HelioCast-L1

Inspiration

We rely entirely on a fragile technological ecosystem. A single, severe solar flare could instantly knock out our GPS, pull satellites out of orbit, and overload the power grids that keep our cities running. For decades, the biggest problem has been a lack of warning time. When ISRO launched the Aditya-L1 mission, placing a permanent observatory at Lagrange Point L1, it felt like an open invitation. We realized we could use India’s own space data to tackle this global vulnerability and engineer a literal planetary defense mechanism that predicts these storms before they reach Earth.

What it does

HelioCast-L1 is an automated early-warning space weather system. It constantly ingests live soft and hard X-ray data from the Aditya-L1 spacecraft and runs it through a predictive machine learning engine. Instead of just reacting to a flare that is already happening, our web dashboard detects early precursor patterns in the X-ray data and triggers a countdown alert. This provides satellite and power grid operators the precise lead time needed to safely power down and protect vital assets.

How we built it

We approached this as an end-to-end data engineering and machine learning pipeline, breaking it down into three core phases.

First, we pulled Level-1 time-series data directly from the ISSDC PRADAN portal, focusing on the SoLEXS (soft X-ray) and HEL1OS (hard X-ray) payloads. We wrote custom Python scripts to ingest the telemetry, strip out anomalous cosmic ray spikes, and time-sync the two distinct data streams.

Second, we grounded our forecasting engine in astrophysics—specifically the Neupert Effect. This principle states that the amount of soft X-rays released is roughly proportional to the buildup of hard X-rays over time during a flare's early phase. Using this relationship, we engineered temporal features to train a Long Short-Term Memory (LSTM) network. Our model analyzes a rolling window of recent X-ray activity and predicts the probability of a flare peaking within a specific future timeframe.

Finally, we wrapped the ML engine in a live web interface that dynamically plots both light curves and flashes a visual red alert the moment precursor signatures are detected.

Challenges we ran into

Working with raw satellite telemetry was a massive reality check. Time-syncing was our first major hurdle because SoLEXS and HEL1OS sample data at different rates and structures. We spent hours rewriting interpolation algorithms to align the data without corrupting the underlying science.

Next, we hit the "needle in a haystack" problem. Severe solar flares are extremely rare compared to normal solar background noise, which caused our initial models to throw terrible false positive rates. We had to implement Synthetic Minority Over-sampling Technique (SMOTE) and aggressively rewrite our loss functions to penalize false alarms.

Accomplishments that we're proud of

We successfully bridged the gap between theoretical astrophysics and applied deep learning. Taking raw, noisy Level-1 satellite data straight from an ISRO portal and turning it into a clean, predictive machine learning model in a highly constrained timeframe is an achievement our entire team is incredibly proud of.

What we learned

We learned exactly how messy and unforgiving raw space data can be, and how to optimize time-series models for real-time anomaly detection rather than just static datasets. Most importantly, we learned how to design a mission-critical system where reliability and speed matter significantly more than flashy UI design.

What's next for HelioCast-L1

Moving forward, we want to integrate additional solar telemetry—such as magnetogram data and extreme ultraviolet imaging—to further increase the lead time and accuracy of our forecasts. We also plan to build a dedicated API so that other developers and grid operators can plug our real-time flare predictions directly into their automated safety protocols.

Built With

• celery
• django
• django-rest-framework
• docker
• framer-motion
• github
• matplotlib
• next.js
• numpy
• pandas
• plotly.js
• postgresql
• python
• pytorch
• react
• redis
• scikit-learn
• scipy
• seaborn
• supabase
• tailwind-css
• tensorflow
• typescript