Physics-Informed Neural Networks vs. SGP4 for orbit prediction of SAOCOM-1A under variable solar activity.
Master's thesis (Trabajo Fin de Máster) — M.Sc. in Artificial Intelligence, Universidad Internacional de La Rioja (UNIR). Author: Franco Bertoldi Mariglio · Status: in progress
This thesis empirically compares three approaches for predicting the future orbit of the Argentine synthetic-aperture-radar satellite SAOCOM-1A (NORAD ID 43641, operated by CONAE / built by INVAP), using only public catalog data:
- SGP4 — the analytical propagator that is the industry standard (baseline).
- MLP — a plain multilayer perceptron with no physics term (experimental control).
- PINN — a physics-informed neural network whose loss combines data with the residual of the perturbed equations of motion (J2 + exponential atmospheric drag). This is the focal technique.
Predictions are evaluated at 1-, 3- and 7-day horizons, segmented by level of solar activity (via the F10.7 index), since atmospheric drag — and therefore orbital decay — grows with solar activity.
- H1 (modest): under nominal solar activity (F10.7 < 120 sfu), the PINN does not beat SGP4 on mean along-track error for 1–7 day horizons. SGP4 has been tuned for decades and is hard to beat in its comfort zone.
- H2 (the bet): under high solar activity (F10.7 > 150 sfu), the PINN reduces mean along-track error by at least 15% vs. SGP4, with statistical significance (paired Wilcoxon, p < 0.05).
- H3 (control): the physics-free MLP underperforms both SGP4 and the PINN in all conditions, confirming that any PINN advantage comes from the physics term, not raw network capacity.
Every outcome — confirmation, refutation, or partial result — is a valid thesis contribution. Methodological honesty takes priority over expectations: nothing is assumed before it is measured.
| Source | Content |
|---|---|
| CelesTrak | Current GP element sets (TLEs) for SAOCOM-1A |
| space-track.org | Full TLE history (~18 months for this work) |
| CelesTrak SpaceWeather | F10.7 and Ap solar/geomagnetic indices |
| EGM96 / EGM2008 (NASA GSFC) | Earth gravity field coefficients |
sgp4 (Python) |
Vallado's official SGP4 implementation |
- Reproducible public-data download pipeline (TLEs + space weather)
- Exploratory data analysis, including interactive 3D Earth + orbit visualizations
- High-fidelity reference propagation (RK45) for ground truth
- SGP4 baseline evaluation
- MLP and PINN models, training, and statistical comparison
Python · NumPy / pandas · PyTorch (planned) · DeepXDE (planned) · sgp4 · SpiceyPy (planned) ·
Matplotlib / Plotly · Jupyter.
data/ # downloaded data (raw/processed are gitignored; managed with DVC later)
notebooks/ # EDA and experiments
scripts/ # public-data download utilities
docs/ # proposal and project notes
This README is in English for accessibility; project working notes are in Spanish.