Grey-box dynamical systems
Hybrid models that combine incomplete ODE/PDE physics with neural components, staying interpretable while matching real observations.
PhD candidate · Machine Learning · University of Geneva
Gurjeet Sangra Singh
I build generative models for dynamical systems that keep physics in the loop, complete imperfect simulators, and infer what we cannot observe directly.
01 About
I am a PhD candidate at the University of Geneva. I am supervised by Prof. Alexandros Kalousis and Prof. Stéphane Marchand-Maillet, and working in the Data Mining and Machine Learning (DMML) group lab, led by A. Kalousis.
My work sits between dynamical systems, conditional generative models and simulation-based (amortised) inference for inverse problems. Recently I have been applying these ideas to weather and climate modelling. I am also a data scientist: turning raw data into meaningful, structured information that supports decisions.
My PhD is funded by MIGRATE, a geothermal-exploration project bridging seismology and machine learning. Before Geneva, I did my MSc in Data Science between the University of Padova and the University of Helsinki, where I wrote my thesis with Luigi Acerbi and developed PyBADS.
02 Research
Conditional generative models
Flow matching and diffusion models as simulation-free learners of stochastic, multi-modal dynamics.
Simulation-based inference
Amortised posterior estimation for inverse problems — including inference under model misspecification.
Weather & climate
Physics-informed flow matching for stable, long-horizon forecasting on ERA5.
03 News
-
Soon releasing another work, stay tuned...
-
Climate Physics Dynamic Matching accepted at the ECCV 2026 workshop on AI for Climate and Conservation. arXiv
-
Presented Climate Grey-Box Flow Matching for Robust Climate and Weather Prediction at the EGU General Assembly 2026 in Vienna.
-
Variational Grey-Box Dynamics Matching accepted at AISTATS 2026. Code
-
MIGRATE workshop Computational Earth (Elba Island, 12–17 Oct), giving a talk on automating dispersion-curve picking for nodal ambient-noise tomography.
-
Presented Deep Grey-Box Meets Optimal Transport at the ICLR 2025 XAI4Science workshop.
-
PyBADS published in the Journal of Open Source Software.
-
PyBADS 1.0 released on PyPI.
04 Selected publications
Variational Grey-Box Dynamics Matching
The 29th International Conference on Artificial Intelligence and Statistics (AISTATS 2026)
05 Funding
Project
MIGRATE — Multidisciplinary and InteGRated Approach for geoThermal Exploration
Machine learning for passive seismic exploration of the Earth's upper crust, towards lower-risk geothermal energy.