Introducing the Differentiable Universe Initiative

What's the idea

  • Vast amounts of data produced by upcoming surveys
     
  • Controlling systematics is paramount
     
  • All observe the same Universe under various modalities

How do we make the most of the available data ?

What if we could jointly forward simulate these surveys, and use the simulator for inference?

One example:
Weak Lensing

This is a Hierarchical Bayesian Model

aka

a stochastic simulator with some unknown tuning parameters

Advantages and Difficulties

  • Provides a full model for the observations, beyond 2 or 3pt statistics, multiprobe
     
  • Many systematics can now be explicitly modeled as part of  the simulation
     
  • We need to perform inference in a very large number of dimensions

=> We need more  information out of the simulator  to make the inference tractable

Let's build a full end-to-end differentiable model of the Universe

Why now?

  • We are at the dawn of Stage IV surveys, it's a good time to work on novel methodologies
     
  • Significant technological advances, with autodiff framework (like TensorFlow) and hardware acceleration
     
  • Significant methodological advances in machine learning, and inference  methodologies which make these problems tractable
     
  • This is not a new idea,  BCCP members have worked on differentiable simulations and optimization-based inference for years

Why an Initiative?

  • This is a large interdisciplinary task, requires coordination and can use  all the person-power it can get.

     
  • Goals of the Initiative:
    • Develop open source and high-quality software tools
       
    • Foster an open, collaborative, and welcoming environment where people will want to contribute to the effort
       
    • Help coordinate our interdisciplinary efforts towards making tools that can actually be used to analyse surveys.
       

Where we are right now