DDFastShowerML

Library with utilities and plugins that allow running fast simulation in a Geant4 application using ML inference from within ddsim (DDG4).

References

1. Fast Simulation of Highly Granular Calorimeters with Generative Models: Towards a First Physics Application, P. McKeown et al., PoS EPS-HEP202 3 (2024) 568, DOI: 10.22323/1.449.0568

2. Development and Performance of a Fast Simulation Tool for Showers in High Granularity Calorimeters based on Deep Generative Models, P.McKeown, DESY-THESIS 186 pp. (2024), DOI: 10.3204/PUBDB-2024-01825

Quick start guide

Environment Setup

Access to cvmfs required.

The library can be run in apptainer with an Ubuntu 22.04 docker image as follows, with bind mounting to cvmfs:

mkdir cvmfs
export APPTAINER_CACHEDIR=<path_to_your_cachedir>
export APPTAINER_TMPDIR=<path_to_your_tmpdir>
apptainer shell --bind /cvmfs:/cvmfs docker://ghcr.io/key4hep/key4hep-images/ubuntu22:latest

Installation

Prerequisites:

• a key4hep release. This will provide:
  • DD4hep, Geant4, podio, EDM4hep, LCIO, ...
• ONNXRuntime
• libTorch.so/dylib

This setup can be achieved with the setup script provided:

source setup_env_nightly.sh

Now build as usual

mkdir build
cd build
cmake ..
make -j4 install

and ensure to set up the library correctly

source ../install/bin/thisDDML.sh

Running the example

The simulation can then be run as usual with ddsim- for example for the ILD detector:

cd ../scripts
ddsim --steeringFile ddsim_steer.py --compactFile $k4geo_DIR/ILD/compact/ILD_l5_o1_v02/ILD_l5_o1_v02.xml

Depending on the setup in ddsim_steer.py, either a .slcio file or a .edm4hep.root file can be written

Events can be visualised in the standard way for .slcio file or a .edm4hep.root files, e.g for .slcio with ILD:

ced2go -d $k4geo_DIR/ILD/compact/ILD_l5_o1_v02/ILD_l5_o1_v02.xml dummyOutput.slcio

Structure of the Library

The library has the following key components:

• Trigger: Kill the full simulation and run the fast simulation, if conditions are fulfilled
• Model: Prepare the model input, and interpret the output
• Geometry: Place spacepoints produced by the model (localToGlobal). Also provide local direction of track at calorimeter face
• Inference: Calls desired inference library for the model. The library currently supports LibTorch and OnnxRuntime
• Hit Maker: Geant4 helper class for placement of energy deposits that land in a sensitive detector region

The structure of the library is summarised in the following class diagram:

And the order of operations is as follows:

More technical details

For running the example with a given DD4hep detector, there has to be a region defined in which the fast ML shower simulation will be run.

Depending on the detector of interest, this would involve making a local copy of lcgeo/k4geo (available under $k4geo_DIR), and modifying the xml description of the detector element:

diff --git a/ILD/compact/ILD_common_v02/SEcal06_hybrid_Barrel.xml b/ILD/compact/ILD_common_v02/SEcal06_hybrid_Barrel.xml
index 5ff2e50..084f7ea 100644
--- a/ILD/compact/ILD_common_v02/SEcal06_hybrid_Barrel.xml
+++ b/ILD/compact/ILD_common_v02/SEcal06_hybrid_Barrel.xml
@@ -1,7 +1,14 @@
 <lccdd>
+
+  <regions>
+    <region name="EcalBarrelRegion">
+    </region>
+  </regions>
+
   <detectors>

-    <detector name="EcalBarrel" type="SEcal06_Barrel" id="ILDDetID_ECAL" readout="EcalBarrelCollection" vis="BlueVis" >
+    <detector name="EcalBarrel" type="SEcal06_Barrel" id="ILDDetID_ECAL" readout="EcalBarrelCollection" vis="BlueVis"
+     region="EcalBarrelRegion">

       <comment>EM Calorimeter Barrel</comment>

This model has to be activated in the ddsim_steer.py file. The relevant functions are:

• For ONNX inference: def aiDance(kernel)
• For Torch inference: def aiDanceTorch(kernel):
• For Loading from a HDF5 File (Experimental interface!): def LoadHdf5(kernel):

And each can be activated respectively by setting:

• SIM.physics.setupUserPhysics( aiDance)
• SIM.physics.setupUserPhysics(aiDanceTorch)
• SIM.physics.setupUserPhysics(LoadHdf5)

Coding style

Coding style is partially enforced. Formatting is done via clang-format using the configuration found in .clang-format. Additionally there are minimal clang-tidy checks that enforce that class names are CamelCase and that private and protected member variables have an m_ prefix and are in camelBack case.