🔬 physics

Online Data Reduction with Spiking Neural Networks: A Temporal-Coincidence Encoder and Distributed SNN for the ePIC dRICH Detector

This paper presents a two-stage online data reduction pipeline utilizing spiking neural networks—a temporal coincidence encoder and a distributed classifier—to effectively filter high-rate dark count noise from the ePIC dRICH detector's SiPM channels while maintaining high signal detection efficiency.

Original authors: Pierpaolo Perticaroli, Roberto Ammendola, Andrea Biagioni, Ottorino Frezza, Francesca Lo Cicero, Michele Martinelli, Pier Stanislao Paolucci, Elena Pastorelli, Luca Pontisso, Cristian Rossi, Francesco

Published 2026-07-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pierpaolo Perticaroli, Roberto Ammendola, Andrea Biagioni, Ottorino Frezza, Francesca Lo Cicero, Michele Martinelli, Pier Stanislao Paolucci, Elena Pastorelli, Luca Pontisso, Cristian Rossi, Francesco Simula, Piero Vicini, Alessandro Lonardo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a single, quiet conversation in a room where 320,000 people are all shouting at once. That is the challenge facing the ePIC experiment, a massive particle physics detector designed to study how electrons and ions collide.

The detector uses a special "eye" called a dRICH, which is covered in 320,000 tiny sensors (SiPMs). These sensors are supposed to catch flashes of light (Cherenkov radiation) that happen when a particle passes through. However, these sensors are also "noisy." Over time, they start generating random, fake flashes of light (called Dark Count Rate or DCR) just because they are old and damaged by radiation.

The Problem: Too Much Noise
The real signal (the particle flash) is very fast—it happens in a tiny 2-nanosecond burst. The fake noise, however, is spread out evenly over time.
Currently, the sensors are so noisy that they are generating 300,000 fake flashes per second, per channel. If the computer tried to process every single flash, the data stream would be so huge it would clog the internet pipes connecting the detector to the main computers. The team needs to throw away at least 80% of this data before it leaves the detector, but they can't throw away the real signals.

The Solution: A "Brain-Inspired" Filter
The authors propose a solution using Spiking Neural Networks (SNNs). Think of a standard computer as a factory assembly line that checks every single item, whether it's a diamond or a piece of trash. An SNN is more like a smart security guard who only reacts when something interesting happens. It waits for a "spike" (a signal) before doing any work.

Their system has two stages, acting like a two-step security checkpoint:

Stage 1: The "Coincidence" Gatekeeper (The Encoder)

Imagine a bouncer at a club door.

The Setup: Each group of 256 sensors (a "Photo Detection Unit" or PDU) has its own bouncer.
The Trick: The bouncer is programmed to ignore single, lonely flashes (noise). But if two or more sensors in that group flash at almost the exact same time (within a tiny window), the bouncer assumes it's a real particle and rings a bell (sends a "spike").
The Result: This step is incredibly effective. Because the real signal flashes happen in a tight cluster, they get through. The random noise, being spread out, usually doesn't trigger the bouncer. This single step throws away more than 90% of the data immediately, leaving only the most promising "spikes" for the next stage.

Stage 2: The Distributed Jury (The Classifier)

Now, the remaining "spikes" are sent to a team of judges.

The Setup: The detector is divided into 30 sections. Each section has its own small "mini-brain" (a small neural network) that looks at the spikes from its area.
The Decision: These mini-brains talk to a "Head Judge" (an aggregation network). They quickly decide: "Is this a real particle event (Signal + Noise) or just random static (Noise Only)?"
The Speed: Because the system is "event-driven," it only does math when there is a spike to process. If there's just noise, the system stays mostly silent, saving massive amounts of energy and time.

The "Early Exit" Strategy
The system is also designed to be impatient in a good way. If the "Signal" neurons start firing very quickly (which real particles do), the system stops thinking immediately and says, "Yes, that's a particle!" It doesn't wait for the full calculation to finish. This cuts the decision time down to just a few tiny steps, making it incredibly fast.

The Hardware Test
The team built a prototype on a specialized computer chip (an FPGA) that mimics a brain's architecture.

The Result: They successfully processed data at a rate of 1.7 million events per second.
The Goal: The detector needs to handle 100 million events per second. While 1.7 million isn't there yet, the test proved the concept works. The team has a clear list of optimizations (like making the "security guard" faster and the "judges" more efficient) to bridge the gap to the full speed.

In Summary
This paper describes a clever way to filter out noise in a giant particle detector. Instead of trying to process everything, they use a "smart filter" that only pays attention to things that happen at the exact same time. This reduces the data flood by over 90% and uses a brain-like network to make split-second decisions, ensuring that only the interesting physics events get sent to the main computers.

Technical Summary: Online Data Reduction with Spiking Neural Networks for the ePIC dRICH Detector

Problem Statement
The dual-radiator Ring Imaging Cherenkov (dRICH) detector of the Electron-Ion Collider (EIC) ePIC experiment faces a severe bandwidth bottleneck. The detector reads out approximately 320,000 Silicon Photomultiplier (SiPM) channels at a bunch-crossing rate of 100 MHz. Over the experiment's lifetime, the SiPM Dark Count Rate (DCR) is projected to rise to ~300 kHz per channel due to radiation damage. This results in a raw aggregate output of ~6.8 Tbit/s, far exceeding the available 30 Gbit/s egress bandwidth to the first off-detector processing tier (Echelon-0).

The core challenge is a temporal discrimination problem: genuine Cherenkov hits from physics signals cluster within a narrow ~2 ns window of the 10 ns bunch-crossing interval, whereas DCR hits are uniformly distributed across the full 10 ns window. The system requires an online data reduction factor of at least 5 to discard crossings containing only uncorrelated DCR while preserving all crossings with potential physics content. While dense neural networks (MLPs) have been proposed for this task, the fine timing structure and high throughput requirements suggest Spiking Neural Networks (SNNs) as a potentially more efficient, event-driven alternative.

Methodology
The authors propose a two-stage online data reduction pipeline implemented on FPGAs, leveraging the intrinsic temporal dynamics of SNNs:

Stage 1: Temporal-Coincidence Encoder (Per-PDU)
- Architecture: A single Leaky-Integrate-and-Fire (LIF) neuron is assigned to each Photo Detection Unit (PDU), which aggregates 256 SiPM channels.
- Mechanism: The encoder integrates incoming hits over time. Due to the uniform distribution of DCR, the membrane potential leaks faster than it accumulates, keeping the neuron silent. Conversely, the clustered Cherenkov hits (within ~2 ns) cause a rapid voltage rise, triggering a spike.
- Constraint: The system enforces a "single-spike-per-bunch-crossing" policy, outputting at most one spike per PDU per crossing.
- Hardware Implementation: A specialized, heavily approximated LIF model is used (uniform weights, single-bit membrane, shift-based leak) to resolve front-end data-format mismatches (packing up to 4 hits per word) into a single clock cycle. This stage achieves >90% data sparsification before any learned classifier is applied.
Stage 2: Distributed SNN Classifier
- Topology: A two-tier network mirrors the physical detector layout.
  - Sub-sector Networks: 30 sub-networks (one per Data Aggregation and Manipulation, DAM, board) process the 42 encoder outputs from their respective sub-sectors. Each uses a 16-neuron hidden layer and a 4-neuron output layer.
  - Aggregation Network: A central network on a dedicated Trigger Processor (TP) board concatenates the 120 local features (30 sub-sectors × 4 features) and classifies the event as "Signal+Noise" or "Noise-Only."
- Inference Strategy: The system employs an "early-exit" strategy. If the "Signal" neuron fires a configurable number of times ( $E_T$ ) early in the inference sequence, the process halts immediately. This reduces average latency significantly, as Signal events drive early firing while Noise events remain largely silent.
- Hardware Platform: The classifier runs on the AIGOR multi-core neuromorphic architecture deployed on AMD Versal Premium FPGAs, utilizing Address-Event Representation (AER) for spike routing.

Key Contributions

End-to-End Methodology: The paper presents a complete pipeline from raw SiPM hits to trigger decisions, integrating a hardware-optimized temporal encoder with a distributed SNN classifier.
Hardware-Software Co-design: The authors demonstrate that the most hardware-efficient encoder configuration (a single-bit coincidence detector) coincides with the optimal point on the accuracy-latency Pareto frontier.
AER-Routed FPGA SNN Characterization: This work provides a detailed characterization of an SNN pipeline at the boundary of a High-Energy Physics (HEP) front-end, identifying synchronization traffic in timestep-driven execution as the primary throughput bottleneck.
Validation: The system is validated through software simulation on simulated ePIC events and hardware execution on a single-sub-sector testbed using an AMD Versal Premium VPK180 FPGA.

Results

Classification Performance: On simulated events across the full DCR range (25–300 kHz), the system achieves a True Positive Rate (TPR) > 94% and a True Negative Rate (TNR) ≥ 80%.
Sparsification: The LIF encoder alone reduces the data volume by >90%, lowering the post-encoder load to ~60–120 spikes per bunch crossing across the full detector.
Latency: With an early-exit threshold of $E_T=1$ , the average classification latency drops to ~2 algorithmic timesteps (from a maximum of 10), with a 99th-percentile latency of ~5 timesteps.
Hardware Throughput: A hardware proof-of-concept on a single sub-sector achieved a measured throughput of ~1.7 MHz. The authors note this is below the target 100 MHz bunch-crossing rate.
Quantization Robustness: The system maintains performance with aggressive fixed-point quantization (down to 4-bit representations), validating the use of low-precision arithmetic for resource efficiency.

Significance and Claims
The paper positions this work as a concrete demonstration of SNNs in a complete HEP Data Acquisition (DAQ) pipeline, a domain where such integration has been limited to early exploratory work. The authors claim that the combination of a temporal-coincidence encoder and a distributed SNN offers a template for other timing-driven detector applications operating at high rates.

The significance lies in addressing the specific constraints of the dRICH detector: the encoder exploits the native temporal strength of SNNs to filter noise before classification, while the distributed architecture respects the physical readout topology. The authors modestly acknowledge that the current hardware throughput (1.7 MHz) does not yet meet the 100 MHz requirement. They identify the primary bottleneck as the synchronization traffic inherent to timestep-driven execution on AER fabrics. The paper concludes that evolving the execution model toward event-driven dynamics, alongside multi-spike AER packets and higher clock frequencies, constitutes the necessary optimization path to reach the target rate.

Stage 1: The "Coincidence" Gatekeeper (The Encoder)

Stage 2: The Distributed Jury (The Classifier)

Technical Summary: Online Data Reduction with Spiking Neural Networks for the ePIC dRICH Detector

More like this