You can switch the language using the menu below:
🌐 Languages: English | Português
Predicting solar flares is essential for mitigating space weather risks that impact critical technological infrastructures on Earth and in space. Solarfall is a machine learning pipeline designed to parse highly imbalanced and high-dimensional astrophysical data to predict these severe solar events. To handle the complexity of real-world space weather data, the project introduces a custom hierarchical cascade architecture built entirely with the XGBoost algorithm. This architecture progressively filters signal from noise through four specialized classification stages—ranging from a high-recall 'Gatekeeper' that flags potential anomalies to a deep 'Specialist MX' model trained specifically to isolate catastrophic, extreme-magnitude events.
- Machine Learning Engine: XGBoost for native class imbalance management, high-dimensional vector robustness, and tabular data scalability.
- Data Acquisition & Web Scraping: Automated data pipelines utilizing Selenium and FTP protocols to reliably extract operational space weather catalogs from NOAA/NCEI and JSOC/Stanford APIs.
- Hyperparameter Optimization: Optuna framework for the Bayesian optimization of tree structures and decision thresholds.
- Hardware Acceleration: CuPy library for direct VRAM allocation, enabling rapid iterative training via GPU acceleration.
Eduardo Oliveira Ferraz de Campos - Software Engineering Student at IFSP Câmpus São Carlos
Advisor: Sérgio Luisir Discola Junior