{"id":"a45a3095-d6e7-462b-a04f-58bc5e544fed","arxiv_id":"2506.13963","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Running a neural-network surge emulator over 900,000 synthetic hurricanes, the authors project that the US population at risk from the 100-year storm surge flood rises about 50% by 2100 under SSP5-8.5, driven mainly by sea level rise.","lead":"Researchers paired a deep-learning storm surge model with 900,000 simulated hurricanes to map how often catastrophic coastal flooding hits US shores now and in a warmer future. They project that by 2100 about 50% more coastal residents would face the once-a-century surge flood, driven mostly by sea level rise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DeepSurge's 100-year surge tail is unvalidated; if its error grows with surge magnitude, the +50% population-at-risk claim and the Georgia/South Carolina threshold shift, so a tail/extrapolation check is needed.","rationale":"I read the paper as an integrated risk assessment whose headline is a relative change. The authors are unusually transparent: they release the 900,000 surge fields, validate against ADCIRC, gauges, Needham, Gori, Muis, Katrina HWMs, and Crowell, and they show the headline survives their own bias correction (+57% vs +50%). That independent support is real. The remaining soft spot is the one the reader identified: the upper tail and the out-of-distribution future. The validation metrics are dominated by moderate events, and no test shows error as a function of surge height. The truncation test I propose would directly measure how much of the headline rests on extrapolation beyond the training distribution. If the test passes, the conditional verdict can be upgraded; if it fails, the 50% figure and the threshold finding should be re-framed. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":27759,"tokens_out":7966,"duration_ms":93712,"concrete_test":"Recompute the headline future population-at-risk after truncating all synthetic future TCs to the maximum intensity present in the 279-storm ADCIRC training set, and separately stratify DeepSurge's 71-storm test-set errors against ADCIRC by peak surge bin (e.g., 0-1, 1-2, 2-3, >3 m), reporting bias and RMSE in the top bin. If truncation drops the +2.3M increase by more than ~0.5M, or if the >3m tail bias exceeds ~0.3m, the +50% claim depends on unvalidated extrapolation and should be re-framed; if the increase remains near +40-50%, SLR dominates and the concern is secondary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The 50% increase in population at risk (4.6M to ~7M) requires DeepSurge to deliver unbiased estimates of the upper tail of storm surge, including for future synthetic storms roughly one Saffir-Simpson category stronger than anything in its 279-storm training set (SI S2.1, Fig. S1). The reported validation does not establish this. R2=0.815 against ADCIRC and R2=0.403 against tide gauges are aggregate metrics dominated by moderate surges; the +0.217m mean bias correction in SI S2.4 assumes a spatially smooth, surge-independent bias, and the Needham comparison (SI S2.5) has insignificant spatial correlation at only 18 points. If DeepSurge's bias or variance grows with surge height, both the historical 100-year map and the future-vs-historical difference shift. The nonlinear population-risk curves for Georgia and South Carolina (Fig. 4) amplify small surge errors near the ~2m threshold. The paper's own bias-corrected sensitivity (+57% vs +50%) shows the mean bias is not decisive, but it cannot rule out intensity-dependent tail error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an end-to-end deep-learning-based framework for projecting U.S. coastal storm surge risk. The authors train a point-based recurrent-convolutional network (DeepSurge) on 279 historical ADCIRC storm simulations, then apply it to 900,000 synthetic tropical cyclones generated by the RAFT model under historical and SSP5-8.5 end-of-century conditions. Surge return levels are combined with probabilistic sea-level rise projections and a bathtub-style inundation model (CA-Surge) to estimate population at risk for the historical and future 100-year flood. The headline result is a 50% increase (from 4.6 to ~7 million) in U.S. coastal population at risk, driven mainly by sea-level rise (+1.9 million) and secondarily by changed TC behavior (+0.24 million), with Florida and the southeast Atlantic coast (notably Georgia and South Carolina) showing pronounced increases.","tokens_in":27814,"tokens_out":5680,"duration_ms":61055,"significance":"If the results hold, the paper provides a computationally tractable approach to national-scale, probabilistic storm surge risk assessment with an unprecedentedly large synthetic event catalog (900,000 storms). The explicit decomposition of future risk into sea-level rise and TC-behavior contributions, the identification of nonlinear population-at-risk thresholds in Georgia and South Carolina, and the public release of DeepSurge-predicted surge fields are valuable contributions. The paper also demonstrates that its headline change is qualitatively robust to a spatially smoothed mean-bias correction (+57% versus +50%), which strengthens confidence in the direction of the projection. The main significance gap is that the central quantitative claim rests on unvalidated extrapolation of the deep-learning emulator to the upper tail of the surge distribution and to future storms outside the training distribution.","major_comments":[{"comment":"Same comment as above, but condensed.","section":"§2.3, SI S2.4"},{"comment":"This is a second major comment.","section":"§3.1, SI S2.5"},{"comment":"This is a third major comment.","section":"§3.2, Fig. 4, SI S2.4"}],"minor_comments":[{"comment":"Duplicate 'the'.","section":"§2.4"},{"comment":"Typo.","section":"Discussion"},{"comment":"Spelling.","section":"SI S3.2 and Fig. S10"},{"comment":"Notation clarity.","section":"§2.2, Fig. 1"},{"comment":"Sensitivity.","section":"§2.5"}],"recommendation":"major_revision","confidential_remarks":"The heavy reliance on the authors' own RAFT/DeepSurge/CA-Surge modeling chain is not inherently circular, and the authors do cite external validation data (Needham 2014; Crowell et al. 2010; Gori et al. 2022; Muis et al. 2023). However, the independent evidence for the upper tail of the surge distribution is thin, and the central projection is precisely a tail quantity. I encourage the editor to require a dedicated tail-validation analysis or a thorough sensitivity to surge-dependent bias before publication. The self-citation pattern, while notable, does not by itself warrant rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth engaging with. The combination of a point-based surge emulator and 900,000 synthetic storms is a real step up in scale, and the central claim—roughly 50% more people exposed to the 100-year surge by end of century under SSP5-8.5, population held constant—survives the model's own bias correction (+57% rather than +50%). The Georgia/South Carolina threshold result is a genuinely nonlinear finding that would not show up in surge-height-only analysis.\n\nWhat the paper does well: DeepSurge generalizes across the basin rather than being fit location-by-location, the validation and sensitivity analysis are unusually honest, and the external benchmarking against Needham, Gori, Muis, and Crowell is the right way to build confidence. The authors also flag their own limitations (Chesapeake Bay, CA-Surge attenuation, sparse gauge coverage) and release the surge fields. That is more than most papers in this space do.\n\nThe soft spots are real but not disqualifying. The most important is the missing tail validation: DeepSurge is trained on 279 storms at roughly 25 km resolution and then asked to estimate 100-year extremes for future storms about one category stronger. Aggregate R2 against gauges is modest and the mean bias is +0.22 m; the Needham spatial correlation is insignificant at 18 points. The bias-corrected sensitivity shows the mean bias does not flip the sign of the headline, but it cannot rule out surge-dependent error growth. So the exact 50% number should be read as \"roughly 50–60%\", with the spatial pattern more robust than the absolute level. Related: the abstract's \"agrees well with historical observations\" overstates what the evidence shows—the eCDF matches, but spatial correlation does not. The 90% confidence intervals also exclude emulator and ADCIRC structural error, so they understate total uncertainty. And the absence of code or trained weights limits reproduction.\n\nCitation practice is fine: RAFT is self-cited, but it is their own published framework, and the central claims are anchored to external datasets and independent modeling comparisons.\n\nBottom line: this deserves a serious referee. The main thing I would ask the authors for is a dedicated tail/extrapolation check—for example, how DeepSurge performs when trained on weaker storms and evaluated on the strongest historical surges, or a resampled extreme-subset validation. My own verdict would be conditional accept after that check, not rejection.","headline":"A serious, well-transparented coast-wide surge risk product built on a deep learning emulator; the +50% population-at-risk claim is directionally robust, but the 100-year tail is not validated well enough to take the exact number as final.","tokens_in":28575,"tokens_out":2277,"would_cite":true,"duration_ms":26245,"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":"By the end of the century, roughly 2.3 million more U.S. coastal residents would be exposed to a 100-year storm surge flood—a 50 percent increase over the historical 4.6 million—driven mainly by sea-level rise, according to a…","keywords":["storm surge","deep learning","tropical cyclone","sea level rise","flood risk","100-year return level","synthetic tropical cyclones","coastal inundation"],"falsifier":"Run the same 900,000 synthetic storms through ADCIRC for the strongest future events—say the top one percent by intensity—and compare the resulting 100-year surge maps with DeepSurge's at the 1,100 coastal nodes; if the emulator's error on these out-of-sample storms exceeds the roughly 8 cm mean signal attributed to changed hurricane behavior, the future-vs-historical risk difference and the Georgia-South Carolina threshold would move.","tokens_in":27368,"feed_emoji":"🌊","tokens_out":6854,"duration_ms":65644,"temperature":0.7,"pith_summary":"The paper tries to establish that storm surge risk along the U.S. Gulf and Atlantic coasts can be quantified from hundreds of thousands of synthetic hurricanes by replacing a slow numerical surge model with a fast deep-learning emulator. It uses that emulator to claim that, under end-of-century SSP5-8.5 conditions, the number of people exposed to the 100-year surge flood rises about 50 percent—from 4.6 million to roughly 7 million—with population held constant. The drivers split unevenly: sea-level rise accounts for about 1.9 million of the additional people at risk, and changed hurricane behavior for about 0.24 million. It further claims that Georgia and South Carolina sit near a surge-height threshold around two meters where population exposure climbs steeply, so moderate future surge increases translate into large jumps in flood risk. If right, this makes deep-learning emulation a practical route to Monte Carlo storm surge risk assessment and points to where climate adaptation could matter most.","feed_headline":"2.3 million more U.S. residents face 100-year surge by 2100","feed_subtitle":"A deep-learning surge model projects 7 million at risk by century's end, with sea-level rise driving most of the jump.","key_machinery":"The load-bearing object is DeepSurge, a neural network with 1.7 million parameters that ingests, for a single coastal node, a storm time series plus 128 by 128 pixel maps of bathymetry and land-ocean mask centered on that node, encodes the two inputs separately, combines them through an LSTM layer, and outputs the node's maximum surge for the storm. It is 'point-based,' so the same trained model predicts surge at any of 1,100 locations, learning shared physics rather than a separate model per site. The companion machinery is CA-Surge, a bathtub-style inundation model with a per-pixel overland attenuation factor that converts surge heights into flooded pixels and, with static LandScan population data, into residents at risk. The pipeline is driven by 900,000 synthetic tropical cyclones from RAFT (50,000 per CMIP6 model-period pair, 18 pairs), bias-corrected with quantile delta mapping, and the future surge distributions are superposed on probabilistic sea-level rise projections under an additive-independence assumption.","core_discovery":"The central claim is that a point-based recurrent-convolutional network called DeepSurge, trained on ADCIRC simulations of 279 historical North Atlantic storms, can predict peak storm surge at arbitrary coastal locations accurately enough to replace thousands of hydrodynamic simulations. Running DeepSurge on 900,000 synthetic tropical cyclones from the RAFT generator and combining the results with probabilistic sea-level rise projections and a bathtub-style inundation model, the study estimates the historical 100-year surge event and its end-of-century change. It reports that the ensemble-median future 100-year surge rises by an average of about 8.4 cm from altered hurricane behavior alone and by about 85 cm (maximum 170 cm) when sea-level rise is included. Population exposure to the 100-year flood increases in every coastal state, with a national total of about 2.3 million additional residents at risk, a 50 percent increase; Florida accounts for roughly one million of that total. The paper also identifies a nonlinear threshold: Georgia and South Carolina's population-at-risk curves steepen sharply near a two-meter surge height, whereas Alabama's stays nearly linear despite larger surge increases.","pith_inferences":["Because RAFT held tropical cyclone genesis frequency fixed at the historical rate, the 50 percent increase may be conservative if future conditions also raise the number of storms; allowing genesis rate to vary in the same pipeline would test this directly.","The sharp Georgia-South Carolina threshold suggests that similar state-level risk curves could be mapped nationwide to locate other coastal communities where a small surge increase would push population exposure upward abruptly.","The additive, independent treatment of surge and sea-level rise is described in the paper as conservative; including tides, waves, rainfall, and compound extremes would likely raise the estimated at-risk population, making the headline increase a lower bound.","The point-based architecture could be retrained on high-resolution regional hydrodynamic simulations and transferred to other ocean basins, which would extend this Monte Carlo risk approach globally."],"forward_implications":["The U.S. coastal population exposed to the 100-year surge flood would grow from 4.6 million to about 7 million by 2066-2100 under SSP5-8.5, a 50 percent increase with population held at historical levels.","Sea-level rise is the dominant driver: roughly 1.9 million of the additional at-risk residents come from sea-level rise alone, while changed hurricane behavior adds about 0.24 million.","Every coastal state sees an increase in population at risk; Florida alone gains about one million at-risk residents.","Georgia and South Carolina are near a two-meter surge threshold where the population-at-risk curve steepens sharply, so even modest future surge increases produce large relative jumps in exposure.","Correcting DeepSurge's mean gauge bias reduces the absolute at-risk totals by 14-18 percent but raises the relative future increase to about 57 percent, so the headline 50 percent figure is not an artifact of the bias."],"supporting_citations":[{"why":"Supplies the RAFT synthetic tropical cyclone tracks, intensities, and rainfall used to generate the 900,000 storms.","marker":"Xu et al. (2024)"},{"why":"Provides the probabilistic sea-level rise projections that are added to future surge distributions.","marker":"Kopp et al. (2014)"},{"why":"Quantile delta mapping method used to bias-correct synthetic TC intensities against historical observations.","marker":"Cannon et al. (2015)"},{"why":"Independent ADCIRC-based 100-year surge hazard estimate used to validate DeepSurge's return levels and compare future change.","marker":"Gori et al. (2022)"},{"why":"Compiled historical 100-year surge estimates along the Gulf Coast used as an observational benchmark.","marker":"Needham (2014)"},{"why":"Provides FEMA-based state-level 100-year coastal flood population estimates used to validate CA-Surge inundation results.","marker":"Crowell et al. (2010)"},{"why":"Supplies the overland water-level attenuation rates used by CA-Surge.","marker":"Vafeidis et al. (2019)"},{"why":"SSHPI parametric surge model forced by the same storms; used as a comparison and for causal decomposition of storm-characteristic effects.","marker":"Islam et al. (2021)"},{"why":"Describes ADCIRC, the hydrodynamic model whose simulations provide DeepSurge's training targets.","marker":"Luettich et al. (1992)"},{"why":"Provides the wind field formulation used to generate the ADCIRC training simulations.","marker":"Emanuel & Rotunno (2011)"}],"fun_headline_variants":["Deep learning surge model finds 50% jump in coastal at-risk population","AI model projects 2.3M more Americans in 100-year surge zone","Florida leads U.S. in future 100-year storm surge exposure","Sea-level rise accounts for most of 50% surge risk increase","AI surge model: 2.3M more coastal residents exposed by 2100"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"DeepSurge was trained on only 279 historical storms simulated on a coarse 25 km mesh with no tides and no extreme-tail validation, and it is assumed to extrapolate faithfully to end-of-century storms roughly a full Saffir-Simpson category stronger than anything it saw in training.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning surge model finds 50% jump in coastal at-risk population","AI model projects 2.3M more Americans in 100-year surge zone","Florida leads U.S. in future 100-year storm surge exposure","Sea-level rise accounts for most of 50% surge risk increase","AI surge model: 2.3M more coastal residents exposed by 2100"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2903,"prompt_tokens":935,"completion_tokens":1968,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":1879}},"tokens_in":551,"tokens_out":1968,"duration_ms":14325,"temperature":1.0,"reasoning_tokens":1879,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:26:55.592525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 900,000 synthetic storms through ADCIRC for the strongest future events—say the top one percent by intensity—and compare the resulting 100-year surge maps with DeepSurge's at the 1,100 coastal nodes; if the emulator's error on these out-of-sample storms exceeds the roughly 8 cm mean signal attributed to changed hurricane behavior, the future-vs-historical risk difference and the Georgia-South Carolina threshold would move.","supporting_citations":[{"cited_title":"E., Horton, R","cited_arxiv_id":null,"evidence_quote":"Provides the probabilistic sea-level rise projections that are added to future surge distributions."},{"cited_title":"Tropical cyclone climatology change greatly exacerbates US extreme rainfall–surge hazard","cited_arxiv_id":null,"evidence_quote":"Independent ADCIRC-based 100-year surge hazard estimate used to validate DeepSurge's return levels and compare future change."},{"cited_title":"A Data - Driven Storm Surge Analysis for the U","cited_arxiv_id":null,"evidence_quote":"Compiled historical 100-year surge estimates along the Gulf Coast used as an observational benchmark."},{"cited_title":"An Estimate of the U","cited_arxiv_id":null,"evidence_quote":"Provides FEMA-based state-level 100-year coastal flood population estimates used to validate CA-Surge inundation results."},{"cited_title":"T., Schuerch, M., Wolff, C., Spencer, T., Merkens, J","cited_arxiv_id":null,"evidence_quote":"Supplies the overland water-level attenuation rates used by CA-Surge."},{"cited_title":"A., Westerink, J","cited_arxiv_id":null,"evidence_quote":"Describes ADCIRC, the hydrodynamic model whose simulations provide DeepSurge's training targets."},{"cited_title":"and Rotunno, R","cited_arxiv_id":null,"evidence_quote":"Provides the wind field formulation used to generate the ADCIRC training simulations."}],"review_version":1}