Interpretable Machine Learning for high-resolution astronomical spectroscopy.
We can combine stellar, telluric, and instrumental models into a unified forward model of your entire high-bandwidth, high-resolution spectrum. We can obtain best-in-class models of Earth's atmosphere, line-by-line, automatically, for free (or cheap).
By using autodiff, we can fit over 10,000 spectral lines simultaneously. This enormous amount of flexibility is unavailable in conventional frameworks that do not have autodiff.
^ We do this for 10,000 lines simultaneously.
We first clone a precomputed synthetic spectrum, such as PHOENIX, and then transfer learn with data. By regularizing to the cloned model, we get the best of both worlds: data driven when the Signal-to-Noise ratio is high, and model-driven when we lack data to say otherwise.
We achieve $>60 \times$ speedups with NVIDIA GPUs, so training takes minutes instead of hours.
Visit our step-by-step tutorials or installation pages to get started. We also have deep dives, or you can read the paper. Have a question or a research project in mind? Open an Issue or email gully.
Copyright 2020, 2021, 2022, 2023 The Authors