interTwin Use Case

A particle detector data-driven Digital Twin for High-Energy Physics

Setting up a Digital Twin for fast particle detector simulation

About the role of data simulation in HEP

Particle detectors measure different particle properties at colliders such as  the Large Hadron Collider (LHC). More specifically, calorimeters are key components of the four LHC detectors, which are responsible for measuring the energy of the particles. In a collider, the emerging particles travel through the detector and interact with the detector material through the fundamental forces. Within electromagnetic or hadronic calorimeters, showers of secondary particles are created due to the interaction of each new particle with the dense calorimeter material.

This secondary particle creation is a complex stochastic process which is then typically modeled using Monte Carlo (MC) techniques. These simulations have a crucial role in High Energy Physics (HEP) experiments, and at the same time are very resource-intensive from a computing perspective: recent estimates show that the HEP community devotes more than 50% of the WLCG computing Grid (which has a total of 1.4 million CPU cores running 24/7/365 worldwide) to simulation-related tasks.


Moreover, Monte Carlo simulations are constrained by the need for accuracy, which will further increase, in the near future with the HL-LHC, which will increase the complexity of the associated detector data. As a consequence, HL-LHC  will increase the demand in terms of simulations, and consequently the need for computing resources.

CERN’s interTwin use case concerns a fast particle detector simulation paradigm using Generative Adversarial Networks (GANs). This use case incorporates a convolutional GAN, which we call 3DGAN, as calorimeter detectors can be regarded as huge cameras taking 3D pictures of particle interactions. The three-dimensional calorimeter cells are generated as monochromatic pixelated images with pixel intensities representing the cell energy depositions. A prototype of the proposed 3DGAN approach has already been developed.

Monte Carlo simulations are essential in HEP data analysis and detector design. A particle detector data-driven DT that accelerates simulations will add value in future HEP experiments, impacting many more domains, such as nuclear medicine and astrophysics. 



With upcoming LHC upgrades, the data simulation throughput will have to increase significantly. However, available computing resources will not increase accordingly.

Simulations play a crucial role in High Energy Physics (HEP) experiments: from the development of the detector to the design of strategies for data analytics, and the comparison of results to theoretical models. With the future High Luminosity upgrade of the LHC (HL-LHC) , a major increase in luminosity and dataset size is expected, thus offering the possibility to probe rarer phenomena more accurately. The need for simulated data will increase enormously: the expectation of the LHC experimental collaborations is that simulation throughput will need to increase by a factor of around 100 compared to today. However, the computing resources available for this are not expected to increase. 


The secondary particle creation is a complex stochastic process which is then typically modeled using Monte Carlo (MC) techniques. These simulations have a crucial role in High Energy Physics (HEP) experiments, and at the same time are very resource-intensive from a computing perspective. 


3DGANFastSim: Fast particle detector simulation with GAN

The simulation toolkit that has been developed and performs particle physics simulations based on MC methods and is also used for this purpose in this  use case, is Geant4. It provides a highly flexible simulation framework in C++. Moreover, Geant4 is used by large scale experiments and projects from the domains of nuclear medicine and astrophysics.

The efficiency of general-purpose Monte Carlo algorithms decreases dramatically when the simulations need to take place near critical points due to critical slowing down.  The community has long since started developing faster alternatives to Monte Carlo, including deep learning-based techniques. 

In the calorimeter case, deep learning based fast simulation directly generates the detector output, without reproducing, step by step, each single particle that interacts with the detector material. More specifically, generative models have been used in related HEP applications, as they are able to combine deep learning with statistical inference and probabilistic modeling. Inference with generative models is much faster than Monte Carlo. 

In the interTwin project, we train a third-party network using GAN-generated data and prove that the response is similar to a network trained with data from the Monte Carlo simulation. The showers generated by GAN present accuracy within 10% of Monte Carlo for a diverse range of physics features, with three orders of magnitude speedup.



The digital twin application’s workflow  consists of two different components, integrating the particle detector simulation framework and the GAN-based framework.

Within interTwin, we are currently working on the integration of these two components, to develop an end-to-end application, where each component feeds the other with data. The final goal is to simulate and optimize particle detectors.

Example of a GEANT4 electron event (left) vs. an event generated by GAN (right) for the same initial conditions

The level of integration promised by interTwin can lead to an unprecedented use of data driven tools for high-energy physics simulations. Overall it will contribute to improving the accuracy of particle collision representations and increase the possibility of new discoveries.

 Sofia Vallecorsa, Researcher CERN


CERN, the European Organization for Nuclear Research, is one of the world’s largest and most respected centres for scientific research.