1 of 16

Project BEAD: Anomaly Detection for Particle Physics and Synergies with PHM

​

Pratik Jawahar

1

Overview:

  • The Problem
  • Evolution of Solutions
  • Connections to PHM

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

2 of 16

Proton Collisions at the LHC

  • We have collisions at 40MHz, with each producing ~O(10k) particles
  • The different particles are measured by different sub-components of the larger ATLAS detector system

2

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

3 of 16

The Problems:

  • We don’t want to waste disk space on noisy events caused by detector defects
    • These events can be discarded; but flagging the issues and diagnosing them to inform detector maintenance is key
      • Similar to fault detection, diagnosis and mitigation in PHM
  • Most commonly occurring events have already been studied to extreme precision
    • Exotic Physics phenomena/unknown signals (eg. dark matter) manifest as missing energy
      • Effectively each collision is a partially observed system interacting with detector systems
  • We physically can’t store all events

3

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

4 of 16

The Problems:

  • We don’t want to waste disk space on noisy events caused by detector defects
    • These events can be discarded; but flagging the issues and diagnosing them to inform detector maintenance is key
      • Similar to fault detection, diagnosis and mitigation in PHM
  • Most commonly occurring events have already been studied to extreme precision
    • Exotic Physics phenomena/unknown signals (eg. dark matter) manifest as missing energy
      • Effectively each collision is a partially observed system interacting with detector systems
  • We physically can’t store all events

4

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

5 of 16

The Problems:

  • We don’t want to waste disk space on noisy events caused by detector defects
    • These events can be discarded; but flagging the issues and diagnosing them to inform detector maintenance is key
      • Similar to fault detection, diagnosis and mitigation in PHM
  • Most commonly occurring events have already been studied to extreme precision
    • Exotic Physics phenomena/unknown signals (eg. dark matter) manifest as missing energy
      • Effectively each collision is a partially observed system interacting with detector systems
  • We physically can’t store all events

5

Source

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

6 of 16

Quantifying AD Performance

  • Mathematically impossible to quantify AD performance on unknown signals
  • Our solution: Build a broad benchmark to train and evaluate perf.
    • Roughly extrapolate performance Expectation to unknown signals
      • Darkmachines Anomaly Dataset
    • We also presented novel evaluation metrics {Total Significance Improvement}: both broad and fine-grained statistical performance evaluation with uncertainty quantification

6

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

7 of 16

Unsupervised Anomaly Detection

  • Variational Autoencoder baseline:
    • Train on “nominal” events
    • During inference, detector anomalies result in high reconstruction loss
  • Problem: The standard Gaussian posterior could not model non-deterministic rollouts (quantum procs.)
      • Connects to autonomous, dynamic systems with external agents / unobserved agents
  • My proposal: Learn a complex non-gaussian multimodal posterior (MS Thesis)
    • VAE+Normalizing Flows
  • Problem: Anomaly Detection strictly improves, but we lose scope for diagnosis of detected anomalies (intractable learned bias)
      • We need to induce bias, not learn it
  • My proposal: Replace flat architecture topology with Graphs mapping to the particle-space (MS Thesis)

7

Source

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

8 of 16

Unsupervised Anomaly Detection

  • Variational Autoencoder baseline:
    • Train on “nominal” events
    • During inference, detector anomalies result in high reconstruction loss
  • Problem: The standard Gaussian posterior could not model non-deterministic rollouts (quantum procs.)
      • Connects to autonomous, dynamic systems with external agents / unobserved agents
  • My proposal: Learn a complex non-gaussian multimodal posterior (MS Thesis)
    • VAE+Normalizing Flows
  • Problem: Anomaly Detection strictly improves, but we lose scope for diagnosis of detected anomalies (intractable learned bias)
      • We need to induce bias, not learn it
  • My proposal: Replace flat architecture topology with Graphs mapping to the particle-space (MS Thesis)

8

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

9 of 16

Inducing Physics Biases

  • Representing input point clouds as Heterogeneous Hierarchical Graphs
    • Physics objects per event:
      • Jets (shower of multiple particles originating from a common vertex)
      • Leptons (charged particles that bend in detectors magnetic field) ; photons (EM calorimeters)
  • The graph is constructed to provide categorical and geometric particle information structurally by design instead of learning these
  • Lorentz invariance is built into the message passing step for the macro-graph via the minkowski inner product
  • Each micro-graph uses separate, unconstrained message passing
  • A GCN-modification is used to consolidated aggregated messages conditioned by the hierarchical level

9

Source

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

10 of 16

Inducing Physics Biases

  • Micro-graphs should nominally model IRC safe variables
    • We induce this using Energy Flow Polynomials (node features) that provide a strictly IRC safe basis
  • Training data is typically simulated using the Pythia generator (MCMC modelling)
    • This however has resulted in a large sim-to-real gap
    • We introduce training data from 3 different generators (observational bias)
      • Contrastive learning helps disentangle implicit biases of the chosen MCMC generator
  • Controllable multi-modality:
    • Dirichlet posterior:
      • Tractable approximation via softmax(z)

10

Source

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

11 of 16

Inducing Physics Biases

  • Micro-graphs should nominally model IRC safe variables
    • We induce this using Energy Flow Polynomials (node features) that provide a strictly IRC safe basis
  • Training data is typically simulated using the Pythia generator (MCMC modelling)
    • This however has resulted in a large sim-to-real gap
    • We introduce training data from 3 different generators (observational bias)
      • Contrastive learning helps disentangle implicit biases of the chosen MCMC generator
  • Controllable multi-modality:
    • Dirichlet posterior:
      • Tractable approximation via softmax(z)

11

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

12 of 16

BEAD Software Package

  • FAIR Software design
  • Packaged with Rust-based tooling stack for high-perf, easy use
    • Multi-GPU distributed training
    • nJIT-based large scale data proc.
  • Modular design with wide range of benchmarks, running modes and ablation experiments
  • Supervised GSoC, NSF student projects and 2 postdocs
  • Extending BEAD as a Foundation Model pre-training framework
    • Collaborating with a combined particle, nuclear and astrophysics teams.

12

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

13 of 16

HEP-PHM Synergies & Common Problems

  • Shortage of anomalous data (vs nominal operation) for AD algorithms to use learned bias
    • Known Physics must be induced with structural priors
  • Sim-to-real gap in data production modes
    • Observational bias mitigation strategies are key
  • Partially observed systems
    • Leverage inherent in/equi-variances of the system
  • Systems being modeled are often a combination of multiple sub-systems that are equally complex
    • Current SOTA generally targets fine grained anomaly detection for each sub-system
      • Results in fragmented AI models
  • How can we leverage shared context, and broader semantic structures to improve anomaly detection performance for partially observed multi-agent/component systems?

13

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

14 of 16

Thank You!

14

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

15 of 16

PhD Thesis:

15

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar

16 of 16

Dynami-CAL-GN-CHLU Hybrid for Partially Observed Systems

16

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

BEAD: Pratik Jawahar

SMARTHEP is funded by the European Union’s Horizon 2020 research and innovation programme, call H2020-MSCA-ITN-2020, under Grant Agreement n. 956086

ESR12: Pratik Jawahar