2025 · Published

VIPER — Predicting equatorial vertical plasma drift from global TEC

Python MLP Monte Carlo Dropout Ionosphere JPL GIM C/NOFS

VIPER (Vertical dRIfts: Predicting EquatoRial dynamics) is a machine learning model that predicts day-to-day equatorial vertical plasma drift from publicly available GPS-derived total electron content (TEC) data — with built-in uncertainty quantification.

Paper (AGU)

Overview

Vertical plasma drift (v_z) governs the redistribution of ionospheric plasma at low latitudes through the E × B mechanism. It controls the altitude of the ionosphere, modulates the Equatorial Ionization Anomaly (EIA), and sets up conditions for space weather irregularities that disrupt GPS and communications. Measuring v_z directly requires in-situ satellites — rare, expensive, and limited in spatial coverage.

VIPER bypasses this bottleneck. It learns the relationship between TEC patterns in global ionospheric maps (GIMs) and simultaneous v_z measurements from the C/NOFS IVM instrument, then uses that mapping to predict drift at any longitude, any day — from ground-based GPS data alone.

Animated JPL Global Ionospheric Map showing the Equatorial Ionization Anomaly — An animated JPL Global Ionospheric Map (GIM). The twin TEC crests at ±15° magnetic latitude are the EIA — the signature that VIPER uses to infer vertical plasma drift.

Motivation

Empirical drift models exist, but they are climatological: they represent average behavior, not the day-to-day variability that matters most for space weather forecasting. Physics-based models can capture variability but require wind and chemical inputs that are rarely available in real time.

VIPER is the first model to predict daily v_z globally from TEC data alone, offering a path to real-time ionospheric nowcasting without in-situ measurements.

The model is capped at quiet geomagnetic conditions (Kp < 3) for this first release, making it a robust quiet-time baseline — and a test bed for future storm-time extensions.

Data & feature engineering

VIPER is trained on six years of data (2009–2015) spanning solar minimum through the ascending phase of solar cycle 24.

Inputs

JPL GIMs — 96 maps per day at 15-min cadence, converted to magnetic coordinates using ApexPy. Each day is averaged to a single 24-hr TEC map, then split into eight longitude sectors.
C/NOFS IVM v_z — in-situ vertical drift measurements from the Ion Velocity Meter, used as the training target.

EIA feature extraction

A novel Gaussian fitting routine identifies the EIA in each TEC profile and extracts three physically motivated features:

P2T (Peak-to-Trough) — EIA prominence; a proxy for E × B drift magnitude.
DOP (Distance of Peaks) — latitudinal separation of the northern and southern crests; relates to the height the plasma fountain reaches.
AOP (Asymmetry of Peaks) — hemispheric asymmetry driven by trans-hemispheric winds and production asymmetries.

Cyclical features (day-of-year, MLT, geographic longitude) are encoded as sine/cosine pairs to respect their periodicity. Solar flux (F10.7) and geomagnetic activity (SYM-H daily min, max, range) complete the input vector. Total training set: ~744 k samples (80/10/10 train/test/validation).

Model

VIPER is a four-layer Multi-Layer Perceptron (MLP) with ReLU activations, optimised via Bayesian hyperparameter search:

Layer widths: 1024 → 512 → 256 → 128
Dropout rate: 0.2
Learning rate: 10⁻³, stepped down by 0.6× every 10 epochs
Batch size: 128

Training ran on a 128-core cluster with a single NVIDIA A-100 GPU at UCL MSSL and converged in under 15 minutes.

Uncertainty quantification

VIPER uses Monte Carlo (MC) Dropout — dropout is kept active at inference time, so each forward pass produces a slightly different prediction. Running 500 passes per input yields a distribution over v_z, from which the Median Absolute Deviation (MAD) is reported as the prediction uncertainty. This means every VIPER output comes with a confidence estimate, without any additional computational overhead beyond the 500 samples.

The uncertainty distributions are non-Gaussian, so mean and standard deviation are not appropriate summary statistics — MAD is used throughout.

Results

Evaluated on a held-out test set of ~93 k samples:

MAE: 8.3 m/s
Pearson R: 0.89
Prediction Efficiency (PE): 0.78

VIPER captures both upward and downward drift with limited systematic bias. Error is consistent across MLT, season, and longitude — but doubles at 6 MLT compared to 15 MLT, reflecting sparse data at post-midnight local times when C/NOFS ion density falls below the IVM quality threshold. Performance is strongest during equinox (symmetric EIA) and degrades slightly when the AOP is large, which points to wind asymmetries as a remaining challenge.

What's next

Extending VIPER to storm-time conditions (Kp ≥ 3) where the EIA structure is more complex and data are sparser.
Daily operational predictions using near-real-time JPL GIM products.
Integrating VIPER outputs as boundary conditions for global physics-based ionospheric models.