Goals and Brainstorming

[EiffL]

This issue is to document our overal goals, ideas, and related publications. I'll keep the top of the issue as a clean summarry of the discussion in this thread.

Goal of the Hackathon

Our goal is to infer information about the star formation history of a galaxy from observables such as its SED or morphology:

For this we will be using data from the Illustris TNG simulation.

For some more details, check out the day 1 slides

Relevant Publications

Astro:

• Nonparametric Star Formation History Reconstruction with Gaussian Processes. I. Counting Major Episodes of Star Formation, Iyer et al. 2019

ML:

• Scalable Gradients for Stochastic Differential Equations, Li et al. 2020
• Infinitely Deep Bayesian Neural Networks with Stochastic Differential Equations, Xu et al. 2021

Relevant tools and packages

• https://github.com/google-research/torchsde
• https://github.com/xwinxu/bayeSDE

[EiffL]

Follow ideas with @mhuertascompany

• Verify information content in image vs SED with supervised learning (TNG100)
• Train with different SED sampling (TNG100)
• Apply to data (trained on TNG100, run on obs. sample)
• Train on TNG100 SED and apply on TNG50 SED

Optional subcontract:

• Run standard SFH MCMC recovery on TNG100 data for comparison