The AstroSpectro project journey reached a major milestone today. What began as a modest baseline with 32% accuracy has become, after a series of rigorous scientific iterations, a high-performance stellar classifier reaching 84% accuracy. This post retraces the key steps, discoveries, and lessons learned along the way.
I built a pipeline that models spectral classification based on the principles of stellar physics. The model autonomously learned that Calcium and Hydrogen line-strength ratios are the best temperature indicators, validating our feature engineering approach and allowing us to reach 56% accuracy.
Since the beginning of the AstroSpectro project, the notebook [02_tools_and_visuals.ipynb](https://github.com/PhD-Brown/AstroSpectro/blob/main/notebooks/02_tools_and_visuals.ipynb) served as an exploratory notebook. It was where I tested spectrum visualizations, line detections, and diagnostics. But as the tools accumulated, the code became monolithic and difficult to maintain.
Today marks a crucial milestone for the AstroSpectro project: the first end-to-end training run of our stellar type classification model. After setting up a robust acquisition and processing pipeline for 5000 spectra from the LAMOST DR5 survey, it was time to see what our first approach could achieve.
Today marks a major milestone for the AstroSpectro project: the structure and core content of the documentation for version v0.1.0-alpha are officially complete!