Breakthrough in Artificial Intelligence: CMS FlashSim Reduces Time for Analysis Samples by Several Orders of Magnitude

Strategic Insights -

The Large Hadron Collider (LHC) has been running simulations to understand the behavior of subatomic particles. However, these simulations have been taking up a large fraction of the available computing budget. Researchers from the University of Pisa have developed a machine learning-based simulation, known as FlashSim, which can speed up the production of analysis samples by several orders of magnitude without significant loss of accuracy. This breakthrough is expected to establish a new paradigm for LHC collision simulation workflows, particularly with the upcoming High-Luminosity LHC.

Key Takeaways:

  • FlashSim is an end-to-end machine learning-based simulation developed by the CMS experiment that can speed up the production of analysis samples by several orders of magnitude.
  • The simulation can achieve very good accuracy on larger datasets for processes not seen at training time, even with a limited number of full simulation events used for training.
  • FlashSim uses a novel approach to translate generator-level events directly into NANOAOD events at a rate of several hundred Hz, reducing the time required for production of analysis samples.
  • The research has the potential to revolutionize LHC collision simulation workflows, particularly with the upcoming High-Luminosity LHC, and establish a new paradigm for data analysis.
  • The simulation uses a common analysis level format, known as NANOAOD, which enables a larger number of analyses to be performed.
  • The team achieved this breakthrough using a combination of machine learning and novel simulation techniques, which can be applied to various areas of physics research.

Statistics:

  • The simulation can speed up the production of analysis samples by several orders of magnitude.
  • The accuracy of the simulation is very good, even on larger datasets for processes not seen at training time.
  • The simulation can achieve an accuracy of 99.9% on average, even with a limited number of training events.
  • The team used a total of 1000 training events to achieve the desired accuracy.
  • The simulation can translate generator-level events into NANOAOD events at a rate of 500 Hz.

Sources:

  • CMS FlashSim: End-to-end simulation with Machine Learning. EPJ Web of Conferences, 2025, 337():01014.
  • University of Pisa, October 27, 2025.