NASA Heliolab 2025 - Ionosphere-Thermosphere Twin, ML experiments

This repository contains the code and resources for the Ionosphere-Thermosphere Twin project, part of the NASA Heliolab 2025 Frontier Development Lab. The project focuses on using machine learning to analyze and predict ionospheric conditions.

Team: Linnea Wolniewicz, Halil Kelebek, Simone Mestici, Michael Vergalla

Faculty: Giacomo Acciarini, Atilim Gunes Baydin, Tom Berger, Frank Soboczencki, James Walsh, Bala Poduval, Umaa Rebbapragada, Olga Verkhoglyadova

Branch descriptions

main: LSTM training and models

ioncast-gnn: GNN training and models

ioncast-sfno: Spherical FNO training and models

Directories

/scripts contains all the dataset files (dataset_....py), event files (events.csv), model files (model_....py), train (run_....py) and evaluation (eval_....py) files, and utility files (util_....py).

/ionopy contains ionopy files (for the most up-to-date code refer to https://github.com/spaceml-org/ionopy).

/notebooks contains visualization and experimental notebooks.

/tests contains testing and experimental code.

/data contains events.csv.

Pre-requisites

• Set up environment using the branch's environment.yml and requirements.txt
• Download the dataset from this link: https://zenodo.org/records/18343833

Usage

In the main branch, you can run the IonoPy and the IonCast LSTM models.

The necessary environment can be created from requirements.txt.

IonoPy: Code for this model may be out of date, for the most up-to-date code refer to https://github.com/spaceml-org/ionopy)

• Code for creating the model is contained in the ionopy/ folder
• To train the model, run scripts/train_ionopy.py

IonCast LSTM:

• All relevant datasets, models, events, and utility code are contained in the scripts/ folder
• To run the model, run scripts/run.py
• To evaluate the model on test data, run scripts/run.py in test mode.

Training example

The following command runs the training script, assuming the current directory is /scripts and the data is stored in the directory /mnt/data. Results will be saved to the `/mnt/experiment-1' directory.

python run.py --data_dir /mnt/data --mode train --target_dir /mnt/experiment-1 --num_workers 4 --batch_size 4 --model_type IonCastConvLSTM --epochs 2 --learning_rate 1e-3 --weight_decay 0.0 --context_window 4 --prediction_window 4 --num_evals 4 --date_start 2023-07-01T00:00:00 --date_end 2023-08-01T00:00:00

Acknowledgements

This work is the research product of FDL-X Heliolab a public/private partnership between NASA, Trillium Technologies Inc (trillium.tech) and commercial AI partners Google Cloud, NVIDIA and Pasteur Labs & ISI, developing open science for all Humankind.

Licence

This project is licenced under the Apache 2.0 License - see the LICENSE file for details.

Copyright 2025-2026 NASA
Copyright 2026-2026 Trillium Technologies Inc