[-1,1]$ on the vertical component; and an event-detection branch that both segments the midpoint of P and S arrival times and regresses the origin time. The load-bearing identity is the station-level origin-time proxy: the midpoint of P and S arrival times approximates the origin time up to a distance-dependent offset, and the regression branch learns that correction, so phase association reduces to grouping picks with similar predicted origin times. This shared-encoder design lets the three tasks reinforce one another, and the origin-time branch is what makes a single forward pass sufficient to provide association constraints.","core_discovery":"PhaseNet+ performs phase arrival-time picking, first-motion polarity determination, and origin-time prediction for phase association in a single forward pass, built on a shared U-Net encoder with three output branches. The phase arrival and polarity branches share hidden layers; polarity is scored in $[-1,1]$ at every sample of the vertical component, and the arrival-time branch determines where on the waveform the polarity score is read. Event detection is framed as a segmentation task that locates the midpoint of P and S arrival times, and the origin time is predicted by regression, so that phase association becomes the grouping of picks with similar predicted origin times. Trained on the California Earthquake Event Dataset, the model matches PhaseNet in pick timing, matches or beats the CNN Ross baseline in polarity classification, and keeps origin-time errors mostly within 2 s. In the 2019 Ridgecrest application, the workflow produced 23,103 events (73-81% more than the SCSN catalog), less spatial scatter than the GaMMA-based reference association, and up to 7,815 focal mechanism solutions from combined polarity and S/P amplitude ratio data.","pith_inferences":["If the origin-time proxy degrades with distance, as the paper's own error growth beyond 15 s travel time suggests, the association shortcut may need a distance-aware correction or a hybrid approach when networks include far stations.","The multitask design could be extended to multi-station input, so that origin-time consistency across stations is learned jointly rather than combined afterwards; the paper notes this as a possibility but does not implement it.","Applying PhaseNet+ to induced or volcanic sequences, where event rates and waveform similarity differ from California, would test whether the polarity-reclassification gain and origin-time association margins transfer beyond the training region.","The reported four-fold increase in focal mechanisms implies that even small events, which routine catalogs leave without mechanisms, may carry usable polarity information; this could change how completeness of mechanism catalogs is measured."],"forward_implications":["A single scan of continuous waveforms can yield picks, polarities, and origin times, so earthquake catalogs can be built without running separate phase-picking, polarity-picking, and association models.","Because polarity is extracted directly from the phase-arrival branch, the model reclassifies roughly 50-60% of manual 'unknown' polarities as up or down, increasing the yield of usable first-motion readings for focal mechanism inversions.","On the 2019 Ridgecrest sequence the end-to-end workflow produced 73-81% more events than the SCSN catalog and focal mechanism counts up to about four times SCSN's, with faulting style and stress directions consistent across catalogs.","The multitask template is not specific to PhaseNet: any sequence-to-sequence phase picker, such as EQTransformer or PhaseNO, can be extended the same way.","The workflow separates perception (picking, polarity, association constraints) from physics-based inversion (location, focal mechanism), so deep learning handles pattern recognition while geophysical inversion preserves physical constraints."],"supporting_citations":[{"why":"PhaseNet is the base architecture and phase-picking formulation that PhaseNet+ extends into a multitask model.","marker":"(W. Zhu & Beroza, 2019)"},{"why":"The CNN Ross model is the polarity-picking baseline for comparison and motivates the polarity branch design.","marker":"(Ross, Meier, & Hauksson, 2018)"},{"why":"The California Earthquake Event Dataset provides the joint phase-pick, polarity, and source labels used to train the multitask network.","marker":"(W. Zhu, Wang, et al., 2025)"},{"why":"ADLoc is the location algorithm applied to PhaseNet+'s picks to build the Ridgecrest catalogs.","marker":"(W. Zhu, Rong, et al., 2025)"},{"why":"GaMMA is the dedicated association baseline that PhaseNet+'s origin-time-based association is compared against.","marker":"(W. Zhu, McBrearty, et al., 2022)"},{"why":"The SCSN catalog supplies the routine event and focal-mechanism counts that the enhanced catalogs are measured against.","marker":"(Yang et al., 2012)"},{"why":"The Cheng 2023 catalog is the deep-learning-enhanced focal-mechanism reference used to validate PhaseNet+ mechanism solutions.","marker":"(Cheng et al., 2023)"},{"why":"SKHASH is the inversion package that turns PhaseNet+ polarities and S/P ratios into focal mechanism solutions.","marker":"(Skoumal et al., 2024)"},{"why":"The QTM catalog's inter-event time statistics justify that 2 s origin-time errors are small enough for phase association.","marker":"(Ross, Trugman, et al., 2019)"}],"fun_headline_variants":["Multitask model does picking, polarity, association in one pass","PhaseNet+ unifies detection, polarity, association in single network","Deep learning model triples tasks: picking, polarity, association","All-in-one seismic AI: pick, polarize, associate in one shot","End-to-end quake monitoring with a single deep network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a station-level origin-time estimate, learned from the average of P and S arrival times with errors mostly within 2 s, is accurate enough to separate distinct earthquakes even in dense aftershock sequences where events can occur nearly simultaneously.","fun_headline_variants_meta":{"raw":{"variants":["Multitask model does picking, polarity, association in one pass","PhaseNet+ unifies detection, polarity, association in single network","Deep learning model triples tasks: picking, polarity, association","All-in-one seismic AI: pick, polarize, associate in one shot","End-to-end quake monitoring with a single deep network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001265,"raw_usage":{"total_tokens":5183,"prompt_tokens":951,"completion_tokens":4232,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":4142}},"tokens_in":567,"tokens_out":4232,"duration_ms":26708,"temperature":1.0,"reasoning_tokens":4142,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:45:02.416560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply PhaseNet+ to a sequence with a high-precision reference catalog that includes near-simultaneous earthquakes, and check whether predicted origin times separate events whose true origin times differ by less than about 2 s; if a substantial fraction of associations merge or split at that threshold, the origin-time short-cut fails. A companion check is to measure origin-time prediction error as a function of epicentral distance beyond about 100 km, since the paper already reports growing errors for travel times exceeding 15 s, and to test whether association accuracy degrades exactly in that regime.","supporting_citations":[{"cited_title":"\\ Beroza, G C","cited_arxiv_id":null,"evidence_quote":"PhaseNet is the base architecture and phase-picking formulation that PhaseNet+ extends into a multitask model."},{"cited_title":", Hauksson, E","cited_arxiv_id":null,"evidence_quote":"The SCSN catalog supplies the routine event and focal-mechanism counts that the enhanced catalogs are measured against."},{"cited_title":", Hauksson, E","cited_arxiv_id":null,"evidence_quote":"The Cheng 2023 catalog is the deep-learning-enhanced focal-mechanism reference used to validate PhaseNet+ mechanism solutions."},{"cited_title":", Hardebeck, J L","cited_arxiv_id":null,"evidence_quote":"SKHASH is the inversion package that turns PhaseNet+ polarities and S/P ratios into focal mechanism solutions."}],"review_version":1}