Map-to-map translation <br> <h2>of simulated galaxies with conditional GANs</h2>

Map-to-map translation

of simulated galaxies with conditional GANs

SKA research at
Zurich University of Applied Sciences (ZHAW)

Centre for Artificial Intelligence (CAI)
Institute for Business Information Technology (IWI)

Neuchâtel, 2024/01/22 Mon

Philipp Denzel, Frank-Peter Schilling, Elena Gavagnin

Slides on my website

https://phdenzel.github.io/

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

Projects at ZHAW

SKA project:
- trained (astro)physicists, focused on ML research
our expertise:
- deep generative modelling of (sky) simulations
- CV, DL, XAI, MLOps, …
recently expanded efforts
- two new projects

Figure 1: ZHAW's SKACH team at CSCS in Lugano

The times they are a-changin'

the end of the analytic era
modern surveys: galaxies are no longer blobs
rethink data analyses: analytic ⟶ data-driven

Deep Generative galaxy modelling

goal is to learn an implicit distribution \(\mathbb{P}\) from which
the training set \(X = \{x_0, x_1, \ldots, x_n \}\) is drawn

Analytic models \(\mathbb{P}_\theta\)

Figure 2: Credit: Mandelbaum et al. (2014)

Simulators \(\mathbb{P}_\theta\)

Figure 3: Credit: Schaye et al. (2023)

Implicit distributions

in both cases, we cannot
- sample from (the true) \(\mathbb{P}\)
- evaluate the likelihood \(p_\theta(x)\)
which means: we cannot generate new plausible galaxies
what for?

For instance: strong lensing

Figure 4: 2006, Credit: NASA, ESA, D. Player (STScI)

Strong lens modelling

input data

⟶

latent representation

⟶

reconstruction

Denzel et al. (2021)

Dataset: SPH simulations

B-field (TNG100), Credit: IllustrisTNG

projected IllustrisTNG galaxies
6 domains:
- dark-matter, stars, gas,
  HI, temperature, magnetic field
∼ 3000 galaxies
∼ 10000 images / domain
augmented:
- up to 5x randomly rotated
scale: 2 dark-matter half-mass radii

Figure 5: Use image domain translation models: observations (21cm) ↔ physical properties

cGANs: pix2pix schema

Figure 6: Use pix2pix to generate dark matter maps from mock observations

Sampling from \(\mathbb{P}_\theta\)

Ground truth

Predictions from pix2pix (from gas projections)

Evaluation

Next steps

physical metrics: radial/elliptical (NFW) profiles
substructure just above the resolution limit
still not able to evaluate \(p_\theta(x)\)
GANs: average performance expected to be slightly worse compared
to autoregressive and score-based methods

Point-cloud experiments

generative models for full 3D+ simulations

Property	SPH data	Point clouds
applications	hydrodynamics	3D scanning, CAD, etc.
list of coordinates	✓	✓
unordered	✓	✓
invariance: vector-row perm.	✓	✓
invariance: geometric transf.	✓	✓
discrete	~	✓
smoothing kernel	✓	✗

AdaPoinTr

initial tests indicate feasibility
application: DM-only simulation, generate baryonic particle types (stars, gas, etc.)

Yu et al. (2023)

Radio source classification

idea developed with Michele Bianco (EPFL)
student Manuel Weiss: tested SOTA classification & detection architectures
- ResNet, EfficientNet, ViT, etc. / YOLOv8, DINO, etc.
goal: testing on the GLEAM survey
Radio Galaxy Zoo Object Detection Data Set (11’836 labelled images)