Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Current capabilities

This page describes the development branch as of 4 September 2026. It summarizes implemented paths, not the performance of any particular checkpoint.

Detector and input support

DetectorCollision settingInputsData version in the default recipeStatus
CMS Run 3Proton-proton, with and without pileupTracks and calorimeter clusters3.2.0Supported
CLDElectron-positronTracks and calorimeter clusters3.2.1Supported
CLICdetElectron-positronTracks and calorimeter clusters3.2.1Supported
CLDElectron-positronTracker and calorimeter hits3.2.1Research workflow
CLICdetElectron-positronTracker and calorimeter hits3.2.1Research workflow
MAIAMuon-collider detector studyPostprocessing onlyPartial
IDEAElectron-positron detector conceptWork in progressPlanned

The version numbers above come from the default recipes in particleflow_spec.yaml. Dataset, code, and checkpoint versions are separate and must be checked together.

End-to-end operations

OperationCMSCLDCLIC
Detector simulation configurationSupportedSupportedSupported
Detector-specific postprocessingSupportedSupportedSupported
TFDS creationSupportedSupportedSupported
Short local software testSupportedSupportedSupported
Standard PyTorch trainingSupportedSupportedSupported
Multi-GPU trainingSupportedSupportedSupported
Checkpoint loading and fine-tuningSupportedSupportedSupported
Standalone ROOT evaluationCMS-specific workflowSupportedSupported
Physics validationCMSSW/site dependentSupportedSupported
ONNX export and comparisonSupportedSupportedSupported

“Supported” means that code and configuration exist in this repository. Large-scale simulation and physics validation still require the relevant CMS or Key4HEP software environment, computing resources, and detector knowledge.

Model support

The standard detector recipes currently use an attention model with separate input encodings for different detector-element types. The shared MLPF interface also contains several advanced backbones:

Model familyIntended useStatus
Multi-head attentionDefault CMS, CLD, and CLIC recipesSupported
GNNLSHScalable graph processing with locality-sensitive hashingResearch workflow
HEPT and HEPTv2Hash-based efficient particle transformersResearch workflow
LitePTSparse point-transformer studiesResearch workflow; Nvidia-oriented dependencies
Shared backbone and task-query readoutCross-task model studiesResearch workflow

The repository supports single-node distributed PyTorch training. Ray Train and Ray Tune paths are available for distributed training and hyperparameter searches, but they require cluster-specific setup and are treated as research workflows.

Validation support

The repository can check four different questions:

  1. Dataset integrity: Are the schemas, detector relationships, target assignments, and energy sums sensible?

  2. Model behavior: Does the loss converge, and are particle-level predictions well formed?

  3. Physics behavior: How do particles, jets, and missing momentum compare with reference targets and rule-based reconstruction?

  4. Deployment behavior: Does an ONNX export agree numerically with PyTorch, and what runtime and memory does it use?

A local smoke test covers only a small part of the first two questions. Published physics claims require the larger validation setup described in the corresponding paper.

Known boundaries