Low-Earth-orbit (LEO) satellite internet services such as Starlink are modeled today as a geometry problem. Latency spikes and throughput dips are blamed on satellite handoffs, dish pointing and obstructions, and the best predictors, most notably T3P, use only orbital and pointing features. Those models were built and tested only in the UK and Spain. Near the magnetic equator, the upper atmosphere behaves very differently. After sunset it develops "plasma bubbles," drifting plumes of depleted, turbulent plasma that make radio signals flicker (scintillate) and that knock navigation receivers out of lock. This proposal asks a narrow question that can be answered either way: do these bubbles leave a detectable, separable signature in the ordinary traffic of LEO broadband users, and does that signature follow the latitude pattern the physics predicts? The measurements are passive, taken from users' own TCP connections without injecting test traffic.
The design turns an easily confounded correlation into a controlled comparison. Bubbles are field-aligned plumes that occupy some directions in the sky and not others. A terminal switches satellites every few seconds to tens of seconds, so within a single 15-minute window its line of sight passes through a bubble for some serving intervals and misses it for others. Congestion, time of day and weather are shared across those intervals. We compare throughput and loss between these interleaved intervals, using a co-located navigation receiver to map where the bubbles are. The prediction is that the effect should be present under the equatorial anomaly crest and absent at mid-latitudes. If that holds, it is the first evidence that geometry-only LEO models are missing a physical term that depends on latitude. If it fails, it is a quantified bound that justifies ignoring the ionosphere.
What the directions contain, and a correction to the obvious idea
Latitudinal ionospheric variability comes from the "fountain effect." Near the magnetic equator, electric fields lift plasma upward, and the plasma then slides down the magnetic field lines to pile up in two crests roughly 15–20° north and south of the magnetic equator. Huang's ionosonde study of Guangzhou placed that city under the northern crest and counted 509 negative and 315 positive disturbances over 1965–1982, far more than stations to its north or south. Huang also showed that crest densities persist into the night at solar maximum and vanish at solar minimum. Forecasting work confirms that this zone is hard to predict. Attentional BiGRU TEC prediction reports accuracy rising steadily from 15° N to 78° N, and deep foF2 forecasting finds the same low-latitude penalty. These papers motivate the latitudinal contrast in our design.
Ionospheric signal distortion is well characterized at navigation frequencies. Han et al. co-registered all-sky airglow images with multi-GNSS tracking at Zhuhai during a bubble on 3 April 2022. They found that scintillation is strongest in the bubble's depleted center rather than at its edges, and that tracking stability varies by constellation and signal. Long Tang et al. identified a circularity problem. The usual turbulence index, ROTI, is computed from carrier phase, and carrier phase is exactly the measurement that breaks during storms. They replaced it with a Doppler Index (DI) that is immune to phase breaks, and report 42.4% and 63.1% horizontal positioning gains in two 2024 storms. Tiantian Tang et al. showed that low-latitude ionospheric residuals are heavy-tailed. Refining their correction grid from about 150 km to about 50 km spacing in Guangdong cut the residual spread by about 30%. That result is evidence that the ionospheric state has to be resolved at sub-hundred-kilometre scales, which is roughly the spacing between the points where different satellites' lines of sight cross the ionosphere. It is the reason this proposal uses direction-resolved rather than site-averaged ionospheric features.
Passive throughput and latency estimation reads network quality from production traffic. Dovrolis et al. infer round-trip time from TCP handshakes and slow-start flights. TimeWeaver mines NTP packets for one-way delay. Lean Algorithms and the Lossy Difference Sketch compress delay and loss tracking into switch memory. All of these were validated on terrestrial traces, and TimeWeaver explicitly filtered out wireless clients.
The obvious idea at this intersection is "make LEO latency models ionosphere-aware by correcting for ionospheric group delay." It fails on arithmetic. The extra propagation path is about 40.3 × TEC / f², with TEC in electrons per square metre and f in hertz. At a Ku-band carrier near 12 GHz, even 100 TEC units of slant content adds only about 0.3 m of path, roughly one nanosecond. That is six orders of magnitude below the 30–50 ms medians T3P reports. Delay is not the channel. Any real effect must arrive through amplitude scintillation, meaning fades that force the link to drop to a slower modulation or lose frames, and so appear as throughput dips and bursty loss. Scintillation also weakens with frequency, so an effect that is dramatic at the L-band frequencies Han et al. measured may shrink toward nothing at Ku/Ka band. Whether it survives is the open empirical question, and it is the one this proposal answers.
Core Claim
Our hypothesis is that at sites under the equatorial anomaly crest, and near solar maximum, the serving intervals of a LEO terminal whose line of sight crosses an actively scintillating region show measurably higher rates of throughput collapse and loss bursts than interleaved intervals in the same window that miss it. The effect should rank crest ≥ near-dip-equator > mid-latitude ≈ 0. A turbulence feature that is resolved by direction and immune to phase breaks should add predictive power to T3P's geometry features. Global space-weather indices should not add comparable power.
To our knowledge, no prior work in the file measures communication-layer throughput or loss against direction-resolved ionospheric state. Both ionosphere–signal papers (Han et al., Tiantian Tang et al.) stop at receiver tracking and positioning integrity. The one LEO transport study (T3P) ignores the ionosphere, uses active pings and API polling, and never leaves the mid-latitudes.
Research Program
The program has four parts: deployment, alignment of the two data streams, the statistical design, and a demonstration of what the findings mean downstream.
Deployment. We place LEO broadband terminals along a magnetic-latitude chain. One site is at mid-latitude, reproducing T3P's regime as a control. One is under an anomaly crest. One is near the dip equator. Starlink does not serve mainland China, so the network sites will be chosen at matched magnetic latitudes in regions where service exists. The Zhuhai airglow–GNSS setup of Han et al. serves as the physics reference for calibrating bubble detection. Each site pairs the terminal with a multi-constellation, multi-frequency GNSS receiver sampling at 1 Hz, the same class of instrument Han et al. used. Traffic is captured passively on the terminal's LAN side from production use by consenting households or a small campus. Collection spans two equinox seasons around the current solar maximum, because Huang's solar-cycle dependence implies the window matters.
Aligning ionospheric state with the network data. For every 1-second step we compute the Starlink serving satellite's line of sight from the terminal's pointing telemetry, the same geometry inputs T3P uses. We then find where that line crosses a thin shell at about 300–350 km altitude, which we call the pierce point. The ionospheric feature at that pierce point is interpolated from the nearby GNSS pierce points. Following Long Tang et al., the primary turbulence feature is the Doppler Index. We also compute S4, the amplitude scintillation index, and ROTI, and each is evaluated as an ablation. Interpolation is weighted along geomagnetic field lines, because bubbles are field-aligned. Tiantian Tang et al.'s 50 km result sets the resolution target.
Passive network signals. From the captured traffic we derive round-trip times using Dovrolis-style handshake and slow-start estimation, loss from retransmission and duplicate-acknowledgment patterns, and delivered throughput from bulk flows. Dovrolis et al. report that 40–50% of connections are skip cases that yield no estimate. We will report that coverage per site, because low-latitude fades could raise the skip rate through retransmission timeouts, and that would itself be evidence of an effect.
Statistical design. The primary estimand is a within-window contrast. Inside each 15-minute window we compare the rate of throughput collapses and loss bursts between serving intervals whose pierce point falls in the scintillating region (high DI) and intervals that fall outside it. The comparison is matched on elevation, obstruction flags and handoff phase. Because the intervals are interleaved, gateway congestion, evening demand peaks and weather cancel.
The secondary estimand is a night-to-night contrast. We compare bubble nights against quiet nights at the same local hour. This contrast matters because bubbles occur at roughly 19–24 local time, exactly when residential demand peaks. Any analysis that compares only evening against daytime traffic is confounded by that demand peak.
Rain fade is the remaining threat at Ku/Ka band in the tropics. It is handled in two ways. First, rain attenuates the signal but produces no GNSS Doppler turbulence. Second, windows with rain-gauge or radar precipitation at the site are excluded as a sensitivity check.
Downstream consequence. If a signature exists, we replay the measured traces through a link emulator. The first comparison is T3P-style tuned BBRv2 against the same controller given one extra bit: whether the current loss is attributed to the ionosphere. Separately, we inject the measured scintillation loss bursts into CAIDA backbone traces. We then check whether Lean Algorithms flags the affected flows as "problematic." That sketch was evaluated only with independent 1–10% random loss, so bursty physical-layer loss is an untested case for it.
Baselines
Naive reference. The do-nothing reference predicts each interval's throughput and loss from the site's median for that hour of day, plus a persistence term that repeats the last observed value. Any ionospheric effect that this reference already captures is really a demand effect, and the night-to-night design exists to expose exactly that.
Strongest existing method. This is T3P's predictor, reimplemented with its LSTM latency model and its XGBoost throughput model on azimuth, elevation, orbital state and terminal orientation. T3P reports a latency MAPE (mean absolute percentage error) of 3.65% at its UK/Spain sites, a 115% throughput gain from LEO-tuned BBRv2, and a 25% QoE improvement for video. No reported result was found on Lacuna for T3P's throughput-prediction error, so our own reimplementation's mid-latitude score will serve as the reference. It is the strongest competitor because it already explains handoff-driven tails, which are the dominant non-ionospheric source of variance. It will be trained on the same site-seasons, with the same train/test split by night, the same feature-engineering effort, and the same hyperparameter search budget as every variant we add.
Strengthened generic baseline. This is T3P plus global space-weather indices (Kp, Dst, F10.7) and vertical TEC from IGS global maps at the site. It tests whether any ionospheric information helps, as opposed to direction-resolved information.
Ablations that isolate the key ingredient. The first ablation replaces DI with ROTI, which tests Long Tang et al.'s argument that phase-derived indices fail precisely when they matter. The second replaces direction-resolved features with a site-averaged value. The third is a placebo: DI series are shuffled across nights while hour-of-day structure is preserved, so any gain the placebo keeps is diurnal leakage rather than physics.
What counts as winning. Beating the baselines requires three things. The DI-augmented model must improve prediction of throughput collapse at the crest site on bubble nights, measured as area under the ROC curve and as tail-quantile error, beyond the space-weather-augmented T3P, with a confidence interval from night-level bootstrap excluding zero. The placebo and mid-latitude runs must show no such gain. The within-window contrast must be significant after matching. We will pre-register the minimum effect we consider operationally meaningful before unblinding the low-latitude data.
Why not just…?
Why not just add hour-of-day and Kp to T3P and stop? This is a real threat, and the strengthened baseline tests it directly. The prior reason to expect it to fall short is that bubbles are local and occur irregularly from night to night. Huang's disturbance counts and the low-latitude forecasting penalty in Lei et al. both indicate that crest behavior is poorly captured by global drivers. Tiantian Tang et al.'s 30% gain from finer grids argues that site-level averages blur the structure. If global indices match the DI features, we will report that and recommend the cheap route.
Why not just rely on the operator's adaptive modulation and active pings: if fades mattered, wouldn't the dish already handle them? Adaptive modulation converts fades into lower throughput rather than hiding them, so a throughput signature should remain even if loss is masked. Active pings at low rates also miss short transients, and T3P's roughly 15-second spikes show how brief the relevant events are. Still, the scaling of scintillation with frequency means the true effect at Ku band may be negligible. The within-window contrast is built to deliver a tight upper bound in that case, and a clean null at the crest near solar maximum would settle the question for the field.
Why not just use a loss-tolerant controller like BBR, which already ignores isolated losses? BBRv2 does respond to loss thresholds, and T3P's 115% gain from retuning them shows that the thresholds matter in LEO. The replay experiment asks whether attribution adds anything beyond T3P's tuning. If it adds nothing, the downstream claim is dropped and the measurement result stands on its own.
Why not just attribute tropical evening dips to rain? Rain and scintillation separate cleanly in the GNSS data. Rain attenuates the signal but produces no Doppler-index turbulence, and the within-window contrast is keyed to turbulence along the specific line of sight, not to overall signal level. Excluding precipitation windows provides a second check.
Decisive Checks
Three results together would establish the claim. First, a matched within-window effect at the crest site on bubble nights. Second, the predicted latitudinal ordering across the chain, with the mid-latitude site reproducing T3P's geometry-only sufficiency. Third, DI features beating both the space-weather-augmented T3P and the shuffled-night placebo, while ROTI performs worse during strong turbulence as Long Tang et al.'s circularity argument predicts. The secondary check is the CAIDA injection: whether scintillation-shaped loss bursts produce false "problematic flow" flags in Lean Algorithms at 40 KB and at 160 KB of memory. That paper reports about 82% and above 90% detection of the top-100 problematic flows at those two memory sizes under random loss. A null on the first two checks is also informative, because it would bound the ionospheric contribution to LEO broadband variance at the latitudes where that contribution should be largest.
References
- T3P: Demystifying Low-Earth Orbit Satellite Broadband
- Anomalies of the Ionospheric F<SUB>2</SUB>-layer above Guangzhou Region
- Ionospheric TEC Prediction Base on Attentional BiGRU
- Forecasting Ionospheric foF2 Based on Deep Learning Method
- Effects of Equatorial Plasma Bubbles on Multi-GNSS Signals: A Case Study over South China
- Precise point positioning with an adaptive geometry-free cycle slip detection threshold under complex ionospheric conditions
- Statistical Modeling of PPP-RTK Derived Ionospheric Residuals for Improved ARAIM MHSS Protection Level Calculation
- Passive estimation of TCP round-trip times
- TimeWeaver: Opportunistic One Way Delay Measurement via NTP
- Memory-Efficient Performance Monitoring on Programmable Switches with Lean Algorithms
- Sketching the delay: tracking temporally uncorrelated flow-level latencies