Learning to Simulate Collisions: Foundation AI Models for Heavy ion Physics

by Manjunath Omana Kuttan

Thursday 13 Nov 2025, 16:00 → 18:00 Europe/Zurich

4/2-037 - TH meeting room (CERN)

Foundation AI models are at the forefront of modern AI research. Unlike conventional deep learning approaches designed for specific tasks, foundation models are large-scale, versatile frameworks that can be adapted to a wide range of downstream applications.

In this talk, I will introduce HEIDi [1,2] , the first deep generative framework capable of producing complete, event-by-event heavy-ion collision outputs up to five orders of magnitude faster than conventional UrQMD simulations. The particles in the generated events accurately reproduce the underlying UrQMD distributions of hadron momenta and multiplicities. Such an ‘AI clone’ of a physics model enables the direct computation of complex and computationally expensive observables from event-by-event data, providing a powerful alternative to Gaussian Process surrogates for Bayesian inference [3]. Moreover, the fully differential nature of the model opens the door to gradient-based approaches for inverse problems. 

The results so far are promising, suggesting that HEIDi-based frameworks could evolve into the first foundation models for heavy-ion collisions. I will discuss the current state of this research, the challenges ahead, and how generative AI offers a new paradigm for experimental analysis, theoretical modeling, and inverse problems in heavy-ion physics and beyond.

[1]. M. Omana Kuttan, K. Zhou, J. Steinheimer and H. Stoecker, arXiv:2412.10352 (accepted in Phys Rev C)

[2]. M. Omana Kuttan, K. Zhou, J. Steinheimer and H. Stoecker, arXiv:2502.16330 (accepted in Phys Rev C)

[3]. M. Omana Kuttan, J. Steinheimer, K. Zhou and H. Stoecker,  Phys. Rev. Lett. 131 2023 no.20, 202303