Generative AI for hydrodynamical simulations:

2D, 3D, or 6D galaxy models?

SKA research at
Zurich University of Applied Sciences (ZHAW)

Centre for Artificial Intelligence (CAI)
Institute for Business Information Technology (IWI)
June 10, 2024
contact_qr.png Philipp Denzel, Frank-Peter Schilling, Elena Gavagnin

Slides on my website

https://phdenzel.github.io/

talk_qr.png

Outlook

Recap:
Generative models
for map-to-map translation

Dataset from IllustrisTNG

  • projected IllustrisTNG galaxies
  • 6 domains:
    • dark-matter, stars, gas,
      HI, temperature, magnetic field
  • ∼ 2'000 galaxies, (across 6 snapshots)
  • ∼ 360'000 images
  • each galaxy \(\ge\) 10'000 particles
  • augmented: up to 5x randomly rotated
  • scale: 2 dark-matter half-mass radii

domains.png

Dataset from IllustrisTNG

  • projected IllustrisTNG galaxies
  • 6 domains:
    • dark-matter, stars, gas,
      HI, temperature, magnetic field
  • ∼ 2'000 galaxies, (across 6 snapshots)
  • ∼ 360'000 images
  • each galaxy \(\ge\) 10'000 particles
  • augmented: up to 5x randomly rotated
  • scale: 2 dark-matter half-mass radii

domains_directions.png

Generative model architectures


Benchmark of generative models we're investigating and comparing:

cGANs

pix2pix_schema.png

Figure 1: pix2pix scheme

skais_gas_dm_data_pred_gt.png

Figure 2: cGAN(Gas) → DM: data, prediction, and ground truth (from top to bottom)

Score-based diffusion (SDM)

diffusion.png

Figure 3: Score-based diffusion: Song et al. (2021)

Noise schedule

skais_indi_rnd_dm_no_formula.png

Inversion by Direct Iteration (InDI)

skais_indi_gas_dm.png

Figure 4: InDI's iteration scheme following Delbracio & Milanfar (2023)

Diffusion Mamba (DiM)

diffusion_mamba.png

Figure 5: DiM architecture Teng et al. (2024)

From 2D to 3D models

  • observations inherently have 2D spatial resolution
  • astrophysical structures are inherently 3D
  • modelling difficulties:
    • inherent 3D features, different 2D perspectives
    • degeneracies
    • computational costs, …

Inherent 3D shapes

M83_ap190629.jpg
APOD 2019 June 29: M83

NGC4565_ap240606.jpg
APOD 2024 June 6: NGC 4565

Degeneracies


original image

J0753_kappa.png


reconstruction

SDSSJ0753+3416.png


⟶

J0956_kappa.png


⟶

SDSSJ0753+3416_recon.png

J0029_kappa.png
All valid model solutions: Denzel et al. (2021)

Point-cloud models for 3D modelling


Data type: point cloud

x y z [Mass] [E]
4 8 1 - -
5 1 6 -  
2 3 4 2 -
3 4 3 5 -
5 9 1 3 -
9 6 9 4 -
… … … … -

For Cosmological Inference

For Emulation of DM simulations (Quijote)

Experiments using transformers

  • AdaPoinTr architecture
  • Task: point cloud "completion"
  • Limitations:
    • input: max. ~10'000 particles
    • output: max. ~16'000 particles
      ⟶ Iterative generation
      ⟶ subsampling input
  • by Master student: Raphael Emberger

The "Good"

adapointr_dm_gas_1.png

Figure 8: AdaPoinTr (Yu et al. 2023) on TNG50 galaxies: DM → gas

adapointr_dm_gas_hist_1.png

Figure 9: Radial profiles of particle numbers

The "Bad"

adapointr_dm_gas_2.png

Figure 10: AdaPoinTr (Yu et al. 2023) on TNG50 galaxies: DM → gas

adapointr_dm_gas_hist_2.png

Figure 11: Radial profiles of particle numbers

The "Ugly"

adapointr_dm_gas_3.png

Figure 12: AdaPoinTr (Yu et al. 2023) on TNG50 galaxies: DM → gas

adapointr_dm_gas_hist_3.png

Figure 13: Radial profiles of particle numbers

Towards "Phase-space-point" models

  • expand feature vector to: mass, momenta/velocities, potential, …
  • problems:
    • already barely computationally tractable

    • more particles needed for accuracy

      ⟶ optimization: quantization, pruning, data parallelism, sharding, …
      ⟶ better subsampling strategies
      ⟶ self-consistency checks? regularizations?

Contact

https://phdenzel.github.io/

talk_qr.png



Email: philipp.denzel@zhaw.ch

References

Created by phdenzel.