Astronomical Imaging and
Hardware Design with
Differentiable Optical Models

Benjamin Pope, University of Queensland

with PhD student Louis Desdoigts (University of Sydney)

Slides available online at

benjaminpope.github.io/talks/uqai

Imaging Exoplanets

Simulated images of a star with JWST: ideal (left) and with misaligned mirrors (right).

Computational Imaging in Astronomy

• Image Analysis
  • phase retrieval
• Optical design
  • coronagraphs

Optics as analogy to neural networks

• Optical systems are made of layers:
• Each layer is a global linear transformation (propagation)
  • or elementwise operation (eg a mask)

Physics-Informed Neural Networks

dLux - Differentiable Optical Models in Jax

Define an optical system just like a neural network:

optical_layers = [
    dl.CreateWavefront    (wf_npix, aperture),
    dl.TiltWavefront      (),
    dl.CircularAperture   (wf_npix),
    dl.NormaliseWavefront (),
    dl.PhysicalMFT        (det_npix, fl, det_pixsize)]

Built in Jax and equinox.

Examples

With Louis Desdoigts we have been working on a few examples: see the docs!

• HMC for simultaneous phase retrieval and deconvolution
• Designing toward an optimal Fisher Information Matrix

Future

Submitting v1.0 soon

• Want to collaborate!
  • with astronomers
  • with hardware designers
  • with other AI researchers

• Applications in:
  • microscopy
  • laser physics
  • your science?