Project 01 / 11
HelioGuard
Explainable solar-flare forecasting. My final-year research project — an explainable, full-stack machine-learning system that forecasts solar flares and tackles the hardest class to predict, intermediate M-class flares, with cost-sensitive learning and live SHAP explanations.
Why solar flares?
Solar flares are sudden bursts of magnetic energy from the Sun. The big ones disrupt power grids, damage satellites, degrade GPS and endanger high-altitude flights. Space-weather centres issue warnings so engineers can protect that infrastructure, but most machine-learning forecasters are black boxes. Operators have to decide whether to take costly defensive action without knowing why the model raised an alert.
The problem: the M-class ambiguity trap
Flares are graded by strength. Models separate quiet periods and extreme X-class flares fairly well, but M-class flares sit in between, and their physical parameters overlap with both neighbours. Recent multi-class models perform worst on exactly this class. Yet the M threshold is where radio-blackout warnings and satellite-safing procedures start.
Two things make it hard:
- Severe class imbalance. Flaring examples are vastly outnumbered by quiet ones (about 49:1). Standard training with symmetric losses like cross-entropy biases models towards the easy majority.
- Messy telemetry. Space-telescope data streams are noisy and full of gaps.
The approach
HelioGuard is an end-to-end, explainable forecasting framework built on the public SWAN-SF / SHARP benchmark: multivariate time series of solar active-region magnetic parameters from NASA’s SDO/HMI.
- Data guard. Unsupervised Isolation Forests filter outlier noise, and Pearson-correlation-based k-NN imputation fills missing values using physically correlated signals.
- Feature engineering. Highly predictive SHARP parameters, such as total unsigned current helicity, free-energy proxies and polarity-inversion-line measures, tracked over a window of historical frames.
- Cost-sensitive learning. Sequence models in PyTorch trained inside a Class-Dependent Reward (CDR) framework. A tuned penalty matrix punishes false alarms and missed flares asymmetrically to sharpen the M-class decision boundary without hurting X-class detection.
- Honest evaluation. Train/test splits grouped by NOAA active region (no leakage), scored with the field’s skill metrics: TSS, HSS, recall and false-alarm rate.
- Explainability you can use. SHAP attributions served through an asynchronous FastAPI layer to an interactive React + Tailwind dashboard, so operators can see which physical parameters drove each forecast.
Research questions
- How can a Class-Dependent Reward framework be tuned to resolve decision-boundary confusion for M-class flares?
- How far does combining Isolation Forests with correlation-based k-NN imputation stabilise models on noisy, multi-active-region telemetry?
- How can time-series explanations be visualised live in a dashboard to build operator trust?
Status
In progress as my final-year BSc (Hons) Computer Science project at IIT / University of Westminster. I’ll share results and lessons on the blog as they land.
Next project