From HEP.TrkX to Exa.TrkX
Paolo
Timelines and Budgets
HEP.TrkX
Oct 2106 to March 2018 (12+6 months)� Six months no-cost extension
Budget:� Caltech $200K� FNAL $275K� LBL $275K
12 Authors�2 Papers�11 Conference talks
Exa.TrkX
Jun 2019 to Nov 2021 (30 months)
Budget:� Caltech $525K� FNAL $525K� LBL $525K� SLAC $525K
Goals and Objectives
Provide a framework to develop and evaluate new algorithms for track finding and classification. The framework will be demonstrated by applying advanced pattern recognition techniques to track candidate formation
Develop Tracking algorithms that can run efficiently and robustly in production on the massively parallel architectures (GPUs, TPUs, etc) which will be used by Exascale-class HPCs, and by the online Trigger and Data Acquisition (TDAQ) systems of HEP experiments.
HEP.TrkX Milestones
Group | Deliverable | Due | Institutions | Status | Notes |
Simulation | 1: Realistic full simulation samples of generic HL-LHC tracking detector. | Q4 16 | LBL, FNAL | ✅ | aCTS, TrackML |
Algorithm | 2: Reference solution for seeding and track candidate formation using state of the art algorithms. | Q2 17 | FNAL, Caltech | ❌ | Not defined, nor run. |
| 3: Pattern recognition algorithm to build track candidates from seeds and space-points. | Q3 17 | Caltech, LBL | ✅ | Seeding much better than track following |
| 4: Adapt track finding algorithm and evaluate in LArTPC environment | Q4 17 | FNAL, Caltech | ❌ | Never started |
Milestone: Demonstrate full simulation and reconstruction chain for generic HL-LHC tracking detector and produce plan for full development and implementation | Q4 17 | ALL | ❌ | Doable, not done |
Exa.TrkX Milestones Year 1 (Jun 19-Jun 20)
| Description | Due | Institutes | Who |
| Paper submission for IPDPS | Oct 19 | LBL | Steve |
| Classical tracking algorithms running on trackML data, possibly in aCTS | Oct 19 | Caltech | JR+Joosep |
D2a | First implementation 3D point making GNN (Proto-DUNE) | Dec 19 | Fermi | Giuseppe |
D2b | First implementation 3D clustering pixel detector DUNE ND | Dec 19 | SLAC | Kazu |
D1 | End-to-end reference solution based on Geometric Deep Learning, initially targeting TrackML dataset. | Dec 19 | Caltech, LBNL | ? |
D3 | Sensitivity studies of LHC and LArTPC tracking models, identifying most important nuisance parameters. | Mar 20 | Caltech, FNAL | FNAL postdoc? |
D4 | Strategy to increase tracking model robustness against nuisance parameters. | Mar 20 | LBNL, SLAC | LHC Exalearn postdoc |
D5 | Identify best approaches (Bayesian, Evolutionary, autoML) to optimize tracking model architecture and hyperparameters. | Jun 20 | Caltech, FNAL | Y2 Q2 |
M1 | Demonstrate distributed training of HEP tracking models. | Jun 20 | All | Y2 Q2 |
Exa.TrkX Milestones Year 2
| Description | Due | Institutes | |
D6 | Optimize tracking model resource usage with HEP Trigger applications in mind. | Sep 20 | FNAL, SLAC | |
D7 | Optimize distributed data delivery, parallelize input data preparation. | Dec 20 | Caltech, LBNL | |
D8 | Demonstration of HPC-scale distributed optimization for HEP tracking models. | Mar 21 | Caltech, LBNL | |
D9 | Optimize training and parameters of LArTPC network; evaluate resulting performance on different experiments. | Mar 21 | FNAL, SLAC | |
M2 | Run HEP Tracking Grand Challenge in collaboration with NERSC Big Data Center. | Jun 21 | All | |
Exa.TrkX Milestones Year 3
| Description | Due | Institutes |
D10 | Document physics performance of fully optimized tracking models in peer reviewed papers, present results to at least two HEP collaborations. | Dec 21 | All |
D11 | Provide interoperability layers with main HEP frameworks (e.g. cmssw, athena/Gaudi, LArSoft). Package Exa.TrkX software as required by experiments. | Dec 21 | All |
D12 | In collaboration with at least one HEP experiment, compare CPU and physics performance of Exa.TrkX tracking models to experiment tracking algorithms. | Mar 22 | All |
M3 | In collaboration with at least one HEP experiment, validate our tracking models for production. | Jun 22 | All |