Real-Time Wiener Deconvolution for feature reconstruction in JUNO

This paper presents a real-time Wiener deconvolution algorithm implemented on JUNO's FPGA-based readout boards to enable online signal reconstruction from photomultiplier tubes, thereby improving the detection of low-energy neutrino events while managing high data throughput and minimizing storage requirements.

The Gist

Imagine you are trying to listen to a single person whispering in a crowded, noisy stadium. Now, imagine that instead of one person, you have 43,000 people (the photomultiplier tubes, or PMTs) all trying to whisper at once, and the "crowd noise" is actually background radiation and electronic static.

This is the daily challenge for the JUNO experiment, a massive underground neutrino detector in China. Neutrinos are ghostly particles that rarely interact with matter, so when they do, they create tiny flashes of light. The detectors need to catch these flashes to study the universe.

However, there's a problem: The data is too heavy.

If the detectors recorded every single whisper and every bit of background noise, they would generate so much data that their hard drives would fill up in minutes, and their internet connection would burst. To solve this, they usually use a "gatekeeper" algorithm (called COTI) that just counts how loud the noise is above a certain level. But this gatekeeper is a bit clumsy. If two whispers happen very close together, the gatekeeper thinks it's just one loud shout, losing crucial details about when exactly the whispers happened.

The Solution: The "Super-Listener" (Real-Time Wiener Deconvolution)

This paper introduces a new, smarter gatekeeper built directly into the detector's brain (an FPGA chip). It's called Real-Time Wiener Deconvolution (RTWD).

Here is how it works, using some everyday analogies:

1. The Problem: The "Echo Chamber"

When a neutrino hits the detector, it creates a flash of light. But the detector itself isn't perfect; it blurs the flash, like looking at a bright light through a foggy window. If two flashes happen close together, their "fogs" overlap, making it look like one big, messy blob. The old system (COTI) just sees the blob and says, "Okay, that's one hit," and moves on.

2. The Tool: The "Noise-Canceling Headphones" (Wiener Filter)

First, the new system puts on a pair of high-tech noise-canceling headphones.

• How it works: It knows exactly what the "static" of the machine sounds like (the noise) and what a perfect "whisper" (a single particle hit) looks like.
• The Magic: It subtracts the static and sharpens the signal. It's like taking a blurry photo and using software to instantly make it crisp and clear, removing the graininess so you can see the edges of the object.

3. The Tool: The "Time-Reverse Engineer" (Deconvolution)

Next, the system acts like a time-traveling editor.

• The Analogy: Imagine you drop a pebble in a pond. The ripples spread out and get messy. If you could play the video of those ripples backwards, the water would flow back into the pebble, and the pebble would pop out of the water.
• The Magic: The "Deconvolution" filter does this mathematically. It reverses the blurring effect of the detector. It takes that messy, overlapping blob of a signal and "un-blurs" it, separating the two close whispers back into two distinct, sharp spikes.

Why This Matters: The "Traffic Cop" vs. The "Smart Director"

Think of the old system (COTI) as a Traffic Cop at a busy intersection.

• If two cars (particles) arrive at the exact same time, the cop gets confused and counts them as one big vehicle. He misses the details.
• If the cars are too close, he can't tell them apart.

The new system (RTWD) is like a Smart Director with a high-speed camera.

• Even if the cars arrive almost simultaneously, the director uses slow-motion replay (the filters) to see exactly where Car A ended and Car B began.
• It counts them correctly: "That's two cars, not one."

The Result: Seeing the Invisible

By installing this "Smart Director" directly onto the detector's chip (the FPGA), the JUNO experiment can:

1. Catch the faint whispers: It can detect very low-energy events that were previously lost in the background noise.
2. Save space: Instead of sending terabytes of raw, messy video to a server, it only sends a simple list: "At 10:00:01, we heard a whisper. At 10:00:02, we heard another." This saves massive amounts of storage.
3. Be faster: It does all this math in real-time, nanoseconds after the event happens, without slowing down the experiment.

The Bottom Line

This paper proves that we can build a "super-smart" filter inside the detector's brain. It turns a blurry, noisy mess of data into a clean, sharp list of events. It's like upgrading from a muddy, grainy security camera to a crystal-clear, high-definition system that can tell you exactly who walked through the door, even if they walked in pairs.

This allows scientists to study the universe's most elusive particles with much greater precision, potentially helping us understand everything from the sun's core to the death of stars.

Technical Summary

1. Problem Statement

The Jiangmen Underground Neutrino Observatory (JUNO) utilizes a massive array of 17,612 Large Photo Multiplier Tubes (LPMTs) to detect neutrino interactions. The data acquisition system (DAQ) faces significant challenges:

• Data Volume: Storing full waveforms (1008 samples per event) for all channels is infeasible due to the high trigger rate induced by cosmic muons and natural radioactivity (approx. 4 Hz global trigger rate would generate ~11 TB/day).
• Current Limitations: The current online reconstruction method, Continuous Over Threshold Integration (COTI), integrates charge only when the signal exceeds a threshold. It fails to distinguish between closely spaced photoelectrons (PEs). If two PEs arrive within the pulse width (~40 ns), COTI merges them into a single "hit," leading to:
  • Underestimation of the total charge.
  • Incorrect hit counting (merging two hits into one).
  • Degraded energy resolution, particularly for low-energy events.
• Resource Constraints: While Field Programmable Gate Arrays (FPGAs) on the readout boards (Global Control Units, GCUs) offer real-time processing, they have limited computational resources (specifically DSP slices) and strict latency requirements (~100 ns).

2. Methodology: Real-Time Wiener Deconvolution (RTWD)

The authors propose a novel Real-Time Wiener Deconvolution (RTWD) algorithm implemented directly on the FPGA to reconstruct Time-Charge (TQ) pairs with high precision. The method consists of four main stages:

A. Signal Characterization

• Template Estimation: A mean Single Photoelectron (SPE) impulse response (template) is derived from experimental data by aligning and averaging SPE waveforms.
• Noise Characterization: The Power Spectral Density (PSD) of the noise is estimated using "empty waveforms" (no signal) to characterize the electronic and environmental noise background.

B. Filter Design

Two Finite Impulse Response (FIR) filters are designed based on the signal and noise PSDs:

1. Wiener Filter (Type I, 11-tap): Designed to maximize the Signal-to-Noise Ratio (SNR) by suppressing noise while preserving the signal shape. It acts as a smoothing filter.
2. Deconvolution Filter (Type III, 7-tap): Designed as the reciprocal of the template's frequency amplitude. It acts as a differentiator to "sharpen" the waveform, effectively removing the PMT's impulse response and converting the broad pulse into a sharp spike corresponding to the original PE arrival time.

C. FPGA Implementation Architecture

The algorithm is implemented in the AMD Kintex-7 FPGA on the JUNO GCU using a pipelined VHDL architecture (TQ_reco):

1. Baseline Tracker: Dynamically estimates and subtracts the baseline to handle drift, while ignoring fast transients (PEs).
2. Over-Threshold Check: Identifies candidate hits and triggers the reconstruction pipeline.
3. FIR Filtering: The baseline-subtracted signal passes through the Wiener filter (denoising) and then the Deconvolution filter (pulse sharpening).
4. Peak Finding: Identifies local maxima in the deconvolved waveform. A peak is validated only if it represents a true SPE shape (increasing before, decreasing after), ensuring robust hit counting.

3. Key Contributions

• Algorithmic Innovation: Introduction of a real-time Wiener Deconvolution approach specifically optimized for FPGA constraints, moving beyond simple threshold integration.
• Hardware Efficiency: Successful implementation of complex signal processing (two FIR filters) on a resource-constrained FPGA (Kintex-7) without exceeding DSP limits, utilizing only ~74% of available DSP slices.
• Resolution of Pile-up: The method effectively disentangles closely spaced PEs (down to 0.5x the pulse width separation) that the current COTI algorithm merges.
• Calibration Strategy: Development of a calibration procedure to correct for systematic biases, including a ~5% negative charge bias (due to non-ideal Wiener filtering) and undershoot effects caused by closely spaced hits.

4. Results

The algorithm was validated using a CAEN detector emulator simulating JUNO conditions. Key performance metrics include:

• Hit Identification:
  • COTI: Fails to resolve two hits unless they are separated by >1.0x the pulse width. At 0.5x separation, it incorrectly identifies 99.5% of double-hit events as single hits.
  • RTWD: Correctly identifies the number of hits in >98% of cases, even at 0.5x separation. Its performance is comparable to Offline Deconvolution (which runs on CPUs with unlimited resources), despite running on an FPGA.
• Charge Reconstruction:
  • Raw RTWD: Shows a ~5% underestimation of charge due to residual noise and filter approximation.
  • Calibrated RTWD: After applying a correction curve derived from the deconvolved template, the charge distribution collapses to a single population centered on the expected value (1 PE).
  • Comparison: The "Charge-per-Hit" metric ($Q/N_{hit}$) for RTWD remains close to 1.0 across all temporal separations, whereas COTI jumps to ~2.0 (indicating merged hits) for close separations.
• Resource Utilization: The design consumes 624 out of 840 available DSP48E1 slices (74.28%), confirming that DSP resources are the primary bottleneck but are sufficient for this implementation.

5. Significance

This work demonstrates that advanced signal processing techniques, previously reserved for offline analysis, can be deployed in real-time on FPGA-based readout electronics for large-scale neutrino experiments.

• Physics Impact: By accurately reconstructing low-energy depositions and disentangling pile-up events, RTWD significantly improves the energy resolution and detection efficiency for transient astrophysical phenomena (e.g., supernovae) and low-energy neutrino interactions.
• Scalability: The approach is applicable to any PMT-based detector with FPGA readout, offering a pathway to handle higher data rates and improve physics reach without increasing DAQ bandwidth or storage costs.
• Future Outlook: The authors note that while the current implementation uses a single template for all channels, future iterations could incorporate per-channel template characterization to further enhance performance.