🔬 physics

Exploiting the latent space of deep AutoEncoders for the identification of signal pulses in noisy time-series

This paper proposes a data-driven method using convolutional variational autoencoders to identify signal pulses in noisy time-series by mapping waveforms to a compressed latent space, where regions corresponding to pure noise are isolated to effectively tag candidate signals for low-energy nuclear recoil detection in Liquid Argon Time Projection Chambers.

Original authors: Gioacchino Alex Anastasi, Sebastiano Francesco Albergo, Marzio De Napoli, Noemi Pino, Sebastiana Maria Puglia, Alessia Rita Tricomi

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gioacchino Alex Anastasi, Sebastiano Francesco Albergo, Marzio De Napoli, Noemi Pino, Sebastiana Maria Puglia, Alessia Rita Tricomi

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 find a single, tiny whisper in a room that is constantly filled with the chaotic, crackling static of a thousand old radios. That is the challenge facing scientists hunting for Dark Matter. They are looking for "ghost particles" called WIMPs that might bump into atoms, creating a faint, delayed flash of light. But in their giant tanks of liquid argon, this tiny signal is often drowned out by random electronic noise that looks suspiciously like a real signal.

The paper you are reading proposes a clever, data-driven way to solve this "needle in a haystack" problem using a type of artificial intelligence called a Variational Convolutional Autoencoder. Think of this AI not as a detective looking for clues, but as a master chef who learns to recognize the exact taste of "noise" so well that it can instantly spot anything that doesn't taste like noise.

The Training Kitchen

To teach this AI, the researchers didn't just feed it real data; they cooked up a massive batch of 7,500 synthetic waveforms (simulated data) to train the model, and then saved another 7,500 to test it.

Here is how they made the food:

The Noise: They created a messy background by randomly adding and subtracting 100 tiny, independent "noise pulses." This creates a fluctuating static that can sometimes look just as loud as a real signal.
The Signal: They added a single, specific type of pulse (shaped like a log-normal curve) that mimics the real light flash from a Dark Matter hit.
The Challenge: They made the signal vary wildly in strength. Some were huge, but others were so faint they were barely above the noise floor—down to the level of a single electron.

The AI's job was to look at these 10,000-sample-long time-series waves, squish them down into a tiny, compressed summary, and then try to rebuild the original wave perfectly. This process is called an Autoencoder.

The "Squish" and the Secret Map

Usually, when you squish a complex image into a tiny file, you lose details. But this AI is special. It uses a Variational approach, which means instead of just memorizing the data, it learns the probability of what the data looks like.

Imagine the AI has a magical map (called the latent space).

When the AI sees a waveform that is pure noise, it doesn't scatter the point randomly on the map. Instead, it consistently drops the point into one specific, crowded neighborhood. It's like how every time you hear a specific type of static, your brain knows exactly where to put that sound in your memory.
When the AI sees a waveform with a real signal pulse, even a tiny one, the point on the map moves away from that crowded noise neighborhood. The stronger the signal, the farther it travels from the center.

The researchers found that after 150 epochs (training rounds), the AI had learned this map perfectly. The "noise-only" waveforms formed a dense, tight cluster. Any wave with a signal, no matter how small, drifted away from this cluster.

The "Distance" Rule

So, how do they find the signals? They invented a simple rule based on distance.

They found the exact center of the "noise neighborhood" (the noise centroid).
They drew a circle around it. The radius of this circle is set so that 99% of all pure noise waves fall inside it. This means there is only a 1% chance of a noise wave accidentally falling outside and being falsely identified as a signal (a "False Positive").
If a new, unknown wave lands outside this circle, the AI tags it as a candidate signal.

What the Results Show

When they tested this on the fresh batch of 7,500 simulated waves:

Huge Signals: If the signal was strong, the AI caught 100% of them.
Tiny Signals: As the signals got smaller and closer to the size of random noise bumps, the AI started missing them. This is expected because if a signal is almost identical to the noise, even a smart AI can't tell them apart.
The "1% Rule": The method successfully kept the false alarms low, ensuring that when they say "we found a signal," they are very confident it's not just a glitch.

What This Is (and Isn't)

It is important to know exactly what this paper proves and what it doesn't.

It is a simulation: The results come entirely from computer-generated data. The paper explicitly states this is a "toy model" and a "Monte Carlo benchmark." They have not yet applied this to real experimental data from the Recoil Directionality (ReD) experiment in Catania, Italy, though that is the ultimate goal.
It rules out simple reconstruction: The authors argue that simply looking at how well the AI can rebuild the wave (reconstruction error) is not the best way to find signals. Instead, they proved that looking at where the data sits on the AI's internal map (the latent space) is a much more robust way to separate noise from signals.
It is not a magic bullet for all noise: The method relies on the noise having specific, consistent properties. If the noise in the real experiment changes drastically, the "noise centroid" would need to be recalculated.

The Bottom Line

This paper suggests that by training an AI to understand the "shape" of noise so deeply that it can map it to a specific spot, scientists can create a highly efficient filter. This filter can spot the faintest whispers of Dark Matter in the roar of the universe, provided the noise behaves the way they simulated. It's a promising new tool for the DarkSide project, designed to help them see the invisible, but for now, it remains a very convincing proof-of-concept in the world of simulations.

Technical Summary: Exploiting the Latent Space of Deep AutoEncoders for Signal Identification

Problem Statement
In the context of direct dark matter searches using dual-phase Liquid Argon Time Projection Chambers (LAr TPCs), such as the Recoil Directionality (ReD) experiment, identifying low-energy nuclear recoil signals is increasingly challenging. For WIMP masses below $\sim$ 1 GeV/ $c^2$ , the prompt scintillation signal (S1) becomes too weak to detect, necessitating reliance on the delayed electroluminescence signal (S2). However, at these low energy thresholds (a few keV), S2 pulses are often comparable in amplitude to stochastic baseline fluctuations, electronic noise, and spurious single-electron extractions. Traditional pulse-finding algorithms struggle to distinguish genuine signals from this non-Gaussian noise background. The paper addresses the need for a data-driven, unsupervised method to identify these low-amplitude signal pulses within long, noisy time-series without relying on labeled training data for every signal intensity.

Methodology
The authors propose a self-supervised framework based on a Variational Convolutional Autoencoder (VCAE). Unlike standard autoencoders that map inputs to discrete points, the VCAE models the underlying probability distributions of the data, enforcing a continuous and regularized latent space.

Synthetic Dataset Generation: A dataset of 15,000 synthetic waveforms (10,000 time-bins each) was generated to mimic ReD experimental conditions. Each waveform $W(t)$ is a superposition of a non-Gaussian noise background $B(t)$ and a log-normal shaped signal pulse $S(t)$ . The noise is constructed by summing 100 random log-normal components, creating a fluctuating background where spurious peaks can match or exceed small signal amplitudes. Signal amplitudes ( $A_S$ ) are sampled to span from 1 down to $\sim$ 0.001, covering the single-photoelectron regime.
Network Architecture: The model utilizes a 1D VCAE with a bottleneck dimension of 2.
- Encoder: Three blocks of 1D Convolutional layers (with ReLU activation) followed by AveragePooling1D, reducing the waveform length while increasing feature channels.
- Bottleneck: The encoder outputs mean ( $\mu$ ) and log-variance ( $\log \sigma^2$ ) vectors for a Gaussian distribution. A reparameterization trick samples the latent vector $z$ .
- Decoder: Mirrors the encoder using UpSampling1D and Transposed Convolutional layers to reconstruct the input waveform.
- Loss Function: The optimization minimizes a combination of Mean Squared Error (MSE) for reconstruction fidelity and Kullback–Leibler (KL) divergence to regularize the latent space toward a standard normal distribution.
Training Strategy: To ensure robustness against random weight initialization (which caused non-convergence in ~20% of cases), a two-step procedure was employed: a "cooldown" phase to select the best-performing initial model, followed by a full training phase using the ADAM optimizer and learning rate scheduling. A 5-fold cross-validation scheme was used to validate statistical robustness.

Key Contributions and Findings
The primary contribution is the demonstration that the latent space representation of pure noise waveforms naturally clusters into a dense, restricted region, distinct from waveforms containing signal pulses.

Latent Space Topology: The study reveals that the mean latent vector ( $z_{mean}$ ) for noise-only waveforms accumulates in a specific, narrow region of the 2D latent plane. As signal amplitude increases, the encoded representations radially diverge from this "noise centroid."
Classification Criterion: Based on this geometric separation, the authors developed a distance-based selection algorithm. A "noise centroid" ( $z_{noise}$ ) is defined as the mode of the noise-only distribution. The Euclidean distance ( $d$ ) of any new waveform's latent encoding from this centroid is calculated.
Thresholding: A threshold ( $d^*$ ) is established at the 99% quantile of the distance distribution for noise-only data. Waveforms with $d \le d^*$ are classified as noise; those with $d > d^*$ are flagged as containing a signal.

Results

Performance: When applied to a test set of 7,500 freshly generated waveforms, the method achieved 100% identification efficiency for events with large pulses.
Sensitivity: The efficiency decreases gradually as signal amplitudes approach the level of accidental noise pulses, which is an expected behavior given the data-driven nature of the threshold.
False Positive Rate: By construction, the method guarantees a 1% False Positive Rate (mislabeling noise as signal), effectively limiting background contamination.
Generalization: The results were consistent across all 5 folds of the cross-validation, indicating that the model learned intrinsic physical features rather than overfitting specific training samples.

Significance and Claims
The paper claims to present an innovative, self-supervised methodology that overcomes the limitations of traditional pulse-finding algorithms for low-energy signals in LAr TPCs. By exploiting the topology of the learned latent space rather than relying on reconstruction error, the approach provides a robust classification criterion for identifying delayed electroluminescence signals.

The authors position this work as a core component of the DAIDREAM initiative, aimed at maximizing the discovery potential of rare-event searches in astroparticle physics. They assert that this framework is designed to fully exploit measurements from experiments like ReD, offering a sensitivity at least comparable to conventional reconstructions while being adaptable to the specific noise characteristics of the detector. The paper concludes that the results validate the feasibility of using unsupervised deep learning for high-efficiency pulse identification down to the physical limits imposed by intrinsic background noise.