Space Weather Challenge

Welcome to the Space Weather Challenge repository. This project is part of the STORM-AI initiative, aiming to develop advanced AI algorithms for predicting space weather-driven changes in atmospheric density across low Earth orbit.

This main branch contains the model we submitted for Phase 1 of the competition. Alternatively, the best_model branch contains an architecturally identical model that outperforms this one overall, especially in the case of extreme solar weather.

Table of Contents

• Project Overview
• Repository Structure
• Getting Started
  • Prerequisites
  • Installation
• Usage
• Troubleshooting
• License

Project Overview

The objective of this project is to create an AI model that can nowcast and forecast atmospheric density variations caused by space weather phenomena. Accurate predictions are crucial for satellite tracking and orbit resilience modeling.

Repository Structure

The repository is organized as follows:

• Codabench_Submission/: Contains Phase 1 submitted model and associated files for Codabench Platform submission.
• devkit/: MIT Development toolkit with utilities and baseline models.
• ml_pipeline/: Directory for our training and processing modules
• .gitignore: Specifies files and directories to be ignored by Git. baseline model.
• data.dvc: DVC file used by Stonecutter for dataset version control
• environment.yml: Conda configuration for package dependency management

Getting Started

Prerequisites

Ensure you have the following installed:

• Conda
• Docker (optional, for containerized environments)

Installation

1. Clone the Repository:

   git clone https://github.com/justin2213/space-weather-challenge.git
   cd space-weather-challenge

2. Set Up a Virtual Environment:

   conda env create -f environment.yml
   conda activate myenv

Usage

1. Data Preparation:

   Ensure that the necessary datasets are available and properly configured.

   The Challenge dataset from Dropbox contains CSV files that include datestamps in the names. However, the loading functions assume those have already been removed. You can use ml_pipeline/preprocessing/truncate_date.sh to remove the datestamps.

   If you are using Stonecutter, when cloning the repo, the dataset will be downloaded and placed in the data/ folder via DVC.

2. Configure Environment Variables

   Create a .env file with variable DATA_PATH that points to where you are storing the dataset.

3. Model Creation:

   Create ml model as a class and store it as a python file inside the ml_pipeline/models/ folder

4. Model Training:

   Train models using the train_model.py file in the ml_pipeline/ directory or by running:

   cd ml_pipeline
   python train_model.py

   This will train a specified model using the parameters that you set inside the train_model.py file. Once trained, the model will be tested against test and validation data and run through a submission generator which will calculate the propagation score compared to the previous best model. You can optionally save this model as the best model inside the Codabench_Submission/ directory or inside of the ml_pipeline/models/saved directory. Each directory will contain the necessary files to submit to the Codabench platform.

5. Submission:

   To generate the zip file that is necessary for Codabench submissions

   cd <model directory name>
   rm -f submission.zip && zip -r submission.zip . -x "submission.zip"

Troubleshooting

Consult TROUBLESHOOTING.md if you encounter errors.

License

This project is licensed under the MIT License. See the LICENSE file for details.