Skip to content

AISSAI Hackathon 2025

Date: November 3 to 7 - 2025

Location: Seolane - Barcelonnette, Alpes de Haute Provence, France

Hackathon Organization

All hackathon materials and organization details are available on GitHub:

Breakout Sessions

We organized two technical breakout sessions during the hackathon:

Hackathon Projects

Symbolic Regression for Stellar Systems

Principal Investigator: Paolo Bianchini
Repository: 2025-phi-so
Focus: Discovering interpretable physics from N-body simulations

Project Overview

Using symbolic regression (Phi-SO) to discover interpretable analytic relations from N-body simulations of globular clusters evolved over 14 Gyr. The goal was to identify fundamental physical mechanisms by deriving analytical formulae for global properties evolution.

Hackathon Results

The hackathon produced three successful working groups, each discovering analytical expressions for different aspects of stellar system dynamics:

Working Group 1: Velocity Dispersion-Mass Relationship
- Confirmed theoretical predictions for evolved clusters: velocity dispersion follows $\sigma_\phi \sim 1/\sqrt{m}$ - Provided direct evidence of mass segregation where heavier objects sink to the cluster core - Code | Presentation

Working Group 2: Global Mass Evolution
- Successfully derived analytical expressions for total mass loss over 14 Gyr - Quantified the impact of stellar evolution and dynamical effects on cluster dissolution - Code | Presentation

Working Group 3: Radial Velocity Profiles
- Derived analytical formulae for azimuthal velocity and velocity dispersion as functions of radius - Performed cross-cluster analysis and temporal evolution studies - Code | Presentation

Final Presentation: Google Slides

Tools & Methods

Foundation Models for Euclid + DESI Analysis

Principal Investigators: Marc Huertas-Company and Malgorzata Siudek
Repository: 2025-fondation-models
Focus: Multimodal embeddings analysis and anomaly detection

Project Overview

Analysis of embeddings from multimodal foundation models (AstroClip, astroPT, AION) applied to Euclid Q1 images and DESI spectra. The project explored how different architectures encode physical galaxy properties and their potential for anomaly detection.

Hackathon Results

Models Compared: - AION: Foundation model with separate spectra-only, images-only, and fused embeddings - astroPT: Transformer model with joint training producing unified embedding space - AstroCLIP: Contrastive image-spectrum alignment (training issues identified)

Key Findings: - Fused embeddings improve separation for age/mass-sensitive features (DN4000, stellar mass) compared to single modalities - Spectra emphasize stellar population/age indicators; images emphasize morphology - Complementary structure when combining modalities - Successfully separated spectral classes (GALAXY / QSO / STAR)

Analyses Performed: - UMAP dimensionality reduction colored by physical properties (mass, SFR, DN4000, colors) - Morphological property analysis (ellipticity, concentration, asymmetry, smoothness, Gini, Sérsic) - Cross-model correlation benchmarking - Cosine similarity analyses between modality-specific and fused embeddings

Dataset: - Working sample: ~15k matched Euclid cutouts + DESI spectra - Planned expansion: ~40k → DR1 scale for improved statistics

Deliverables

  • aion_embeddings_analysis.ipynb: AION multimodal model analysis
  • astropt_embeddings_analysis.ipynb: astroPT transformer analysis
  • model_comparison_astropt_aion.ipynb: Cross-model comparison

Future Directions: Isolation Forest, Local Outlier Factor, targeted anomaly recovery (lenses, dual AGN), fine-tuning AION encoder

Notes

  • Breakfast from 8:00 to 9:00
  • Morning tag up from 9:30 to 10:00 by group
  • Morning coffee break from 10:30 to 11:00
  • Lunch break from 12:30 to 14:00
  • Breakout sessions from 13:30 to 14:00
  • Afternoon coffee break from 15:30 to 16:00
  • Wrap-up session from 17:00 to 17:30 all together
  • Dinner from 19:00 to 20:30
  • Last day will devote to the final presentation of the hackathon