1 of 16

A3D3 HEP activities

Mia

2 of 16

HEP Activities in A3D3

HEP PIs: Shih-Chieh, Phil, Mark, Javier, Mia

Projects (A complete list can be found in later slides):

  • Algorithms for reconstruction, simulation, applications spanning online (<< ms)/offline (ns - s, high throughput).
  • Active collaborations with HAC/Targeted systems/heterogeneous
    • E.g. interpretable GNNs for point cloud data.
    • E.g. developing hls4ml and applying to use cases: recursive NN/GNN in hls4ml.
    • Planned activities : Fast ML workshop in October.

3 of 16

2022 Q1 Publications, Conference submissions planned

  • MLPF with CMS data, ACAT 2021, https://arxiv.org/abs/2203.00330
  • “Semi-supervised Graph Neural Networks for Pileup Noise Removal”, preliminary results shown at BOOST/NeurIPS AI4Scicence. To be submitted to PRD next week.
  • “Information bottleneck-Guided Stochastic Attention Mechanism for Interpretable Graph Learning”: submitted to ICML 2022. https://arxiv.org/abs/2201.12987
  • Conferences:
    • Inter-experiment Machine Learning workshop (CERN): “Semi-supervised Graph Neural Networks for Pileup Noise Removal”
    • APS April meeting: “Calibration of electrons and photons in the CMS ECAL with graph neural networks”

4 of 16

Semi-supervised Graph PUPPI

People: Yongbin Feng (Fermilab), Tianchun Liu (Purdue), Shikun Li(Purdue), Lisa Paspalaki (Purdue), Pan Li (Purdue), Mia Liu(Purdue), Nhan Tran (Fermilab)

Goal: advanced pileup mitigation with graph neural networks

Status:

Short paper presented in NeurIPS 2021 Workshop AI4Science: Including particle label classification performance

Long paper (PRD) in preparation, including improvements in physics observables.

Training with CMS full simulation.

5 of 16

Particle flow reconstruction using a GNN

Goal: perform particle-flow (PF) reconstruction using a graph neural network (GNN) and apply explainable AI techniques to interpret the GNN model

Motivation: PF may be casted as a ML task to take advantage of highly parallelizable chips to perform accurate and fast reconstruction. GNNs are able to take advantage of the irregular geometry of the LHC detector to achieve better reconstruction. Furthermore, interpreting the GNN may help in increasing our confidence of the model and ensure robustness under changing conditions.

People: Javier Duarte, Raghav Kansal, Joosep Pata

Meeting: Thursdays 9am PST (biweekly)

Document/Code: https://github.com/faroukmokhtar/particleflow

Paper: https://arxiv.org/abs/2111.12840

6 of 16

Calorimetry Clustering with SPVCNN

Goal: Develop low-latency, scalable and accelerable ML algorithms for calorimetry clustering in HLT.

Motivation: Pushing the energy, intensity frontiers at colliders w/ conventional algorithms is limited by computational budget. High pileup is a rich environment for ML, opportunity for improvement over conventional approach.

People: Alex Schuy (UW), Shih-Chieh Hsu (UW), Scott Hauck (UW), Zhijian Liu (MIT), Phil Harris (MIT), Song Han (MIT), Jeff Krupa (MIT)

Meeting: 3:30 pm (PST) Wed https://indico.cern.ch/event/1111394/

Document/Code: mit-han-lab/spvnas-dev at dev/calo (github.com)

7 of 16

Using Autoencoder Latent Spaces to Derive L1 Trigger Physics Quantities

Goal: Derive more effective L1 Trigger information for the upcoming HGCAL installation at the HL-LHC

Motivation: Data transmitted off the Compact Muon Solenoid (CMS) experiment at LHC will arrive at such a fast rate that there is little time to derive meaningful physics quantities used to run the trigger systems. This can be done from the latent spaces which are produced in the hardwares that run HGCAL wafers.

People: Phil Harris (MIT), Duc Hoang (MIT), Christin Herwig (Fermilab), Cristina Mantilla Suarez (Fermilab), Nhan Tran (Fermilab)

Meeting: Biweekly Thursday 3PM EST

Document/Code: https://github.com/eric-moreno/ECON-Regression

8 of 16

Machine learning for HEP simulations

Goal: Use machine learning (ML) and graph neural networks (GNNs) for simulating LHC collisions

Motivation: Detector simulation takes up ~40% of LHC computing resources, and simulation complexity is only going to increase over the next decade with the upcoming HL-LHC upgrades. ML has the potential to both provide a significant (up to five-orders-of-magnitude) speed-up and increase simulation performance.

People: Javier Duarte (UCSD), Raghav Kansal (UCSD), Maurizio Pierini (CERN)

Meeting: Thursdays 8am PST (Weekly)

Document/Code: https://github.com/rkansal47/MPGAN/

Paper: Kansal et. al., Particle Cloud Generation with Message Passing Generative Adversarial Networks, NeurIPS 2021, arXiv:2106.11535

9 of 16

Energy corrections with graph neural networks

Goal: Develop new approaches to energy corrections in the CMS calorimeters to provide improved energy resolutions and reduced sensitivity to pileup, noise, nonfunctioning detector elements, etc.

Motivation: New approaches will be needed to handle increase HL-LHC pileup, and to provide improved sensitivity in many flagship analyses, eg Higgs->diphoton

People (in no particular order): Simon Rothman, Phil Harris (MIT), Rajdeep Chatterjee, Bhargav Joshi, Roger Rusack (UMN), Alpana, Nitish Kumar, Seema Sharma, Shubham Pandey (IISER-Pune), Lindsey Gray (FNAL)

Meeting: TBA

Document/Code: TBA

10 of 16

Tau3Mu reconstruction using GNNs

Goal: Develop neural network models for real-time global event reconstruction in CMS L1 trigger.

Motivation: Hardware upgrades in the HL-LHC will allow for more sophisticated algorithms in the online filtering systems. Neural networks can be used to achieve high efficiency in identifying exotic events that indicate new physics.

People: Miaoyuan Liu (PU), Pan Li (PU), Jacobo Konigsberg (UF), Daniel Guerrero (UF), Lisa Paspalaki (PU), Siqi Miao (PU), Hyeon-Seo Yun (PU), Benjamin Simon (PU), Eric Reinhardt (PU)

Meeting: Wednesday, 3:30pm (EST)

Document/Code: TBA

11 of 16

CMS L1BTag

Goal: Develop a NN based tagger for identifying b jets in the CMS L1 Trigger

Motivation: Numerous applications for b quark identification; di-Higgs Golden Channel (HH4b), b Meson Decay, etc.

People: Aidan Chambers, Dylan Rankin, Phil Harris, Duc Hoang

Meeting: Tuesday 12PM EST

Document/Code: Some basic code here: https://github.com/aidan-dc/L1BTag

12 of 16

SONIC in CMS

People: Phil, Patrick, Javier, Mia, Yongbin, Nhan, Kevin + many others

Goal: CMS data processing workflow based on SONIC running in CMS central production (tier-2 sites)

Status: Two CMS data processing workflows developed. Tests on-going with servers manually set up. Infrastructure development for production systems planned for 2022.

13 of 16

Interpretable Models for Point Cloud Data

Goal: Study models that are inherently interpretable when trained with point cloud data

Motivation: ML models trained on point cloud data (especially in HEP) lack interpretability due to the complex nature of physical processes. Yet point cloud data interpretation has not been investigated in depth. With interpretability, ML models are more trustworthy in science, and they can also provide us with more insights regarding the datasets used.

People: Siqi Miao, Miaoyuan Liu, Pan Li

Meeting: Friday 4pm EST (biweekly)

Document/Code: https://github.com/Graph-COM/GSAT

Paper: https://arxiv.org/abs/2201.12987

14 of 16

ONNX Support of LSTM/GRU Networks for HLS4ML Implementations

Goal: Establish a LSTM/GRU->ONNX->HLS4ML workflow.

Motivation: LSTM is a type of recurrent neural network for deep learning. The work we are doing for QONNX is collaboration between the hls4ml team and Xilinx/AMD. LSTM can be applied to our HLS4ML workflow.

People: Xiaohan Liu(UW), Jeffery Xu(UW), Aidan Yokuda(UW), Andrew(Yihui) Chen(UW), Scott Hauck(UW), Shih-Chieh Hsu(UW)

Meeting: Tuesday 10:30am PST (Weekly)

Document/Code:

LSTM: https://people.ece.uw.edu/hauck/publications/Thesis-RichaRao.pdf

QONNX: https://xilinx.github.io/finn/2021/11/03/qonnx-and-finn.html

15 of 16

GNNs in hls4ml

Goal: GNNs on FPGAs with hls4ml

People (currently active): Javier Duarte (UCSD), Suki Krishna (UCSD), Simon Poon (UCSD), Tony Aportela (UCSD), Abdel Elabd (Penn), Shih-Chieh Hsu (UW), Mia Liu (Purdue), Jan Schulte (Purdue).

Short Summary:

  • PyG to HLS:
    • Example with tracking arXiv:2112.02048
    • hls4ml PR in progress: #379
  • Fully-connected adjacency interaction network based model
    • Use for L1 anomaly detection
    • Use for L1 LLP jet tagging
    • Use for L1 MET

16 of 16

Recursive NNs (LSTM & GRU) in hls4ml

Goal: Fast LSTM/GRU on FPGAs with hls4ml

People (currently active): Elham E Khoda (UW), Rafael Teixeira de Lima (SLAC), Dylan Rankin (MIT), Phil Harris (MIT), Michael Kegan (SLAC), Shih-Chieh Hsu (UW), Scott Hauck (UW)

Short Summary:

  • Working with Keras/TensorFlow models and targeting Xilinx board
  • Demonstrate the performance on three models of different size and complexity
    • Small model (~5k parameters): binary classifier (top-tagging)
    • Medium model (~20-50k parameters): 3-class classifier (b-tagging)
    • Large model (~150k parameters): 5-class classifier (QuickDraw dataset)