snn-online-data-reduction-physics

name	snn-online-data-reduction-physics
description	Spiking Neural Networks for online data reduction in high-energy physics detectors. Temporal-coincidence encoding and distributed SNN architecture for the ePIC dRICH detector at the Electron-Ion Collider. Achieves 5x data reduction while preserving genuine Cherenkov photon signals against SiPM dark counts.
category	ai_collection
trigger_words	["online data reduction SNN","temporal-coincidence encoder","dRICH detector SNN","ePIC experiment data reduction","SiPM dark count filtering","SNN particle physics","Cherenkov photon detection SNN","distributed SNN for detector readout","neuromorphic particle physics","100 MHz detector readout","silicon photomultiplier SNN","online trigger SNN","event-driven detector readout","neuromorphic high energy physics"]

Spiking Neural Networks for Online Data Reduction in Particle Physics Detectors

Source

Perticaroli, P., Ammendola, R., Biagioni, A., Frezza, O., Lo Cicero, F., Martinelli, M., Paolucci, P.S., Pastorelli, E., Pontisso, L., Rossi, C., Simula, F., Vicini, P., & Lonardo, A. (2026). Online Data Reduction with Spiking Neural Networks: A Temporal-Coincidence Encoder and Distributed SNN for the ePIC dRICH Detector. arXiv:2607.03492

Categories: physics.ins-det arXiv: https://arxiv.org/abs/2607.03492

Problem Statement

The dual-radiator Ring Imaging Cherenkov (dRICH) detector at the ePIC experiment (Electron-Ion Collider) faces a critical data reduction challenge:

• 320,000 SiPM channels read out at 100 MHz bunch-crossing rate
• Dark count rate (DCR) rises to 300 kHz per channel over experiment lifetime
• DCR saturates output bandwidth → requires online data reduction factor ≥ 5×
• Most crossings contain only uncorrelated DCR hits; genuine Cherenkov photons produce temporally coincident signals

Solution Architecture

Two-Stage SNN Pipeline

SiPM Channel Signals
    │
    ▼
┌─────────────────────────┐
│  Temporal-Coincidence    │  Stage 1: Feature Encoding
│  Encoder                 │  Converts raw SiPM hits to spike trains
│                          │  based on temporal coincidence detection
└─────────────────────────┘
    │
    ▼
┌─────────────────────────┐
│  Distributed SNN         │  Stage 2: Classification
│  Classifier              │  Identifies genuine Cherenkov photon
│                          │  patterns vs. uncorrelated DCR noise
└─────────────────────────┘
    │
    ▼
Reduced Data Stream
(≥ 5× reduction, preserving genuine hits)

Stage 1: Temporal-Coincidence Encoder

Key Insight: Genuine Cherenkov photons arrive within a narrow time window; dark counts are temporally uncorrelated.

Mechanism:

• Monitors SiPM channel outputs for temporal coincidence (multiple hits within narrow window)
• Converts coincident hit patterns into spike trains for SNN input
• Acts as both feature extractor and noise pre-filter

Advantages over traditional methods:

• Event-driven: only processes when hits occur
• Temporal precision: captures sub-ns timing information
• Low computational overhead compared to full waveform processing

Stage 2: Distributed SNN Classifier

Architecture:

• Spiking neural network trained to classify temporal-coincidence patterns
• Distributed across processing nodes for scalability
• Event-driven inference: no computation on empty crossings

Key Design Decisions:

• Neuron model selection optimized for latency vs. accuracy
• Spike encoding preserves temporal information from encoder
• Distributed architecture matches detector channel topology

Performance Characteristics

Data Reduction Target

• Required: ≥ 5× reduction in output data rate
• Genuine signal preservation: Cherenkov photon patterns must be retained
• DCR rejection: Uncorrelated dark counts must be filtered

Latency Requirements

• Must operate at 100 MHz bunch-crossing rate
• Decision latency << bunch crossing period (10 ns)
• Event-driven processing eliminates idle-time computation

Scalability

• Distributed across 320,000 SiPM channels
• Each processing node handles subset of channels
• Inter-node communication minimal (only coincidence events)

Implementation Patterns

Temporal-Coincidence Detection

For each SiPM channel:
    Monitor hit timestamps
    If N hits within time window Δt:
        Generate spike event
        Forward to SNN classifier
    Else:
        Suppress (DCR noise)

SNN Classification Pipeline

Temporal-coincidence spikes → SNN input layer
    → Hidden spiking layers (LIF neurons)
    → Output layer: Cherenkov pattern probability
    → Decision: pass-through or suppress

Application Domains

Primary

• High-energy physics detectors: dRICH, other Cherenkov detectors
• SiPM-based readout systems: LHC upgrades, future colliders
• High-rate particle detectors: Any detector with bandwidth saturation

Secondary

• Astronomical detectors: Photon-counting instruments with dark noise
• Medical imaging: PET detectors with high dark count rates
• LIDAR systems: Time-of-flight sensors with background noise

Integration with Hardware

FPGA Implementation

• SNN inference on FPGA for low-latency operation
• Temporal-coincidence encoder in programmable logic
• Distributed processing matches channel topology

Neuromorphic Hardware

• Event-driven SNN naturally suited for neuromorphic chips
• Loihi, SpiNNaker, or custom neuromorphic ASICs
• Energy efficiency advantage over GPU/CPU solutions

Key Innovations

1. First application of SNNs to real-time particle physics data reduction
2. Temporal-coincidence encoding as a physics-informed feature extractor
3. Distributed SNN architecture matching detector channel topology
4. Event-driven processing eliminates idle-time computation at 100 MHz
5. Scalable design for 320,000+ SiPM channels

Challenges

1. Training data: Generating realistic SiPM + Cherenkov datasets for SNN training
2. Latency constraints: Sub-10ns decision time required
3. Radiation hardness: Electronics must survive detector environment
4. Calibration drift: SiPM characteristics change over experiment lifetime

Related Skills

• aigor-modular-neuromorphic-architecture (same research group, related hardware architecture)
• event-driven-neuromorphic-transceiver
• spiking-neural-network-analysis
• snn-fpga-hardware-software-codesign

Activation

snn online data reduction, temporal-coincidence encoder, dRICH detector, ePIC experiment, SiPM dark count filtering, Cherenkov photon detection, neuromorphic particle physics, distributed SNN detector, high-energy physics trigger, event-driven detector readout, 100 MHz readout, silicon photomultiplier SNN