{"id":"0f647411-10ad-4d0f-8aee-4a37dcc9c0bf","arxiv_id":"2507.17298","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"This paper presents a multi-modal encoder-decoder network that forecasts daily averaged solar wind speed at L1 up to four days ahead, with validation RMSE around 55 to 58 km/s and a four-day-ahead test RMSE of 53 km/s.","lead":"A neural network that combines historical solar wind readings with four days of solar EUV images forecasts solar wind speed at Earth one to four days ahead. The model reports lower errors than earlier machine-learning forecasters and is tested on data from a later solar cycle it never saw during training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Daily averaging of the target makes the claimed 20% RMSE improvement over Brown et al. (2022) not apples-to-apples; the RMSE gap may be a smoothing artifact rather than forecasting skill.","rationale":"Read in good faith, the paper's core contribution is a two-stage encoder-decoder that uses 27/54-day solar-wind history and four days of AIA 211 images, with a genuinely held-out 2019-2024 test set and consistent improvement over the 27-day persistence baseline. Those are real strengths. The central claim emphasized in the abstract, however, is the 'much improved RMSE' relative to Upendran et al. (2020) and Brown et al. (2022), and that claim rests on Table 3. Its weakest load-bearing premise is target-scale comparability. Brown et al. and Upendran et al. optimize and evaluate on hourly averaged solar wind speed; this model predicts daily averages. Because RMSE is a second-moment statistic of the error between forecast and target, changing the target from hourly to daily averages removes high-frequency variance and lowers the achievable RMSE floor. The day-4 correlation is identical to Brown et al. (0.63), which is what one would expect if the RMSE advantage came mostly from smoothing. The paper acknowledges different validation splits in Sec. 3.2, but the abstract's claim is unqualified. A concrete check—computing how much daily averaging lowers the persistence baseline or a prior model's RMSE on the same window—would settle whether the 20% figure is skill or artifact. The secondary issues (underspecified decoder handling at inference, no released code, no uncertainty estimates) are real but do not change the verdict; they affect reproducibility rather than the logical status of the comparative claim. The appropriate verdict remains CONDITIONAL: the model results are plausible and the held-out test is commendable, but the headline comparison needs either a corrected apples-to-apples evaluation or a qualified wording before it should be accepted at face value.","tokens_in":19190,"tokens_out":6428,"duration_ms":66408,"concrete_test":"Compute the four-day-ahead 27-day persistence baseline (Owens et al. 2013) on the paper's exact validation window (2014-2017) for two target definitions: hourly-averaged solar wind speed and daily-averaged solar wind speed. If daily averaging lowers the baseline RMSE by roughly 20% (from ~72 to ~58 km/s), the same reduction applies to any hourly RMSE, and the Table 3 improvement over Brown et al. (2022) is explained by target smoothing rather than by model skill. Independently, if the authors can rerun Brown et al.'s architecture on daily averages—or if Brown et al.'s hourly predictions are publicly available, daily-average them and recompute RMSE—the comparison becomes apples-to-apples; a persistent ~20% gap after matching cadence would validate the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim—'much improved RMSE values compared to the previous works' (Abstract; echoed in Sec. 3.2 and Sec. 4)—is built on Table 3, which compares this model's daily-averaged validation RMSE (57.6 km/s at day 4) with hourly-averaged RMSEs from Upendran et al. (2020) and Brown et al. (2022). RMSE is measured on the target series, so smoothing the target by taking daily averages removes sub-day variance and mechanically lowers RMSE; a model with identical information can report a smaller RMSE on daily averages than on hourly averages. The equal day-4 correlation (r=0.63 in both this work and Brown et al.) is consistent with this: the RMSE gap is not accompanied by a correlation gain. The paper's Sec. 2.2.2 motivates daily averaging by arguing that 'within a day, the solar wind speed usually does not show significant variability,' but Table 3 treats the two error scales as directly comparable. Sec. 3.2 concedes the comparison is 'only indicative' due to different validation splits, but the abstract and Discussion present the 20% improvement as established. The validation-window difference (five-fold splits across cycle 24 versus a single balanced 2014-2017 split) is a second confound. Until the RMSE reduction is shown to survive target-cadence matching, the headline comparative claim is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a multi-modal encoder-decoder neural network that forecasts daily-averaged solar wind speed at L1 for four days ahead. The model combines two input modalities: a history encoder processing solar wind measurements from approximately 27 and 54 days before the current time, and an AIA 211 Å encoder processing EUV image representations from the four preceding days. A decoder, conditioned on the recent four days of solar wind speed with future values masked, outputs speeds from four days in the past to four days in the future. The authors train in two stages: first with solar wind history alone from 1996 onward, then with the AIA encoder added, using a validation split from solar cycle 24 that includes both active and quiet phases. They report validation RMSEs of 54.7, 58.3, 58.2, and 57.6 km/s with correlations of 0.78, 0.66, 0.64, and 0.63 for days 1 to 4, and test-period (2019-2024) RMSEs of 46.9, 50.9, 52.5, and 53.3 km/s. The paper's headline claim is that the model achieves 'much improved RMSE values compared to the previous works,' specifically about a 20% RMSE improvement over Brown et al. (2022) for the four-day-ahead forecast.","tokens_in":19524,"tokens_out":2359,"duration_ms":25873,"significance":"If the stated results hold, the paper would be a useful contribution to solar wind forecasting: it demonstrates a two-stage transfer-learning approach that leverages long solar-rotation-scale autocorrelations, incorporates EUV imagery, and evaluates on a genuinely held-out period from solar cycle 25 that was not used in training. The validation-set design that deliberately balances active and quiet phases is a thoughtful improvement over naive random splits, and the time-series and heatmap diagnostics give a concrete picture of where the model succeeds and fails. The fast inference time is also a practical strength for operational use. However, the significance of the central comparative claim is currently undermined by the inconsistent error metrics between this work and the prior works it compares against, and by the differing validation windows, as detailed below.","major_comments":[{"comment":"The central comparative claim, that the model yields 'much improved RMSE values compared to the previous works,' is not yet supported because Table 3 compares this model's daily-averaged RMSEs with hourly-averaged RMSEs from Upendran et al. (2020) and Brown et al. (2022). RMSE is computed on the target variable, and daily averaging removes sub-day variability, which mechanically tends to lower RMSE even if the underlying forecasting skill is unchanged. The fact that the day-4 correlation is identical (r=0.63) in this work and in Brown et al. (2022) is consistent with the RMSE gap being partly a smoothing artifact. The authors should recompute their results on hourly averaged output, or compare against daily averaged baselines and daily averaged versions of the previous models, and report both cadences. Without such a matched comparison, the 20% improvement in Section 4 and the abstract should be removed or explicitly restated as not directly comparable.","section":"§3.2, Table 3; Abstract"},{"comment":"The comparison with prior work is also confounded by the validation-window difference. Upendran et al. (2020) and Brown et al. (2022) use five-fold cross-validation splits covering different phases of solar cycle 24, whereas the current model uses a single validation period from January 2014 to February 2017 selected to balance active and quiet phases. Because solar wind predictability, persistence correlations, and RMSE magnitudes all vary strongly with solar cycle phase, as the authors themselves show in Figure 4, the claimed RMSE improvement may reflect the choice of validation window rather than model skill. The manuscript acknowledges in Section 3.2 that the comparison is 'only indicative,' but the abstract and Section 4 present the improvement as established. A matched-period evaluation, or at minimum a clear statement in the abstract that the comparison is not apples-to-apples, is required.","section":"§3.2, Table 3; §2.5.1"},{"comment":"The day-4 validation RMSE of 57.60 km/s in this work is compared with Brown et al.'s 72.21 km/s, but the two numbers are not on the same target cadence. The paper justifies daily averaging by stating that 'within a day, the solar wind speed usually does not show significant variability,' yet if that is true, the daily-versus-hourly distinction should have a small effect and the authors should be able to demonstrate this quantitatively. As written, the statement in Section 3.2 that the RMSE is 'significantly lower' by 20% is not established. A simple test, such as computing the RMSE of the daily-averaged version of the Brown et al. predictions or of an hourly persistence model both before and after averaging, would clarify whether the improvement is real forecasting skill or an artifact of the averaging window.","section":"§2.2.2, §3.2"}],"minor_comments":[{"comment":"The Stage 1 day-3 and day-4 validation RMSEs are both listed as 63.25 km/s; this may be a typographical error and should be checked against the original outputs.","section":"Table 2"},{"comment":"The text says the loss weighting and other hyperparameters were 'fine-tuned to obtain optimal predictions,' but exact values for learning rate, batch size, dropout, weight decay, and the loss-scaling factor (beyond noting the factor of 100) are not reported. Providing these values, or a link to code, would materially improve reproducibility.","section":"§2.5.3"},{"comment":"The paper refers to 'Appendix 2' in Section 2.5.3, but the manuscript as provided has no appendices; the reference should be fixed or the relevant distribution plot included.","section":"§2.5.4 and §3.3"},{"comment":"The caption states that the model yields 'significantly lower RMSE values compared to the 27-day persistence baseline,' but the text in Section 3.2 says the r values are only 'approximately equal' for Stage 1; the wording should be made consistent across the figure caption and main text.","section":"Figure 4"},{"comment":"The timestamp encoding uses 'Y ear: (Year-1996)/50.0' and a fixed solar-cycle-number assignment; the arbitrary choice of 50 years and 11 years for cycle 25 should be flagged in the text as a potential limitation if the model is used beyond cycle 25, since the normalization and solar cycle phase would become inaccurate.","section":"§2.2.1"}],"recommendation":"major_revision","confidential_remarks":"The core modeling contribution is solid and the held-out test is a real strength, but the abstract's headline comparative claim currently rests on a cadence-mismatched RMSE comparison. In my view this is fixable within the scope of the manuscript by recomputing or clearly qualifying the comparison; it is not a reason to reject. I would also encourage the editor to ask the authors to make the training hyperparameters and, if possible, the code available, since the field would benefit from reproducibility here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a sensible engineering paper with a genuine held-out test set (2019–2024), a real improvement over 27-day persistence, and an architecture that is new for this problem: two history encoders on the 27/54-day windows plus an AIA 211 encoder, trained in two stages to handle the missing AIA mode before 2010. The decoder masks future values, so there is no circularity. The loss weighting is a tuned free parameter, but that is normal for ML. The paper is honest enough to call the prior-work comparison “only indicative” in Section 3.2.\n\nThe soft spot is the abstract and Discussion. The headline claim—“much improved RMSE” of about 20% over Brown et al.—compares daily averaged RMSE against hourly averaged RMSE. Daily averaging removes short-timescale variance and mechanically lowers RMSE. The fact that the day-4 correlation ties Brown et al. at 0.63 while RMSE drops from 72 to 58 km/s is exactly what you would expect from smoothing the target, not from forecasting skill. The validation-window difference (a balanced 2014–2017 split versus five-fold splits across cycle 24) is a second confound that the paper acknowledges, but the abstract does not. To make the claim stick, they need to evaluate on hourly averages or report the daily-averaged RMSE of the prior models on the same window.\n\nAlso underspecified: decoder inference-time handling when future values are masked is not fully described, no code is released, and there are no uncertainty estimates. Those are minor-to-moderate, not fatal. The persistence-baseline comparison, the unseen test period, and the phase-balanced validation split are real strengths.\n\nIf the comparison is fixed, the paper’s own result—consistent gains over 27-day persistence across solar cycle phases with a truly unseen test set—is useful for space weather operators. I would send it to peer review, with a request for cadence-matched comparison and code/data release before acceptance. The architecture and the held-out test are solid enough to be worth the referee time.","headline":"Solid held-out forecasting result, but the headline RMSE gain over prior work is likely an artifact of daily versus hourly averaging; the architecture is still worth engaging.","tokens_in":20088,"tokens_out":1991,"would_cite":false,"duration_ms":19916,"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":"An encoder-decoder network that pairs prior-rotation solar wind history with EUV coronal images forecasts daily solar wind speed four days ahead with validation RMSEs near 55 km/s and an unseen-data RMSE of 53 km/s.","keywords":["solar wind forecasting","encoder-decoder network","SDO/AIA 211 Angstrom imagery","27-day persistence","coronal holes","space weather","sequence-to-sequence model","L1 in-situ measurements"],"falsifier":"Evaluate the same encoder-decoder architecture on the hourly-averaged solar wind data and the exact five-fold validation splits of Brown et al. (2022) and Upendran et al. (2020), reporting RMSE and Pearson r for days 1 through 4 at the same hourly cadence; if the day-4 RMSE does not fall below roughly 72 km/s for Brown et al. and 81 km/s for Upendran et al. on those splits, the claimed ~20% and ~29% improvements would not be confirmed.","tokens_in":19001,"feed_emoji":"☀️","tokens_out":7371,"duration_ms":66558,"temperature":0.7,"pith_summary":"The paper proposes that the solar-wind speed at L1, measured as a daily average up to four days ahead, can be forecast simultaneously by a two-stream encoder-decoder neural network that learns to 'time-translate' wind patterns from earlier solar rotations and to read present coronal structure from EUV images. The authors train the network in two stages: first on solar-wind history alone (solar wind state, flare magnitudes, sunspot number from the two prior rotations), then with four days of SDO/AIA 211 Angstrom coronal images added. On the validation period 2014-2017 they report RMSEs of 54.7, 58.3, 58.2 and 57.6 km/s and correlations of 0.78, 0.66, 0.64 and 0.63 for one to four days ahead; on unseen data from 2019-2024 the four-day-ahead RMSE is 53.25 km/s with correlation 0.55. They interpret these scores as roughly 20% lower RMSE than the model of Brown et al. (2022), with the caveat that the comparison is only indicative because forecast horizons and validation splits differ. A reliable four-day ambient solar-wind forecast matters because this window is the main lead time for preparing for space-weather effects on satellites and power grids.","feed_headline":"Two-stream network cuts 4-day solar-wind forecast error to ~53 km/s","feed_subtitle":"History from prior solar rotations plus 211 Angstrom coronal images yields a day-4 RMSE of 53 km/s on unseen 2019-2024 data.","key_machinery":"The central object is the multi-modal encoder-decoder network built on a sequence-to-sequence framework: two identical history encoders process 1D time series of seven solar-wind and solar-activity features sampled over eight days around the prior solar rotations; an AIA encoder processes GoogleNet representations of four days of 211 Angstrom EUV images; and a decoder, seeded with the past four days of wind speed with future entries masked to zero, produces a time sequence reaching four days into the future. Hand-encoded timestamps (hour, day, month, year, relative day, solar-cycle phase and cycle number) plus positional encoding give the network temporal and solar-cycle context. The two-stage training is the mechanism that makes the multimodal setup trainable from only roughly 50k samples: Stage 1 learns the history-to-wind translation, and Stage 2 freezes those weights initially, adds the AIA encoder, and refines the whole model.","core_discovery":"The central claim is that combining two kinds of solar observations in one sequence-to-sequence architecture beats prior single-modality deep-learning forecasts of ambient solar wind speed. The two inputs are the past two solar rotations of wind history (eight-day windows centered at days 27 and 54 before the current time, each with solar wind speed, proton density, three magnetic-field components, flare magnitudes and sunspot number) and the most recent four days of SDO/AIA 211 Angstrom EUV images, condensed through a pre-trained GoogleNet. The decoder receives the past four days of wind speed with the future values masked and outputs sixteen values spanning the past four days to the next four days at 12-hour cadence; the trained model is used for the four future days. After Stage 2 training with the EUV encoder, the model gives validation RMSEs of 54.67, 58.30, 58.17 and 57.60 km/s with correlations of 0.78, 0.66, 0.64 and 0.63 for days 1 to 4 ahead, and on the unseen 2019-2024 test period a day-4 RMSE of 53.25 km/s with r = 0.55. The authors interpret the RMSE reduction relative to Brown et al. (2022) as roughly 20%, while noting that the comparison is only indicative, and they show the model consistently beats the 27-day persistence baseline across solar-cycle phases.","pith_inferences":["An obvious extension the paper leaves implicit is to apply the same two-stage encoder-decoder to the full solar-wind vector (density, temperature, magnetic field) or to produce probabilistic forecasts via ensembles, since the decoder already outputs a full future sequence.","The 20% RMSE improvement claim could be tested directly by re-running the model with an output head that emits hourly values and evaluating on the five-fold splits used by the comparison papers; such a benchmark would separate the architectural gain from the benefit of coarser daily averaging.","Because the model relies on the 27-day and 54-day autocorrelation from co-rotating interaction regions, its advantage should degrade during solar maximum; the paper's own heatmaps suggest adding flare, CME, or SEP indicators, or additional EUV channels such as 193 Angstrom, would target exactly the regime where it currently fails.","The masked-decoder formulation also reconstructs the past four days, which provides a free internal consistency check; a future model could exploit reconstruction error as an anomaly signal for unusual solar-wind events."],"forward_implications":["A four-day-ahead daily-averaged solar-wind forecast is available in about 10 ms per sample once trained, which is fast enough for operational monitoring.","On unseen data from cycle 25 (2019-2024), the day-4 forecast RMSE is 53.25 km/s, so the model generalizes beyond the cycle-24 training and validation window.","Adding the EUV image encoder improves days-2-to-4 forecasts over history-only training, particularly during the declining and quiet phases where coronal holes drive high-speed streams.","The model under-predicts extreme wind speeds above about 600 km/s, which occur mostly during high solar activity, so its practical value is strongest for ambient background wind rather than storm-time extremes.","If kept updated, this architecture can serve as a data-driven complement to physics-based WSA-ENLIL forecasts, which are slower, more costly, and have larger reported errors."],"supporting_citations":[{"why":"Supplies the AIA-211-image deep-learning baseline and shows GoogleNet features capture coronal holes; this work compares against its r and RMSE values.","marker":"Upendran et al. (2020)"},{"why":"The main comparison baseline, combining EUV images with the 27-day-prior wind speed; the claimed ~20% RMSE improvement is measured against its four-day forecast.","marker":"Brown et al. (2022)"},{"why":"Establishes the 27-day persistence correlation and one-day persistence baselines and motivates the history-encoder design.","marker":"Owens et al. (2013)"},{"why":"Benchmarks physics-based operational models at roughly 57% correlation and ~100 km/s error, providing the context for improvement.","marker":"Jian et al. (2015)"},{"why":"Provides the SDOML standardized SDO/AIA 211 Angstrom image dataset used as model input.","marker":"Galvez et al. (2019)"},{"why":"Provides the pre-trained GoogleNet inception architecture that extracts the 1000-dimensional EUV image representations.","marker":"Szegedy et al. (2015)"},{"why":"Shows that combining PFSS coronal magnetic field attributes with the 27-day-prior wind speed improves machine-learning forecasts, motivating the multimodal input.","marker":"Yang et al. (2018)"},{"why":"Supplies the encoder-decoder sequence-to-sequence formulation that the architecture is based on.","marker":"Sutskever et al. (2014)"}],"fun_headline_variants":["Multimodal net forecasts solar wind 4 days ahead to 53 km/s","Dual-input AI reduces 4-day solar wind error to 53 km/s","EUV images and wind history improve solar wind forecast to 53 km/s","Encoder-decoder model achieves 53 km/s error on 4-day solar wind forecast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central improvement claim, roughly 20% lower RMSE than Brown et al. (2022), rests on the assumption that daily-averaged forecasts evaluated on the balanced 2014-2017 validation window are directly comparable to hourly-averaged forecasts evaluated on different five-fold splits of solar cycle 24, even though the forecast cadence and difficulty differ.","fun_headline_variants_meta":{"raw":{"variants":["Multimodal net forecasts solar wind 4 days ahead to 53 km/s","Dual-input AI reduces 4-day solar wind error to 53 km/s","EUV images and wind history improve solar wind forecast to 53 km/s","Encoder-decoder model achieves 53 km/s error on 4-day solar wind forecast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001331,"raw_usage":{"total_tokens":5530,"prompt_tokens":1175,"completion_tokens":4355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":791,"completion_tokens_details":{"reasoning_tokens":4268}},"tokens_in":791,"tokens_out":4355,"duration_ms":28355,"temperature":1.0,"reasoning_tokens":4268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:51:53.216704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the same encoder-decoder architecture on the hourly-averaged solar wind data and the exact five-fold validation splits of Brown et al. (2022) and Upendran et al. (2020), reporting RMSE and Pearson r for days 1 through 4 at the same hourly cadence; if the day-4 RMSE does not fall below roughly 72 km/s for Brown et al. and 81 km/s for Upendran et al. on those splits, the claimed ~20% and ~29% improvements would not be confirmed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The main comparison baseline, combining EUV images with the 27-day-prior wind speed; the claimed ~20% RMSE improvement is measured against its four-day forecast."},{"cited_title":"F., Jin, M., et al","cited_arxiv_id":null,"evidence_quote":"Provides the SDOML standardized SDO/AIA 211 Angstrom image dataset used as model input."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the encoder-decoder sequence-to-sequence formulation that the architecture is based on."}],"review_version":1}