The Engineering Behind AuroraLens
AuroraLens began as a personal challenge — could a student with an interest in space weather build a genuinely accurate, location-specific aurora forecast using nothing but open NASA data and applied machine learning?
The answer required learning FastAPI for the inference backend, building a dual-stage XGBoost pipeline from scratch on the NASA OMNI2 dataset, integrating real-time DSCOVR telemetry, and designing a frontend that made complex geomagnetic data feel intuitive for non-scientists.
The 81.0% weighted F1-score across 1.2 million hours of test data is not the end goal — it is a baseline. The model is open source, documented, and designed to be improved by anyone who wants to contribute.
Mosin Mushtaq
B.Tech AI/ML Engineering
Space Weather Intelligence Platform
A dedicated infrastructure-focused monitoring system for geomagnetic storm detection, CME tracking, and real-time risk assessment for power grids, GPS networks, and satellite operators.
Technical Architecture
From deep-space telemetry to local forecasts—how the AuroraLens engine processes magnetic flux into predictive insights.
NASA/NOAA Data
Live DSCOVR Telemetry (L1 Point)
Data Pipeline
Preprocessing & Feature Engineering
XGBoost Engine
Dual-Stage Gradient Boosting
Forecaster HUD
Real-time UX Visualization
Inference Engine
XGBoost Dual-Stage
The engine utilizes two distinct Gradient Boosting models trained on the NASA OMNI dataset. Stage 1 determines geomagnetic activity probability, while Stage 2 classifies intensity.— Stage 1 output (Activity Probability) feeds Stage 2 as a primary classification feature.
Empirical Validation
Validated against 1.2M hours of NASA OMNI telemetry records, achieving an 81.0% weighted F1-score across KP classification bins.
View Project Artifactsarrow_forwardEngine Specifications
Dual-Stage XGBoost
Two-stage Gradient Boosting pipeline. Stage 1 predicts geomagnetic state probability; Stage 2 classifies intensity (KP-index).
Feature Engineering
Inputs include IMF Bz (GSM), Solar Wind Speed/Density, and local tropospheric cloud density updated every 60s.
81% F1-Score
Validated against 1.2M hours of NASA OMNI telemetry. Optimized for recall on high-intensity (G1-G5) storm events.