BiMA: Bias-Corrected LightGBM Model for Atmospheric Density Forecasting

Bias-Corrected Atmospheric Density Forecasting Using NRLMSIS 2.1 and LightGBM for the MIT ARCLab Prize for AI Innovation in Space 2025 Challenge

Setup  •  Usage  •  Examples  •  Model Report  •  Challenge Page  •  Dataset  •  Leaderboard

Project structure

├── configs/                    # Model experiment config
├── data
│   ├── interim/                # Intermediate processed data
│   └── raw/                    # Raw data of the challenge
├── models/                     # Final model training directory
├── reports                     # Model reports and images
│   ├── images/
│   └── model_report.pdf
├── runs/                       # Cross-validated model artefacts and log
├── src                         # Main source code
│   ├── config.py 
│   ├── dataset.py  
│   ├── eval.py           
│   ├── features.py         
│   ├── models.py
│   ├── qa.py
│   ├── utils.py
│   └── viz.py
├── model_performance.ipynb
├── submission.py
├── train.py
├── train.sh
└── validate.sh

Setup

conda create -n orb python=3.10
conda activate orb
pip install -r requirements.txt

Usage

Model training was executed in Windows/WSL environment

Training

1. Download STORM-AI Phase 1 Public Training Dataset and locate under data/raw with this structure:

   └── data
       └── raw
           ├── meta
           │   ├── 00000_to_02284-initial_states.csv
           │   └── ...
           ├── omni2
           │   ├── omni2-00000-20000603_to_20000802.csv
           │   └── ...
           └── sat_density
               ├── champ_-00000-20000802_to_20000805.csv
               └── ...
   

2. Generate combined training dataset (saved into data/interim)

   python -m src.dataset
   

3. [OPTIONAL] Validate the models using grouped satellite-CV (artefacts and log will be saved into runs/lag3det_final)

   # validate.sh
   python train.py -c raw_ratio -dm lag3det_final --save_prediction --validate_future --train_mode validate
   python train.py -c raw_ratio_log -dm lag3det_final --save_prediction --validate_future --train_mode validate
   

4. Train the models using all data (final models will be saved into models/lag3det_final)

   # train.sh
   python train.py -c raw_ratio -dm lag3det_final --train_mode full_train
   python train.py -c raw_ratio_log -dm lag3det_final --train_mode full_train
   

Inference

Inference for public and private test data was executed in Codabench platform (tutorial).

Running this command below locally will execute the code in DEBUG mode (check if .gitignore exist or not)

python submission.py

Examples

Example of Model Forecasts for CHAMP satellite during 2003 Halloween Solar Storms, plot was generated from model_performance.ipynb