Slides on my website

https://phdenzel.github.io/

Link/QR code to the slides for later or to follow along

Current status

• ZHAW still only has 1 SKACH project: deep learning for SKA 🔗
  • hydrodynamical simulations   ⇿   SKA mock observations
  • more projects to come…

Goal

• compress the knowledge from hydrodynamical and mock simulations to
  
  • map properties from simulations to mock observations
  • infer (hidden) astrophysical properties from observables
• explore the usability of various deep learning techniques
  for scientific (high-precision) data

Hydrodynamical simulations

• cosmological & astrophysical processes from first principle
• latest simulations reach (almost) petabyte sizes   ⇾   ideal for deep learning
  • IllustrisTNG, Simba, FIRE, EAGLE, Phoebos, and others

• dark matter
• gas (HI, HII, H₂, He, etc.)
• velocities
• stars
• temperature
• metallicity
• turbulence
• magnetic field strength
• X-ray luminosity
• Credit: IllustrisTNG Collaboration

Last time: CycleGAN

Zhu et al. (2017)

• two generator - discriminator pairs
• learn the mapping from domain A   ⇿   B and vice versa

CycleGAN experiments

• dataset: roughly 10'000 galaxies from Illustris TNG50-1
• brightness temperature of the gas   \(T_b(\mathbf{x}) = 189 h \frac{H_0}{a^2H(a)} \frac{\rho_{\text{HI}}(\mathbf{x})}{\rho_c}\,\text{mK}\)

Problem with training GANs

• GANs: \(\quad \mathbb{E}_{x\sim p_\text{data}}[\log{D_\theta(x)}] + \mathbb{E}_{z\sim q(z)}[1-\log{D_\theta(G_\theta(z))}]\)
  • fast inference and high quality results
  • implicit density and difficult to diagnose
  • mode collapse ⇾ not so much an issue for conditional GANs (such as Pix2Pix)
  • vanishing gradients ⇾ regularization (trades quality for stability)

Failure mode

Pile of data ⇾ AI system

More parameters, better models?

• hype over generative models: GPT-4, Vicuna, Stable Diffusion, etc.
  • larger, more complex   ⇾   better
  • sidenote: No Moat (interesting article on this topic)
• better: adjust the complexity of your model
  to the size of your dataset and task at hand

Pix2Pix

• pix2pix by I. Phillipi (originally in lua torch)
• blog post by Ch. Hesse (difference: Pix2Pix & CycleGAN)

Domains

Pix2Pix vs. CycleGAN

• tested on a set of 500 TNG50-1 galaxies
  • evaluation metric: \(\chi_{\nu}^{2} = \frac{(D_{i,\text{model}} - D_{i,\text{data}})^{2}}{N\sigma_{i}^{2}}\)
    (L2 loss normalized with Poisson noise)

Future plans

• better systematics with Karabo
• compare with actual strong gravitational lensing results
• integrate normalizing flow and diffusion networks