pith. sign in

arxiv: 2606.12661 · v1 · pith:FV25EKFQnew · submitted 2026-06-10 · 🌌 astro-ph.SR · physics.plasm-ph· physics.space-ph

Finding Novel Precursors for Solar Wind Stream Interaction Regions with Interpretable Deep Learning

Prateek Mayank , Yogesh , Enrico Camporeale , D. Chakrabarty , Lan K Jian , Gregory G. Howes , Thomas E. Berger This is my paper

Pith reviewed 2026-06-27 08:02 UTC · model grok-4.3

classification 🌌 astro-ph.SR physics.plasm-phphysics.space-ph
keywords solar windstream interaction regionsSIR detectionTransformerIntegrated Gradientsinterpretabilityflow deflectionspace weather
0
0 comments X

The pith

A lightweight Transformer detects solar wind stream interaction regions and ranks proton density, magnetic field, and flow deflection as top precursors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces SIREN, a two-layer Transformer encoder that ingests sequences of eleven solar wind parameters to output a calibrated probability of stream interaction region presence at each time step. Self-attention weights concentrate on the relevant interval of each sequence, while Integrated Gradients attribution produces a feature ranking that places proton density and magnetic field magnitude first, followed by temperature and bulk speed, with transverse velocity and east-west flow angle together contributing 13 to 17 percent. The model reaches a ROC-AUC of 0.93 on a held-out set of 102 events and supplies continuous probabilities rather than fixed labels. This supplies both an operational detection tool and a quantitative basis for treating flow deflection as a consistent SIR signature.

Core claim

SIREN is a compact Transformer with approximately 100000 parameters that processes eleven in-situ solar wind measurements spanning magnetic field, velocity, and thermodynamic quantities. Trained with weighted binary cross-entropy and cosine annealing, then calibrated via Platt scaling, the model yields per-timestep SIR probabilities. On 102 held-out events it records ROC-AUC 0.93, F1 0.78, and true skill statistic 0.67. Self-attention focuses on the interaction region, and Integrated Gradients attribution assigns 24.3 percent to proton density, 21.6 percent to magnetic field magnitude, 13.9 percent to temperature, 12.1 percent to bulk speed, and 13-17 percent combined to Vy and flow angle.

What carries the argument

SIREN two-layer Transformer encoder with self-attention weights and Integrated Gradients attribution applied to sequences of eleven solar wind parameters.

If this is right

  • Continuous probability outputs permit operators to select detection thresholds matched to specific operational requirements.
  • The quantified contribution of transverse velocity and flow angle establishes flow deflection as a measurable and consistent SIR signature.
  • The lightweight architecture demonstrates that compact interpretable models can reach high performance on space-weather classification tasks.
  • Attribution methods confirm that decisions rest on the physically relevant portion of each input sequence.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported feature hierarchy could be used to derive simplified, physics-informed detection rules that incorporate flow deflection explicitly.
  • Retraining or testing the same architecture on data from additional spacecraft would indicate whether the precursor ranking changes with radial distance from the Sun.
  • Embedding the continuous probabilities into ensemble forecasting systems could improve lead-time estimates for recurrent geomagnetic storms.
  • The combination of attention and gradient attribution supplies a reusable template for uncovering new signatures in other solar-wind time series.

Load-bearing premise

The human-generated or catalog-derived labels used to train and evaluate the model are accurate and free of systematic bias, and the post-hoc attributions correctly isolate physically causal features rather than spurious correlations present in the training distribution.

What would settle it

Re-labeling the 102 test events by an independent panel of experts or running the model on simultaneous multi-spacecraft observations at separated heliocentric distances would test whether the reported performance and feature ranking persist.

Figures

Figures reproduced from arXiv: 2606.12661 by D. Chakrabarty, Enrico Camporeale, Gregory G. Howes, Lan K Jian, Prateek Mayank, Thomas E. Berger, Yogesh.

Figure 1
Figure 1. Figure 1: Overview of the SIR detection framework. (a) Model architecture processed by a two-layer Transformer encoder. (b) Training and evaluation pipeline: data preparation, normalization, and splitting feed into a training loop. to encode temporal order. Third, the projected sequence passes through a two-layer Transformer encoder, where each layer contains a multi-head self-attention block with 4 heads (head dime… view at source ↗
Figure 2
Figure 2. Figure 2: Training history of SIREN over 200 epochs. Top row: training and validation cross-entropy loss (left), precision and recall (center), and F1 score (right). Bottom row: ROC-AUC (left), true skill statistic (TSS) and Heidke skill score (HSS) (center), and validation accuracy (right). All metrics are computed on the validation set except training loss. 3.1 Training Convergence and Test-Set Evaluation [PITH_F… view at source ↗
Figure 3
Figure 3. Figure 3: Model performance metrics across training, validation, and test sets. ceiver operating characteristic area under the curve (ROC-AUC) stabilizes above 0.90 by epoch 100, confirming strong discrimination between SIR and ambient-wind timesteps across all probability thresholds [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Self-attention weights from the two Transformer encoder layers for a representative test-set sample. Top row is for Layer 1 and bottom row for Layer 2. For each layer, the left panel shows the full attention matrix, with key time steps on the vertical axis and query time steps on the horizontal axis, and the right panel shows the mean attention weight at each key timestep, averaged over all query positions… view at source ↗
Figure 5
Figure 5. Figure 5: Global feature attribution from Integrated Gradients. (a) Fractional importance of each input feature, computed as mean absolute IG attribution across all 102 test-set samples. (b) Temporal attribution magnitude averaged across all test-set samples, with features on the vertical axis and timesteps on the horizontal axis (log scale). Figure 5a presents the global feature importance ranking, computed as the … view at source ↗
Figure 6
Figure 6. Figure 6: Case studies of two contrasting test-set events. Left column (a1-a7): a strong and well-defined SIR. Right column: a weak and complex SIR. For each event, panels show (top to bottom): solar wind time series (velocity, magnetic field, density and temperature, flow angles), calibrated SIR probability with catalog mask, signed IG attribution heatmap, and per-feature fractional importance for the full sequence… view at source ↗
Figure 7
Figure 7. Figure 7: Real-time response of SIREN to evolving SIR properties. Panels (a-e) demonstrate the con￾tinuous ingestion of solar wind data. The model probability (black) and catalogue SIR interval (red shaded region) are shown along with velocity (blue) and density (green) time series. –12– [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Superposed epoch analysis (SPA) of 676 SIR events. The black vertical dashed line denotes the epoch time, defined by the stream interface. Panels (a–f) in the left column show the magnetic field magnitude and components (|B|, Bx, By, Bz), temperature (T), and the north–south flow angle (θ). Panels (g–l) in the right column display the velocity magnitude and components (|V |, Vx, Vy, Vz), proton number dens… view at source ↗
read the original abstract

Solar wind stream interaction regions (SIRs) drive recurrent geomagnetic storms, yet most existing catalogs rely on expert inspection and simple thresholds that are subjective and can miss events with complex morphologies. We present SIREN (SIR Encoder Network), a lightweight Transformer based model for per timestep SIR detection from in situ solar wind observations. The model ingests sequences of 11 solar wind parameters, spanning magnetic field, velocity, and thermodynamic properties. With approximately 100,000 trainable parameters in a two layer encoder architecture, SIREN is trained using weighted binary cross entropy loss and a cosine annealing learning rate. Platt scaling is applied to produce well-calibrated detection probabilities. On a held-out test set of 102 events, the calibrated model achieves a ROC-AUC of 0.93, F1 score of 0.78, and true skill statistic of 0.67. Analysis of the self-attention weights confirms that the model concentrates on the SIR, grounding its decisions in the physically relevant portion of each sequence. Integrated Gradients attribution reveals a quantifiable feature hierarchy: proton density (24.3%) and magnetic field magnitude (21.6%) dominate, followed by temperature (13.9%) and bulk speed (12.1%). Notably, the transverse velocity component Vy and east-west flow angle together contribute 13-17%, identifying flow deflection as a consistent but previously under-quantified SIR signature. By producing continuous probabilities rather than binary labels, SIREN enables flexible threshold tuning for operational use and provides a template for compact, interpretable deep-learning systems in space weather.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

4 major / 3 minor

Summary. The paper introduces SIREN, a lightweight two-layer Transformer encoder (~100k parameters) that ingests sequences of 11 in-situ solar wind parameters to perform per-timestep binary detection of stream interaction regions (SIRs). Trained with weighted binary cross-entropy and cosine annealing, followed by Platt scaling, the model reports ROC-AUC 0.93, F1 0.78 and TSS 0.67 on a held-out test set of 102 events. Self-attention weights are shown to focus on the SIR interval, while Integrated Gradients attributions rank proton density (24.3 %) and |B| (21.6 %) highest, with transverse velocity Vy and flow angle together contributing 13–17 % and thereby identifying flow deflection as a previously under-quantified signature. The work positions the continuous probability output as operationally flexible and the attribution pipeline as a template for interpretable space-weather models.

Significance. If the reported performance and attributions survive independent label validation and temporal-split verification, the result would supply a compact, reproducible, and physically grounded alternative to subjective catalog construction, while the quantified contribution of flow deflection could motivate targeted observational or simulation studies of SIR morphology.

major comments (4)
  1. [Abstract, §3] Abstract and §3 (data & labels): the central performance figures (ROC-AUC 0.93, F1 0.78) and the Integrated Gradients ranking rest on the assumption that the human- or catalog-derived labels are an unbiased proxy for true SIR occurrence. The abstract itself states that existing catalogs rely on subjective thresholds and miss complex events, yet no cross-catalog comparison, independent expert re-labeling, or sensitivity test to label noise is described; any systematic bias in the labels is therefore inherited by both the metrics and the claimed “novel precursor.”
  2. [§4] §4 (experimental setup): no information is supplied on whether the train–test split of the time-series data respects chronological order. In recurrent solar-wind data, non-temporal splitting risks leakage of future SIR signatures into the training set, which would directly inflate the reported test-set metrics and render the attribution results circular.
  3. [Results] Results paragraph and Table 2 (if present): the abstract gives point estimates for ROC-AUC, F1 and TSS on the 102-event test set but supplies neither bootstrap or binomial confidence intervals nor a breakdown by event morphology; without these, it is impossible to judge whether the claimed superiority over catalog thresholds is statistically robust.
  4. [§5.2] §5.2 (attribution analysis): the claim that transverse velocity and flow angle constitute a “consistent but previously under-quantified SIR signature” is supported only by post-hoc Integrated Gradients on a model trained on the same labels; no external physical validation (e.g., comparison with MHD simulations or independent in-situ studies) is provided to distinguish causal contribution from correlation induced by the labeling procedure itself.
minor comments (3)
  1. [Abstract, §2] The abstract states “approximately 100,000 trainable parameters” but does not give the exact count or the embedding/hidden dimension that would allow reproduction; add the precise architecture table.
  2. [§2] Notation for the 11 input parameters is introduced only in the abstract; a compact table listing each variable, its units, and its source instrument would improve clarity for readers outside the immediate sub-field.
  3. [Abstract, §5.2] The phrase “flow deflection as a consistent but previously under-quantified SIR signature” appears without a supporting citation to prior literature that quantified the same feature; add the relevant references or qualify the novelty claim.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for their thorough and constructive review. The comments correctly identify gaps in label validation, experimental transparency, statistical reporting, and external attribution support. We respond point-by-point below and indicate the revisions that will be incorporated.

read point-by-point responses

  1. Referee: [Abstract, §3] Abstract and §3 (data & labels): the central performance figures (ROC-AUC 0.93, F1 0.78) and the Integrated Gradients ranking rest on the assumption that the human- or catalog-derived labels are an unbiased proxy for true SIR occurrence. The abstract itself states that existing catalogs rely on subjective thresholds and miss complex events, yet no cross-catalog comparison, independent expert re-labeling, or sensitivity test to label noise is described; any systematic bias in the labels is therefore inherited by both the metrics and the claimed “novel precursor.”

    Authors: We agree that label quality is a foundational issue. The manuscript notes subjectivity in existing catalogs but does not quantify its impact through cross-catalog comparison or noise sensitivity tests. In revision we will expand §3 with a discussion of label limitations, add a limited cross-catalog consistency check using an independent SIR list where overlap exists, and include a simple label-perturbation sensitivity experiment. A full independent expert re-labeling campaign lies outside the scope of the present study and will be noted as future work. revision: partial

  2. Referee: [§4] §4 (experimental setup): no information is supplied on whether the train–test split of the time-series data respects chronological order. In recurrent solar-wind data, non-temporal splitting risks leakage of future SIR signatures into the training set, which would directly inflate the reported test-set metrics and render the attribution results circular.

    Authors: The referee correctly flags an important omission. Our split was performed chronologically (training data ending before the test interval begins) to avoid leakage; however, this was not stated in the text. We will revise §4 to document the exact temporal boundaries, the rationale for the split, and verification that no future information entered training. revision: yes

  3. Referee: [Results] Results paragraph and Table 2 (if present): the abstract gives point estimates for ROC-AUC, F1 and TSS on the 102-event test set but supplies neither bootstrap or binomial confidence intervals nor a breakdown by event morphology; without these, it is impossible to judge whether the claimed superiority over catalog thresholds is statistically robust.

    Authors: We accept that point estimates alone limit assessment of robustness. In the revised results section we will report bootstrap confidence intervals for ROC-AUC, F1 and TSS and add a morphology-stratified performance table (simple vs. complex SIRs) to allow readers to evaluate stability across event types. revision: yes

  4. Referee: [§5.2] §5.2 (attribution analysis): the claim that transverse velocity and flow angle constitute a “consistent but previously under-quantified SIR signature” is supported only by post-hoc Integrated Gradients on a model trained on the same labels; no external physical validation (e.g., comparison with MHD simulations or independent in-situ studies) is provided to distinguish causal contribution from correlation induced by the labeling procedure itself.

    Authors: The attribution is indeed post-hoc and label-dependent. We will revise §5.2 to moderate the language, explicitly state that the result is model-derived, add citations to existing in-situ studies of SIR flow deflection, and frame the finding as a hypothesis-generating observation that warrants targeted MHD or observational follow-up. No new external validation data will be added at this stage. revision: partial

Circularity Check

0 steps flagged

No circularity: standard ML training and post-hoc attribution on held-out data

full rationale

The paper describes training a Transformer model on catalog-derived labels, evaluating performance metrics on a held-out test set of 102 events, and applying Integrated Gradients for feature attribution. No equations, self-citations, or derivations reduce the reported ROC-AUC, F1, or feature percentages to the inputs by construction. The central results are empirical outputs of supervised learning rather than tautological renamings or fitted parameters relabeled as predictions. Label quality is an external assumption but does not create a self-referential derivation chain within the paper's own steps.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim depends on the correctness of supervised labels and the reliability of post-hoc attribution methods. The model itself introduces ~100k fitted parameters whose values are determined entirely by the training data.

free parameters (2)
  • ~100000 model parameters
    Weights of the two-layer Transformer encoder learned via gradient descent on the training set.
  • Platt scaling parameters
    Two parameters fitted to map raw model outputs to calibrated probabilities.
axioms (2)
  • domain assumption Training and test labels for SIR presence are accurate and representative of the underlying physical distribution.
    The entire supervised pipeline rests on the quality of these labels.
  • ad hoc to paper Integrated Gradients and self-attention weights recover physically meaningful causal contributions rather than dataset-specific correlations.
    This is an untested assumption required to interpret the 24.3% density and 13-17% velocity attributions as novel precursors.

pith-pipeline@v0.9.1-grok · 5849 in / 1683 out tokens · 31223 ms · 2026-06-27T08:02:51.388539+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    B., & Taroyan, Y

    Alielden, K., Camporeale, E., Kors´ os, M. B., & Taroyan, Y. (2023, March). Prediction Inter- val of Interface Regions: Machine Learning Nowcasting Approach.Space Weather,21(3), e2022SW003326. doi: 10.1029/2022SW003326 Alves, M. V., Echer, E., & Gonzalez, W. D. (2006, July). Geoeffectiveness of corotating interaction regions as measured by Dst index.Journ...

    work page doi:10.1029/2022sw003326 2023

  2. [2]

    doi: 10.1088/0004-637X/798/2/135 Camporeale, E., Car` e, A., & Borovsky, J. E. (2017, November). Classification of Solar Wind With Machine Learning.Journal of Geophysical Research (Space Physics),122(11), 10,910-10,920. doi: 10.1002/2017JA024383 Chi, Y., Shen, C., Luo, B., Wang, Y., & Xu, M. (2018, December). Geoeffectiveness of Stream Interaction Regions...

    work page doi:10.1088/0004-637x/798/2/135 2017

  3. [3]

    T., & Gonzalez, W

    Echer, E., Tsurutani, B. T., & Gonzalez, W. D. (2013, January). Interplanetary origins of moderate (-100 nT ¡ Dst≤-50 nT) geomagnetic storms during solar cycle 23 (1996-2008).Journal of Geophysical Research (Space Physics),118(1), 385-392. doi: 10.1029/2012JA018086 Gonz´ alez-Avil´ es, J. J., Riley, P., Ben-Nun, M., Mayank, P., & Vaidya, B. (2024, August)...

    work page doi:10.1029/2012ja018086 2013

  4. [4]

    T., Asbridge, J

    doi: 10.1051/swsc/2024014 Gosling, J. T., Asbridge, J. R., Bame, S. J., & Feldman, W. C. (1978, April). Solar wind stream inter- faces.Journal of Geophysical Research,83(A4), 1401-1412. doi: 10.1029/JA083iA04p01401 Gosling, J. T., & Pizzo, V. J. (1999, July). Formation and Evolution of Corotating Interaction Regions and their Three Dimensional Structure.S...

    work page doi:10.1051/swsc/2024014 1978

  5. [5]

    Decoupled Weight Decay Regularization

    doi: 10.1007/ s11207-022-01962-1 Hu, A., Camporeale, E., & Swiger, B. (2023, April). Multi-Hour-Ahead Dst Index Prediction Using Multi-Fidelity Boosted Neural Networks.Space Weather,21(4), e2022SW003286. doi: 10.1029/2022SW003286 Ishii, M., Costa, J. E. R., Kuznetsova, M. M., Andries, J., Gopalswamy, N., Belehaki, A., . . . Inter- national Space Weather C...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1029/2022sw003286 2023

  6. [6]

    (2024, November)

    doi: 10.1007/s12036-025-10107-2 Mayank, P., Lotz, S., Vaidya, B., Mishra, W., & Chakrabarty, D. (2024, November). Study of Evolution and Geo-effectiveness of Coronal Mass Ejection–Coronal Mass Ejection Interactions Using Magnetohydrodynamic Simulations with SWASTi Framework.The Astrophysical Jour- nal,976(1),

    work page doi:10.1007/s12036-025-10107-2 2024

  7. [7]

    (2022, 09)

    doi: 10.3847/1538-4357/ad8084 Mayank, P., Vaidya, B., & Chakrabarty, D. (2022, 09). Swasti-sw: Space weather adaptive simulation framework for solar wind and its relevance to the aditya-l1 mission.The Astrophysical Journal Supplement Series,262,

    work page doi:10.3847/1538-4357/ad8084 2022

  8. [8]

    (2023, 12)

    Retrieved 2023-01-31, fromhttps://doi.org/10.3847/1538 -4365/ac8551doi: 10.3847/1538-4365/ac8551 Mayank, P., Vaidya, B., Mishra, W., & Chakrabarty, D. (2023, 12). Swasti-cme: A physics-based model to study coronal mass ejection evolution and its interaction with solar wind.The Astrophysical Journal Supplement Series,270, 10-10. doi: 10.3847/1538-4365/ad08...

    work page doi:10.3847/1538 2023

  9. [9]

    The Astro- physical Journal Letters940(1), 25 (2022) https://doi.org/10.3847/2041-8213/ ac9f38

    doi: 10.1007/ s12036-025-10064-w National Academies of Sciences, Engineering, and Medicine. (2024).The next decade of discovery in solar and space physics: Exploring and safeguarding humanity’s home in space. Ofman, L., Yogesh, & Giordano, S. (2024, jul). Understanding the variability of helium abundance in the solar corona using three-fluid modeling and ...

    work page doi:10.3847/2041-8213/ 2024

  10. [10]

    doi: 10.3847/1538-4357/ad54c3 Platt, J. (1999). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods.Advances in large margin classifiers,10(3), 61–74. Riley, P., & Lionello, R. (2011, 05). Mapping solar wind streams from the sun to 1 au: A comparison of techniques.Solar Physics,270, 575-592. doi: 10.1007/s11...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538-4357/ad54c3 1999

  11. [11]

    T., Gonzalez, W

    doi: 10.3847/1538-4357/ada93c Tsurutani, B. T., Gonzalez, W. D., Gonzalez, A. L. C., Guarnieri, F. L., Gopalswamy, N., Grande, M., . . . Vasyliunas, V. (2006, July). Corotating solar wind streams and recurrent geomagnetic activity: A review.Journal of Geophysical Research (Space Physics),111(A7), A07S01. doi: 10.1029/2005JA011273 Upendran, V., Cheung, M. ...

    work page doi:10.3847/1538-4357/ada93c 2006

  12. [12]

    V., Meng, X., et al

    doi: 10.1088/0004-637X/782/2/81 Vysakh, P. A., & Mayank, P. (2023, November). Solar Flare Prediction and Feature Selection Using a Light-Gradient-Boosting Machine Algorithm.Solar Physics,298(11),

    work page doi:10.1088/0004-637x/782/2/81 2023

  13. [13]

    (2021, May)

    doi: 10.1007/ s11207-023-02223-5 Yogesh, Chakrabarty, D., & Srivastava, N. (2021, May). Evidence for distinctive changes in the solar wind helium abundance in solar cycle 24.Monthly Notices of the Royal Astronomical Society,503(1), L17-L22. doi: 10.1093/mnrasl/slab016 Yogesh, Chakrabarty, D., & Srivastava, N. (2023, November). New insights on the behaviou...

    work page doi:10.1093/mnrasl/slab016 2021

  14. [14]

    G., Niranjana, Martinovi´ c, M., Howes, G

    doi: 10.3847/1538-4357/ad84d6 Yogesh, Ofman, L., Klein, K. G., Niranjana, Martinovi´ c, M., Howes, G. G., . . . et al. (2026, March). Solar Wind Heating near the Sun: A Radial Evolution Approach.The Astrophysical Journal, 999(2),

    work page doi:10.3847/1538-4357/ad84d6 2026

  15. [15]

    S., Deehr, C

    doi: 10.3847/1538-4357/ae4582 Zhang, Y., Sun, W., Feng, X. S., Deehr, C. S., Fry, C. D., & Dryer, M. (2008, August). Statistical analysis of corotating interaction regions and their geoeffectiveness during solar cycle 23.Jour- nal of Geophysical Research (Space Physics),113(A8), A08106. doi: 10.1029/2008JA013095 –22–

    work page doi:10.3847/1538-4357/ae4582 2008