pith. sign in

arxiv: 2606.29644 · v1 · pith:URDYTI3Hnew · submitted 2026-06-28 · 💻 cs.LG · physics.space-ph

t-STEP: An interpretable model for Total Electron Content predictions and irregularities estimations

Stephen Tete , Carl Shneider , Maxime Cordy , Claudio Cesaroni , Andreas Hein , Vasily Petrov This is my paper

Pith reviewed 2026-06-30 07:06 UTC · model grok-4.3

classification 💻 cs.LG physics.space-ph
keywords Total Electron ContentTEC predictionionospheric irregularitiesmachine learninggeomagnetic stormsROT indexROTIinterpretable model
0
0 comments X

The pith

t-STEP predicts Total Electron Content every 30 seconds at 91% accuracy while estimating ionospheric irregularities.

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

The paper presents t-STEP as a machine learning model that predicts Total Electron Content at 30-second intervals from GPS data. It seeks to demonstrate that this high temporal resolution allows the model to also estimate signatures of TEC irregularities during geomagnetic storms. A reader would care because accurate short-term TEC forecasts could help mitigate disruptions to GPS and satellite communications caused by ionospheric disturbances. The model is trained on data from one station during solar cycle 24 and uses SHAP for feature interpretation and dynamic time warping for evaluation. It claims to outperform both the IRI-2020 model on hourly scales and an attention-based LSTM on irregularity capture.

Core claim

t-STEP is an interpretable model that predicts TEC at 30-second resolution using GPS observations from a station at 5.49°S, 47.49°W. During high solar activity in 2015, the predictions reach 91% accuracy with a mean absolute error of 4.38 TECU. The high-cadence output enables calculation of the Rate of TEC Index (ROTI) to monitor irregularities. The model captures the dynamics and morphologies of TEC irregularities during geomagnetic storms of varying intensities and outperforms an attention-based Long Short-Term Memory model. The hourly version of the model improves upon the International Reference Ionosphere (IRI-2020) by increasing accuracy 35%, reducing errors 57%, and boosting predictio

What carries the argument

The t-STEP model, which generates 30-second TEC predictions from which ROT and ROTI are derived to indicate ionospheric variability.

If this is right

  • The high temporal resolution preserves small-scale TEC irregularities in the predicted signals.
  • A single model framework supports both TEC prediction and irregularity monitoring without separate event-specific models.
  • The approach improves accuracy over the IRI-2020 reference model for hourly predictions.
  • SHAP explanations reveal feature contributions to the predictions.
  • Multi-metric evaluation including dynamic time warping confirms robustness.

Where Pith is reading between the lines

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

  • Testing the model on data from additional stations could reveal how well the irregularity signatures generalize beyond the training location.
  • Integration with real-time GPS networks might enable operational space weather alerts for communication systems.
  • The single-station training raises questions about performance during different solar cycles or at other latitudes.
  • Combining t-STEP with physics-based models could further constrain predictions under physical laws.

Load-bearing premise

The assumption that observations from a single station during solar cycle 24 suffice to establish the model's ability to capture general TEC irregularity signatures across conditions.

What would settle it

A comparison on data from a different geographic station or solar cycle where the t-STEP model shows lower accuracy or fails to match observed irregularity morphologies compared to the LSTM baseline.

read the original abstract

Earth system infrastructures relying on satellite-based technologies, such as Global Positioning System (GPS) communications, are affected by ionospheric Total Electron Content (TEC) gradients. Modeling these gradients under physical constraints remains challenging due to their dynamic and transient nature. While existing machine learning (ML) models can predict hourly TEC variations, it remains unclear whether their temporal resolution is sufficient to preserve small-scale TEC irregularities within predicted signals. To address this gap, we introduce an interpretable ML-based model, t-STEP, designed to predict TEC at a 30-second resolution and estimate irregularity signatures from the modeled signals. This high cadence enables the derivation of Rate of TEC changes (ROT) and the ROT Index (ROTI) as diagnostic indicators of ionospheric variability. The model is developed using GPS observations from solar cycle 24 at a station located at 5.49{\deg}S, 47.49{\deg}W. A multi-metric evaluation framework, including dynamic time warping, is used for robustness assessment, while SHAP (SHapley Additive exPlanations) provides insight into feature contributions. The 30-second TEC predictions achieve 91% accuracy with a mean absolute error (MAE) of 4.38 TECU during high solar activity (2015). Compared with the International Reference Ionosphere (IRI-2020), the hourly model improves accuracy by 35%, reduces absolute errors by 57%, and increases prediction skill by 54%. More importantly, the 30-second model captures TEC irregularity dynamics and morphologies during geomagnetic storms of different intensities, outperforming an attention-based Long Short-Term Memory model under the same experimental conditions. This study demonstrates the potential of a single TEC prediction framework for scalable irregularity monitoring without requiring separate models for individual transient events.

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

3 major / 1 minor

Summary. The manuscript introduces t-STEP, an interpretable ML model trained on GPS observations from a single equatorial station (5.49°S, 47.49°W) during solar cycle 24 to predict TEC at 30-second resolution. From the predicted signals it derives ROT and ROTI to estimate ionospheric irregularities, reports 91% accuracy and 4.38 TECU MAE for 2015 high solar activity, claims a 35% accuracy improvement, 57% error reduction and 54% skill increase over IRI-2020 for the hourly version, states that the 30-second model captures storm-time irregularity dynamics and morphologies while outperforming an attention-based LSTM under identical conditions, and concludes that the approach constitutes a scalable single-framework solution for irregularity monitoring.

Significance. If the generalization and validation claims hold, the work could supply a practical high-cadence TEC predictor with integrated irregularity diagnostics useful for GPS infrastructure and space-weather applications. Positive elements include the explicit use of SHAP for feature attribution, a multi-metric evaluation that incorporates dynamic time warping, and the attempt to unify prediction and irregularity estimation in one model rather than requiring separate event-specific models.

major comments (3)
  1. [Abstract] Abstract: the assertion of 'a single TEC prediction framework for scalable irregularity monitoring' rests on data from only one station at 5.49°S, 47.49°W during solar cycle 24; no cross-station, cross-latitude, or cross-cycle validation is described, so the extrapolation to arbitrary geomagnetic conditions and transient events is unsupported.
  2. [Abstract] Abstract: the claim that the 30-second model 'captures TEC irregularity dynamics and morphologies during geomagnetic storms of different intensities' and outperforms the attention-based LSTM is evaluated on the identical single-station dataset used for training; without an independent out-of-sample test set or external irregularity benchmark, it is unclear whether the ROT/ROTI signatures reflect genuine predictive skill or are recovered by construction from the fitted signals.
  3. [Abstract] Abstract: the quantitative improvements over IRI-2020 (35% accuracy, 57% absolute-error reduction, 54% skill increase) are stated only for the hourly model; the relationship of these metrics to the 30-second t-STEP predictions and to the irregularity estimation task is not specified.
minor comments (1)
  1. Ensure that the methods section explicitly reports data-split strategy, hyperparameter search, error bars or statistical tests for all reported metrics, and the precise definition of the 91% accuracy figure.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed review. We address each major comment below and will revise the abstract to ensure claims accurately reflect the scope and evidence in the study.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion of 'a single TEC prediction framework for scalable irregularity monitoring' rests on data from only one station at 5.49°S, 47.49°W during solar cycle 24; no cross-station, cross-latitude, or cross-cycle validation is described, so the extrapolation to arbitrary geomagnetic conditions and transient events is unsupported.

    Authors: We agree that the study uses observations from only one equatorial station during solar cycle 24. The phrasing 'scalable irregularity monitoring' is meant to describe the unified single-framework design rather than to assert validated performance across all latitudes, longitudes, or solar cycles. We will revise the abstract to explicitly qualify the dataset scope and present scalability as a proposed advantage of the approach. revision: partial

  2. Referee: [Abstract] Abstract: the claim that the 30-second model 'captures TEC irregularity dynamics and morphologies during geomagnetic storms of different intensities' and outperforms the attention-based LSTM is evaluated on the identical single-station dataset used for training; without an independent out-of-sample test set or external irregularity benchmark, it is unclear whether the ROT/ROTI signatures reflect genuine predictive skill or are recovered by construction from the fitted signals.

    Authors: The 30-second model was assessed via a temporal train-test split on the station data, with multiple geomagnetic storm intervals reserved for testing. We will revise the abstract to state the evaluation protocol clearly and to note the lack of an external irregularity benchmark, while retaining the reported outperformance relative to the LSTM under the same conditions. revision: yes

  3. Referee: [Abstract] Abstract: the quantitative improvements over IRI-2020 (35% accuracy, 57% absolute-error reduction, 54% skill increase) are stated only for the hourly model; the relationship of these metrics to the 30-second t-STEP predictions and to the irregularity estimation task is not specified.

    Authors: The cited improvements versus IRI-2020 apply exclusively to the hourly t-STEP model; the 30-second model is compared against the attention-based LSTM, and irregularity diagnostics are derived directly from the 30-second predictions. We will revise the abstract to distinguish the two model cadences and to indicate which metrics pertain to each component of the work. revision: yes

standing simulated objections not resolved
  • Cross-station, cross-latitude, or cross-cycle validation, as the study is restricted to a single station and solar cycle 24

Circularity Check

0 steps flagged

No significant circularity; standard supervised ML pipeline with external benchmarks

full rationale

The paper trains t-STEP on GPS observations from one station to predict 30 s TEC, then computes ROT/ROTI from the predicted time series as a post-processing diagnostic. This is ordinary supervised regression followed by derived metrics; performance is reported against held-out data, IRI-2020, and an attention-LSTM baseline, with SHAP for feature attribution. No equations, self-citations, or uniqueness claims reduce the central result to a tautology or to the training inputs by construction. The single-station scope is a generalization limit, not a circularity defect.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on fitted machine-learning parameters and the domain assumption that single-station training data suffices for general irregularity detection; no invented physical entities are introduced.

free parameters (1)
  • ML model parameters
    All weights and hyperparameters of the t-STEP model are fitted to the GPS training data from the single station.
axioms (1)
  • domain assumption 30-second temporal resolution preserves small-scale TEC irregularities within predicted signals
    Explicitly stated as the motivation for moving beyond hourly models.

pith-pipeline@v0.9.1-grok · 5870 in / 1210 out tokens · 45564 ms · 2026-06-30T07:06:08.386794+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

2 extracted references · 2 canonical work pages

  1. [1]

    Space Sci Rev 63:209–243

    Aarons J (1993) The longitudinal morphology of equatorial F-layer irregularities relevant to their occurrence. Space Sci Rev 63:209–243. https://doi.org/10.1007/BF00750769 Adolfs M, Hoque MM, Shprits YY (2022) Storm-Time Relative Total Electron Content Modelling Using Machine Learning Techniques. Remote Sens 14:1–17. https://doi.org/10.3390/rs14236155 Bil...

    work page doi:10.1007/bf00750769 1993

  2. [2]

    Radio Sci 52:439–460

    https://doi.org/10.1002/2015SW001182 de Oliveira Moraes A, Costa E, Abdu MA, et al (2017) The variability of low-latitude ionospheric amplitude and phase scintillation detected by a triple-frequency GPS receiver. Radio Sci 52:439–460. https://doi.org/10.1002/2016RS006165 Gonzalez WD, Joselyn JA, Kamide Y, et al (1994) What is a geomagnetic storm? J Geophy...

    work page doi:10.1002/2015sw001182 2017