Unified Framework
Self-supervised learning, weak alignment, supervised bottlenecks, multimodal fusion, and domain-invariant encoding.
University of Geneva cross-faculty programme
AI Foundation Models for Scientific Discovery
FUNDIS develops a unified, information-theoretic foundation-model framework for reliable scientific discovery in high-volume, heterogeneous domains where labels are scarce, noisy, or expensive.
4faculties and institutes
10foundation and application sub-projects
1shared scientific AI platform
Openreproducible workflows where agreements permit
Scientific objective
Transform pre-trained models into scientific instruments.
FUNDIS connects computer science, astronomy, particle physics, statistics, econometrics, digital humanities, and global studies through a shared model stack.
The programme uses SKA interferometry, JWST/EUCLID galaxy evolution, ATLAS event dynamics, weather and climate forecasting, cultural evolution, and global governance as stress tests for a common representation layer.
Operating model
From scientific data to validated discovery workflows.
Data
Self-supervision
Latent hierarchy
World model
Scientific task
Validation
Work packages
Four foundations, six scientific applications.
Hierarchical Latents & Information Bottlenecks & Reasoning
Minimal sufficient statistics for scarce-label scientific tasks.
World Models & Dynamics
Latent-space operators for dynamic reconstruction, anomaly detection, and counterfactual simulation.
Memory & Bio-Inspired Reasoning
Stable latent dynamics, long-range dependencies, and interpretable routing.
SKA & Radio Interferometric Imaging & Inference
Latent representations of sparse visibilities, reconstruction, uncertainty, detection, classification, and segmentation.
First Galaxies & Observational Cosmology
JWST/EUCLID morphology embeddings, redshift-aware latents, galaxy-evolution forecasting, and parameter extraction.
ATLAS Event Dynamics
Particle-shower world models, anomaly detection, and detector optimisation.
Weather & Climate Forecasting & Causality
Multi-scale forecasting, extreme-event modelling, causal interpretation, uncertainty, and robustness.
Cultural Evolution
Semantic drift, visual change, handwriting recognition, and stylistic/cultural pattern analysis.
Global Governance
Geopolitical and governance dynamics, policy diffusion, multilateral interactions, negotiation, and computational diplomacy.
Delivery
Shared infrastructure with measurable outputs.
Deliverables
- Shared PyTorch/JAX model library
- Benchmarks across all application domains
- Reusable latent spaces, world-model operators, and bottleneck modules
- Model cards, validation reports, and reproducible pipelines
Governance rhythm
- Monthly SP1-SP4 and SP5-SP10 integration meetings
- Quarterly steering committee reviews
- Shared documentation and validation reports
- Co-supervised PhD and postdoctoral positions
News
Seminars, papers, datasets, and platform releases.
Seminars
Talks, invited lectures, and reading groups.
Papers
Preprints, journal articles, and reports.
Datasets
Benchmark releases and documentation.
Software and models
Code, weights, and reproducibility packages.
Collaboration
Work with centers, research groups, and individual researchers.
Research centers
Joint programmes and strategic links.
Research groups
Shared benchmarks and cross-domain pilots.
Individual researchers
Focused collaborations around datasets and methods.
Infrastructure partners
Compute, repositories, data governance, and open science.
Scientific Advisory Group
External scientific guidance for quality, relevance, and reach.
Members to be announced
Senior researchers and domain experts will advise FUNDIS on scientific priorities, methodological rigor, responsible AI, and international collaboration.
Scope
Scientific strategy, technical evaluation, domain relevance, benchmark design, and external connections.
Consortium
A cross-faculty team at UNIGE.
Slava Voloshynovskiy
Computer Science, Faculty of Science
Francois Fleuret
Machine Learning, Computer Science
Roland Bouffanais
Computational Diplomacy, Computer Science
Taras Holotyak
Digital Forensics and Vision, Computer Science
Tobias Golling
Experimental Particle Physics, ATLAS
Sebastian Engelke
Statistics and Data Science, GSEM
Stefan Sperlich
Statistics and Econometrics, GSEM
Beatrice Joyeux-Prunel
Digital Humanities, Faculty of Letters
Stanislav Smirnov
Mathematics, Faculty of Science
Daniel Schaerer
Astronomy, Faculty of Science
Stephane Paltani
Astronomy, Faculty of Science
Pascal Oesch
Astronomy, Faculty of Science
Contact
Coordinate a dataset, benchmark, seminar, demonstrator, or collaboration.
FUNDIS is organised for practical collaboration between domain scientists and foundation-model teams.