{"id":"69950200-3f61-45e2-9188-979dfa8d4eab","arxiv_id":"2505.04918","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PASSAT combines spherical graph neural networks with analytic advection and Navier-Stokes solvers and reports the best RMSE and ACC among parameter-matched deep learning baselines and IFS T42 on 5.625° ERA5 forecasts.","lead":"PASSAT is a weather forecasting model that couples the classical advection and Navier-Stokes equations on a spherical Earth with a spherical graph neural network, and it reports better scores than several deep learning baselines and the IFS T42 numerical model on coarse-resolution ERA5 data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Own Tables 8 and 9 contradict the Section 4 claim that PASSAT outperforms every deep learning baseline at every lead time: GraphCast and FourCastNet win several t2m cells.","rationale":"The reader's weakest-assumption pick was that per-variable learned velocity fields for scalar variables such as t2m and z500 are not validated against physical winds, making the Navier-Stokes component an architectural prior rather than a physical law. That is a legitimate interpretive concern about the physics contribution, but it does not directly falsify the headline empirical result; even a purely architectural prior could improve accuracy. The more load-bearing concern is internal: the paper's own Tables 8 and 9 contradict the unqualified dominance claim in Section 4 and the abstract. A reviewer can verify this by checking the reported t2m entries against the surrounding text. This matters because the abstract and Section 4 stake the paper's value on PASSAT being strictly better than the deep learning baselines. Once the comparison is restricted to 'competitive or better on most metrics', the contribution changes from a state-of-the-art result to a hybrid architecture with mixed performance. The reader's CONDITIONAL verdict remains appropriate, but the acceptance conditions should explicitly require correcting the dominance claim and reporting per-cell uncertainty or significance. I did not find reason to move beyond the reader's CONDITIONAL verdict because the method, ablations, and the IFS T42 comparison still provide useful evidence; the overclaim is fixable rather than fatal to the underlying approach.","tokens_in":18870,"tokens_out":8798,"duration_ms":86467,"concrete_test":"Reproduce the t2m rows of Tables 8 and 9 at lead times 24, 48, and 120 h by running the released checkpoint on the 2017-2018 ERA5 5.625-degree test set and computing latitude-weighted RMSE and ACC exactly as specified in Equations 17-19 in Appendix C. If the entries reproduce, PASSAT is not strictly better than GraphCast at 24/48 h or FourCastNet at 120 h ACC, and the Section 4 dominance sentence must be revised. If the entries do not reproduce and PASSAT becomes strictly best, then the manuscript's tables are the source of the error and must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 states: 'PASSAT outperforms the other deep learning models in all weather variables across different lead times.' Tables 8 and 9 directly contradict this strict-dominance version of the central claim. For t2m at 24 h, Table 8 reports RMSE 1.25 for PASSAT versus 1.21 for GraphCast; at 48 h, PASSAT is 1.54 versus GraphCast's 1.52. Table 9 reports t2m ACC 0.970 versus 0.972 at 24 h and 0.955 versus 0.956 at 48 h; at 120 h FourCastNet's ACC 0.888 exceeds PASSAT's 0.886. t850 at 24 h is tied on both RMSE and ACC. Thus PASSAT is not the best deep learning model in every variable and lead time even in the paper's own reported results. The abstract's claim that PASSAT 'outperforms the state-of-the-art deep learning-based weather prediction models' is therefore an overstatement; the defensible claim is that PASSAT improves on most variables and lead times, especially z500, u10, v10 and longer leads, and beats IFS T42 at the reported 72 and 120 h cells. This is an internal inconsistency between the text and the reported tables, not a disagreement with external consensus, so it bears directly on the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PASSAT, a deep learning weather-prediction model that combines a spherical graph neural network with numerical solves of the advection equation and the Navier-Stokes equations on the sphere. The GNN estimates the Earth-atmosphere interaction tendency and the initial velocity fields, while the PDE solver advances the variables and their velocities in time. Experiments are conducted on the 5.625-degree ERA5 dataset for t2m, t850, z500, u10, and v10, with lead times up to 144 hours. The paper claims that PASSAT outperforms both state-of-the-art deep learning weather models and the operational IFS T42 model, and ablation studies are used to argue that both physics and topology information contribute to the gains. The appendices provide architecture details, training procedures, modifications of the baseline models, and full comparison tables.","tokens_in":19179,"tokens_out":5331,"duration_ms":53519,"significance":"If the central claims were fully supported, the paper would be a useful contribution to physics-assisted, topology-aware data-driven weather forecasting. The most valuable parts are the spherical treatment of the advection equation, the attempt to couple learned interaction terms with PDE-based advection, and the availability of code and checkpoints. The ablation results are internally consistent and show clear gains over the authors' own variants. However, the headline claim of outperforming state-of-the-art deep learning models is not supported as written, because the baseline models were downscaled to roughly 1.15 million parameters before comparison, and because the paper's own tables contain counterexamples to the strict-dominance statement. The comparison with IFS T42 is also limited to three variables and two lead times. With recalibrated claims and additional validation of the learned velocity fields, the core idea remains worth publishing.","major_comments":[{"comment":"The statement in Section 4 that 'PASSAT outperforms the other deep learning models in all weather variables across different lead times' is directly contradicted by the paper's own results in Tables 8 and 9. For t2m, GraphCast has lower RMSE than PASSAT at 24 h (1.21 vs 1.25), 48 h (1.52 vs 1.54), and 72 h (1.84 vs 1.85), while FourCastNet has lower RMSE at 144 h (2.68 vs 2.69) and ties at 96 h and 120 h. On ACC, GraphCast leads t2m at 24 h (0.972 vs 0.970), 48 h (0.956 vs 0.955), and 72 h (0.937 vs 0.935), and FourCastNet leads at 120 h (0.888 vs 0.886) and 144 h (0.865 vs 0.862). The t850 result at 24 h is tied on both RMSE and ACC. The abstract and the conclusion should be revised to state that PASSAT improves on most variables and lead times, especially z500, u10, v10, and longer leads, rather than claiming superiority in every cell.","section":"Section 4 and Appendix F, Tables 8-9"},{"comment":"The comparison with 'state-of-the-art' deep learning models is not a comparison with the published models themselves. GraphCast, Pangu, FourCastNet, and SFNO were reduced from 36.7M, 64M, 59.1M, and 107M parameters, respectively, to about 1.15M parameters and retrained from scratch at 5.625 degrees. A 1.15M-parameter GraphCast is not the GraphCast from the Science paper, and the same holds for the other baselines. This makes the controlled comparison legitimate but does not support the abstract's claim that PASSAT 'outperforms the state-of-the-art deep learning-based weather prediction models.' The claim should be explicitly limited to parameter-matched reimplementations trained on this dataset, or the authors should add comparisons with the original pretrained models at their native parameter counts.","section":"Section 4, 'Baseline deep learning models'; Appendix E, Tables 3-7"},{"comment":"The paper does not provide evidence that the learned per-weather-variable velocity fields v(t,s) correspond to physical atmospheric winds. Each variable t2m, t850, z500, u10, and v10 is assigned its own velocity field, and the Navier-Stokes equations (7)-(8) are then solved for these fields. For variables such as t2m and z500, there is no separately measured physical velocity in the dataset, so these fields are learned latents whose physical interpretation is unclear. If the learned velocities do not resemble real winds, then equations (7)-(8) act as a learned regularizer rather than as a physical law, and the claim that PASSAT 'seamlessly integrates the physics of weather evolution' is overstated. A concrete test would be to compare the velocity-branch output for u10 and v10 against ERA5 wind fields, or to couple all advected variables to a single shared wind field. The authors should either add such validation or substantially soften the physics-language claims.","section":"Section 3.5, Section 3.7, and Algorithm 1"}],"minor_comments":[{"comment":"The viscosity coefficient mu appears in equations (7) and (8) and in Algorithm 1 but its value is never reported in the paper. Since this coefficient controls the strength of the friction term, the authors should state its value and, ideally, report sensitivity to it.","section":"Section 3.5"},{"comment":"There is a typo in 'GeForece RTX 2080'; it should read 'GeForce RTX 2080.' Additional typos include 'Naiver-Stokes' in Section 5 and 'planner' in Figure 2's caption, and 'Pacth Size' in Table 4.","section":"Appendix D.2"},{"comment":"The ClimODE rows show ACC values that collapse to 0.510 at 72 h and 0.020 at 96 h before becoming NAN at 120 h and 144 h. If this reflects numerical instability or a training failure, it should be stated explicitly so that readers do not interpret the missing entries as an unqualified failure of the baseline.","section":"Tables 8 and 9"},{"comment":"The comparison with IFS T42 is limited to t2m, t850, and z500 at 72 h and 120 h, as the paper notes. The abstract's claim that PASSAT outperforms the operational NWP model IFS T42 should be qualified accordingly, since no claim is made for u10, v10, or other lead times.","section":"Section 4, comparison with IFS T42"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: PASSAT combines spherical advection and a Navier-Stokes update on the sphere with a spherical graph network. That combination is new relative to ClimODE and NowcastNet, and the paper presents it cleanly. The experiments are more complete than usual for this area: five variables, six lead times, five retrained baselines, ablation studies for topology and physics, and a working code link. The ablation story is coherent, and the strongest gains are on z500, u10, v10, and longer lead times, which is where the physics prior should help.\n\nThe soft spots are real but not fatal. First, the abstract and Section 4 claim PASSAT outperforms all deep learning baselines in all weather variables across lead times. Their own Table 8 shows GraphCast beating PASSAT on t2m at 24h and 48h by small margins (RMSE 1.21 vs 1.25; 1.52 vs 1.54), and Table 9 shows FourCastNet ahead on t2m ACC at 120h. So strict dominance is false; the defensible claim is that PASSAT improves most variables and leads, especially at longer range, and beats IFS T42 at the reported 72h and 120h cells. That is still a useful result, but smaller, and the text should say it.\n\nSecond, the baseline comparison is weaker than it looks. The original GraphCast, Pangu, FourCastNet, and SFNO have tens to hundreds of millions of parameters; the authors shrink all of them to around 1.15M and retrain from scratch. That is not an unreasonable way to compare architectures, but calling those compressed versions \"state-of-the-art\" overstates them. The fair summary is that PASSAT beats comparably sized versions of those architectures, not the operational-scale models.\n\nThird, the Navier-Stokes viscosity mu is never given. It appears as a constant in Eqs. (7)-(8) and in Algorithm 1. A fixed value or fitted value would both be acceptable; an unstated value is missing information for a paper whose hook is physics.\n\nThe weakest conceptual point is not the math but the interpretation. Velocity fields are learned separately for each weather variable, and the paper never checks whether these fields resemble physical winds. If they are merely reparameterizations, the Navier-Stokes block is functioning as a smoothness regularizer rather than a physical law. That weakens the \"seamlessly integrates physics\" claim, though it does not invalidate the empirical result. A simple diagnostic comparing the learned velocities to ERA5 winds would settle it.\n\nBottom line: the paper deserves a serious referee. The central claim needs rewriting, mu must be reported, and a physical plausibility check of the velocities would meaningfully strengthen the argument. I would send it to review with those conditions.","headline":"PASSAT is a genuinely new hybrid architecture with honest, extensive experiments, but the paper's headline claim of outperforming all deep learning baselines is contradicted by its own tables.","tokens_in":19715,"tokens_out":1732,"would_cite":true,"duration_ms":17087,"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":"Spherical physics-AI weather model beats operational IFS T42","keywords":["weather prediction","physics-informed deep learning","spherical graph neural network","advection equation","Navier-Stokes equations","topology-aware forecasting","ERA5","medium-range forecasting"],"falsifier":"Ablate the Navier-Stokes velocity update by replacing it with a neural-network velocity updater of matched capacity, or simple persistence of velocities, while keeping the advection solver and spherical graph fixed; if RMSE does not degrade, the Navier-Stokes equations are not load-bearing. A second check would correlate PASSAT's learned initial velocity fields for t2m and z500 with ERA5 wind components; near-zero correlation would indicate the physics is a learned latent, not atmospheric advection.","tokens_in":18637,"feed_emoji":"🌦️","tokens_out":7644,"duration_ms":68386,"temperature":0.7,"pith_summary":"PASSAT is a deep learning weather model that builds the two recognized drivers of atmospheric evolution, advection and the Earth-atmosphere interaction, directly into its architecture, and solves the governing equations on the sphere rather than on a flat latitude-longitude grid. The paper claims that by numerically integrating the advection equation for each weather variable and the Navier-Stokes equations for the associated velocity fields, while a spherical graph neural network supplies the interaction tendency and initial velocities, the model outperforms both state-of-the-art deep learning baselines and the operational numerical model IFS T42 at coarse 5.625-degree resolution. The central claim matters because it suggests physics constraints and spherical geometry can make data-driven forecasts more accurate and robust than either black-box deep learning or coarse operational NWP.","feed_headline":"Spherical physics-AI weather model beats operational IFS T42","feed_subtitle":"On coarse 5.625° ERA5 data it also tops parameter-matched deep learning baselines through 144-hour leads.","key_machinery":"The central machinery is the coupling of a spherical advection-Navier-Stokes solver with a two-branch spherical graph neural network. For each weather variable $u$, the advection equation gives the advective tendency $- (v_\\theta \\partial u/\\partial\\theta + \\frac{v_\\varphi}{\\cos\\theta}\\partial u/\\partial\\varphi)$, and the spherical Navier-Stokes equations (7)-(8) update the velocity components $v_\\theta$ and $v_\\varphi$ under advection, curvature, Coriolis, pressure-gradient (replaced by the z500 geopotential gradient) and viscous terms. The spherical graph neural network supplies the initial velocity fields and the hourly interaction tendency. Solving on the sphere rather than the planar latitude-longitude grid eliminates the $\\sec\\theta$ amplification that distorts velocities near the poles, and the 0.2-hour Euler time stepping with an hourly-constant interaction term keeps the autoregressive rollout stable without backpropagating through every sub-step.","core_discovery":"PASSAT decomposes the total tendency of any weather variable into an advective part and an interaction part. The advective part is computed by solving the advection equation on the unit sphere, with future velocity fields obtained from the Navier-Stokes equations on the sphere, closed by replacing the pressure-gradient force with the gradient of the 500 hPa geopotential. The interaction part, radiation, clouds, and subgrid turbulence, is estimated by a spherical graph neural network whose interaction branch runs once per hour, while its velocity branch generates the initial velocity fields. All equations are integrated with a 0.2-hour Euler step, and the graph is built with nodes at the grid points on the sphere using a Haversine/Gaussian kernel. The paper's central discovery is that this physics-assisted, topology-informed architecture achieves the lowest root-mean-square error and highest anomaly correlation among the five compared deep learning models (GraphCast, ClimODE, Pangu, FourCastNet, SFNO) for all five variables, at all lead times up to 144 hours, and also beats IFS T42 at the two reported NWP lead times of 72 and 120 hours. Ablation studies show that removing either the physics equations or the spherical topology degrades accuracy, with most variables benefiting more from the topology correction.","pith_inferences":["Inference: the learned velocity fields for scalar variables such as t2m and z500 are probably not physical winds; if the observed wind fields were used inside the advection equation instead, the results would reveal whether the Navier-Stokes update is genuinely physical or merely a structured smoothing prior.","Inference: the topology benefit likely comes mostly from the $1/\\cos\\theta$ metric factor in the spherical advective derivative, which prevents the polar amplification of planar velocities; this could be tested by keeping the spherical graph but switching the advection derivative back to the planar form.","Inference: because the machinery is variable-agnostic, the same architecture could be applied to other advection-dominated spherical fields, such as ocean surface tracers or atmospheric chemical concentrations, provided interaction terms can be learned.","Inference: the comparison with IFS T42 is limited to two lead times and one coarse resolution; a natural extension would test PASSAT at higher input resolution and against IFS T63-equivalent skill, where the paper reports the gap closes."],"forward_implications":["PASSAT achieves the best RMSE and ACC among the five parameter-matched deep learning baselines for all five variables (t2m, t850, z500, u10, v10) at every lead time from 6 to 144 hours.","It outperforms the operational IFS T42 model at the two reported lead times (72 and 120 hours) for t2m, t850, and z500, despite IFS T42 running at a finer 2.8-degree resolution.","Ablation results show that both the physics equations and the spherical topology contribute to the gain; for t2m the two contributions are roughly equal, while for t850, z500, u10, and v10 the topology information is the larger contributor.","The 0.2-hour integration step with hourly interaction updates is sufficient for stable 144-hour forecasts, suggesting the Euler scheme is not a practical bottleneck at this resolution."],"supporting_citations":[{"why":"Supplies the spherical formulation of the advection and primitive equations used in Eqs. (5), (7), and (8).","marker":"[LTW92]"},{"why":"Gives the linear advection equation that defines the advective tendency in Eq. (4).","marker":"[Cha22]"},{"why":"Provides the Navier-Stokes theory that justifies evolving the velocity fields via Eqs. (7)-(8).","marker":"[Tem84]"},{"why":"ClimODE is the closest physics-assisted baseline, and the paper contrasts its continuity-equation and learned-velocity approach with PASSAT's advection and Navier-Stokes approach.","marker":"[VHG24]"},{"why":"GraphCast is the topology-informed deep learning baseline that PASSAT most closely rivals and must outperform.","marker":"[Lam+23]"},{"why":"WeatherBench supplies the 5.625-degree ERA5 dataset and the IFS T42/T63 RMSE and ACC reference values used for the NWP comparison.","marker":"[Ras+20]"},{"why":"Provides the ERA5 reanalysis data that the benchmark is built on.","marker":"[Her+20]"},{"why":"SFNO is the spherical-Fourier baseline whose topology-informed design is compared and outperformed.","marker":"[Bon+23]"}],"fun_headline_variants":["Physics + topology boost AI weather model past IFS T42","PASSAT: spherical deep learning beats IFS T42 in forecast","Topology-aware physics AI outperforms operational NWP","AI weather model with sphere physics dethrones IFS T42","Graph neural net solves sphere equations to beat IFS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned velocity fields that drive advection behave enough like real atmospheric winds that the Navier-Stokes equations genuinely govern their evolution, rather than the equations merely acting as a smoothness regularizer on learned latent quantities.","fun_headline_variants_meta":{"raw":{"variants":["Physics + topology boost AI weather model past IFS T42","PASSAT: spherical deep learning beats IFS T42 in forecast","Topology-aware physics AI outperforms operational NWP","AI weather model with sphere physics dethrones IFS T42","Graph neural net solves sphere equations to beat IFS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1361,"prompt_tokens":1054,"completion_tokens":307,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":222}},"tokens_in":670,"tokens_out":307,"duration_ms":3146,"temperature":1.0,"reasoning_tokens":222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:18:21.730085+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablate the Navier-Stokes velocity update by replacing it with a neural-network velocity updater of matched capacity, or simple persistence of velocities, while keeping the advection solver and spherical graph fixed; if RMSE does not degrade, the Navier-Stokes equations are not load-bearing. A second check would correlate PASSAT's learned initial velocity fields for t2m and z500 with ERA5 wind components; near-zero correlation would indicate the physics is a learned latent, not atmospheric advection.","supporting_citations":[],"review_version":1}