astro automata focuses on automating astrophysics research with AI. It is both a personal research blog and collection of related content.
Research (full list)
Rediscovering orbital mechanics with machine learning
Could we discover the law of gravitation without even knowing the masses of planets in the solar system? In this paper we show how.
Unsupervised Resource Allocation with GNNs
We show how to optimize resource allocation without knowing the true utility. We use this to learn a mock telescope observational survey from scratch.
Bayesian neural networks for planetary instability
We describe a Bayesian neural network architecture that can accurately learn to predict dynamical (chaotic) instability in compact planetary systems. The network demonstrates surprisingly robust generalization to 5-planet systems.
Discovering Symbolic Models from Deep Learning
We describe a technique for converting a deep learning model into an analytic equation, focusing on graph networks. We validate it on injected force laws and Hamiltonians, and then discover a new equation to accurately predict the overdensity of dark matter from its environment.
Lagrangian Neural Nets
We show how one can learn a Lagrangian from data with a Neural Network. Such an architecture conserves energy in a learned simulator without requiring canonical coordinates.
Normalizing Flows for Stellar Isochrones
A normalizing flow is a highly flexible probability distribution, perfect for the nonlinear surface of an HR diagram. Here, we show how to learn a flow for Gaia’s HR diagram.
Software (full list)
High-Performance Symbolic Regression in Python
A library for interpretable machine learning, using symbolic regression, in Python
Posts
An introduction to likelihood-free inference
I give an introduction to likelihood-free (simulation-based) inference for scientists.
The use and abuse of machine learning in astronomy
I argue that machine learning is both underused and overused in astrophysics.
Symmetries in Neural Networks
Summary of a Twitter discussion about symmetries in Neural Networks
Related articles
Powerful ‘Machine Scientists’ Distill the Laws of Physics From Raw Data
By Charlie Wood. How can machine learning help us discover new theories of physics, rather than simply fitting data?
How Self-Driving Telescopes Could Transform Astronomy
By Ryan F. Mandelbaum. What if an autonomously operating telescope, free from human biases and complications, could find the solutions we’ve been missing?
AI Will Help Scientists Ask More Powerful Questions
By Pushmeet Kohli. Self-learning systems can discover hidden patterns in immense data sets, transcending what humans could ever find on their own