⚛️ high-energy experiments

Robustness of the \textit{smartpixels} classifier for different simulated sensor geometries and non-ideal detector conditions

This paper demonstrates that the \textit{smartpixels} neural network classifier for real-time track filtering in high-granularity pixel detectors maintains robust performance across varying sensor geometries, magnetic fields, and noise levels, with significant efficiency losses due to radiation damage being largely recoverable through network retraining.

Original authors: Danush Shekar, Ben Weiss, Morris Swartz, Corrinne Mills, Jennet Dickinson, Lindsey Gray, David Jiang, Mohammad Abrar Wadud, Daniel Abadjiev, Anthony Badea, Douglas Berry, Alec Cauper, Arghya Ranjan Da

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Danush Shekar, Ben Weiss, Morris Swartz, Corrinne Mills, Jennet Dickinson, Lindsey Gray, David Jiang, Mohammad Abrar Wadud, Daniel Abadjiev, Anthony Badea, Douglas Berry, Alec Cauper, Arghya Ranjan Das, Karri Folan DiPetrillo, Farah Fahim, Rachel Kovach Fuentes, Abhijith Gandrakota, Giuseppe Di Guglielmo, Eliza Howard, Shiqi Kuang, Carissa Kumar, Mia Liu, Petar Maksimovic, Nick Manganelli, Mark S Neubauer, Aidan Nicholas, Emily Pan, Benjamin Parpillon, Jannicke Pearkes, Ricardo Silvestre, Chinar Syal, Amit Trivedi, Keith Ulmer, Jieun Yoo, Eric You

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

In the heart of particle physics, where scientists smash atoms together to understand the fundamental building blocks of the universe, a new challenge has emerged. The machines designed for these collisions are becoming so powerful that they create more data than any computer can possibly process in real time. Imagine a camera taking millions of photos every second; if you tried to save every single image, your hard drive would fill up instantly, and the most interesting moments would be lost in the noise. To solve this, physicists are building smarter detectors that can make decisions before the data even leaves the machine. These detectors, known as pixel sensors, act like a grid of tiny light switches. When a particle passes through, it leaves a trail of electrical signals, or a "cluster," on these switches. The goal is to teach the detector to recognize the shape of that trail and instantly decide which particles are important enough to keep and which are just background clutter to be discarded. This process, happening directly on the sensor chip, is called on-detector data reduction, and it is essential for the next generation of experiments that will explore the deepest secrets of matter.

A team of researchers set out to test how well a specific type of artificial intelligence, a neural network, could perform this filtering job under a wide variety of real-world conditions. They were not just looking at a perfect, idealized sensor in a vacuum; they wanted to know if their "smart pixel" system would survive the messy, harsh environment of a particle collider. They simulated how the system would behave when the physical size of the sensor pixels changed, when the sensors were placed in different parts of the detector where magnetic fields act differently, and when the sensors were damaged by years of radiation. They also tested how the system handled electrical noise, the static that naturally occurs in electronic circuits. Their work focused on a specific algorithm that looks at the shape of the particle's trail to guess its momentum, or how hard it is moving sideways. If the trail is short and straight, the particle is likely moving slowly and can be ignored. If the trail is long and curved, the particle is fast and valuable, so the system keeps it.

The researchers began by building a digital model of the sensor and training the artificial intelligence to recognize these trails. They tested the system with seven different sensor designs, changing the spacing between the pixels and the thickness of the sensor material. They found that the shape of the sensor mattered a great deal. When the pixels were spaced further apart, the system had a harder time seeing the fine details of the particle's trail, and its ability to tell fast particles from slow ones dropped. However, they discovered a clever workaround: making the sensor thicker helped compensate for the wider spacing. A thicker sensor allowed the particle to leave a longer, more distinct trail, which gave the computer enough information to make a good decision even with larger pixels. For the specific sensor design planned for a major upgrade at the Large Hadron Collider, the system successfully identified the fast particles about 92% of the time while discarding nearly 30% of the useless data. This balance is crucial; it means the system keeps the valuable physics data while significantly reducing the amount of information that needs to be sent to the main computers.

The study also looked at how the system performed in different parts of the detector. In the central barrel section, the magnetic field helps bend the paths of the particles, making their trails easier to distinguish. In the endcap sections, where the sensors are oriented differently, the magnetic field does not bend the particles in the same way, making the task harder. The researchers found that the system was indeed less effective in these endcap regions, but it still worked well enough to be useful. They also simulated what happens when the sensors are bombarded by radiation over many years, a process that damages the silicon and changes how the electrical signals travel. When they used the original computer program trained on fresh sensors, the system failed miserably on the damaged ones, throwing away good data and keeping bad data. But here is the key finding: when they simply retrained the computer program using data from the damaged sensors, the system recovered almost all of its performance. This proves that the system is not fragile; it can adapt to the aging of the hardware simply by updating its internal settings.

Finally, the team tested the system's resilience to electrical noise, the random static that can confuse electronic signals. They found that if the noise was high, the system's performance dropped significantly. However, just like with radiation damage, they found that retraining the system on data that included this noise allowed it to maintain its accuracy. Even with noise levels that would typically ruin a standard detector, the retrained system stayed within a few percentage points of its original performance. The researchers also checked if the system could handle a few broken pixels, which are common in real detectors, and found that it was robust enough to ignore them without needing any special adjustments. These results suggest that the technology is ready for the real world. The system is not a one-time solution that breaks when conditions change; it is a flexible tool that can be updated as the detector ages and as the environment shifts. By proving that this intelligent filtering works across different designs, magnetic conditions, and levels of damage, the study provides a solid foundation for building the next generation of particle detectors. These smart sensors will allow physicists to keep the data flowing at unprecedented rates, ensuring that no matter how many particles collide, the most interesting ones are never lost.

Technical Summary: Robustness of the Smartpixels Classifier for Different Simulated Sensor Geometries and Non-Ideal Detector Conditions

Problem Statement
Next-generation collider experiments, such as the High-Luminosity Large Hadron Collider (HL-LHC), will face unprecedented charged-particle densities, particularly near the interaction point. This high density necessitates extremely fine pixel pitches in silicon tracking detectors, which in turn exacerbates bandwidth requirements for data readout. The "Smartpixels" project aims to mitigate this by deploying machine learning (ML) algorithms directly on the detector readout chip (ASIC) to perform real-time data reduction. Specifically, the goal is to filter out low transverse momentum ( $p_T$ ) tracks based on the geometrical shape of charge deposition clusters before they are transmitted off-detector. While previous work established the feasibility of this approach for a single benchmark sensor geometry, a critical gap remained in understanding how the classifier's performance depends on varying sensor designs, operating conditions, and non-ideal detector states (such as radiation damage and noise). Without this understanding, the deployment of such systems in diverse detector regions (barrel vs. endcap) and over the detector's lifetime is uncertain.

Methodology
The authors utilized the PixelAV simulation framework to generate datasets for training and validating a neural network (NN) classifier. The simulation included detailed models of silicon detectors, accounting for drift, diffusion, the Hall effect, and radiation-induced charge trapping.

Datasets: Two primary datasets were created: a "flat- $p_T$ " dataset (uniform distribution from 0.03 to 5 GeV) for training to avoid momentum bias, and a "physical- $p_T$ " dataset (falling distribution typical of 13 TeV $pp$ collisions) for evaluation.
Sensor Configurations: Seven distinct sensor geometries were surveyed, varying in pixel pitch (along $x$ and $y$ ) and sensor thickness. These included the baseline $50 \times 12.5 \times 100$ $\mu m^3$ , the CMS Phase-2 upgrade geometry ( $100 \times 25 \times 100$ $\mu m^3$ ), and a thicker variant ( $100 \times 25 \times 150$ $\mu m^3$ ).
Non-Ideal Conditions: Simulations covered:
- Geometry: Barrel regions (where magnetic field is perpendicular to the sensor) and Endcap regions (where the field is parallel, eliminating Lorentz drift).
- Irradiation: Sensors were simulated at fluences corresponding to 0, 370, and 1100 $fb^{-1}$ of integrated luminosity.
- Noise: Gaussian noise was injected into the charge collection simulation based on Equivalent Noise Charge (ENC) calculations derived from Cadence Virtuoso/Spectre ASIC simulations.
Algorithm: A neural network classifier was employed, taking as input the "cluster $y$ -profile" (integrated charge projected onto the $y$ -axis) and the incident azimuthal position ( $y_0$ ). The network outputs probabilities for three classes: high $p_T$ tracks, and two low $p_T$ categories (distinguishing between oppositely charged tracks curving in opposite directions). The network architecture consists of one dense hidden layer (128 neurons) and one output layer (3 neurons).

Key Contributions
This paper provides the first systematic study of the Smartpixels classifier's robustness across a wide range of design parameters and operational conditions. Key contributions include:

Geometry Dependence: Quantifying how pixel pitch and sensor thickness affect $p_T$ discrimination.
Region Comparison: Evaluating performance differences between barrel and endcap regions, specifically isolating the impact of Lorentz drift.
Radiation Hardness: Demonstrating the degradation of a fixed-weight classifier under irradiation and proving that performance can be recovered through retraining.
Noise Resilience: Assessing the impact of electronic noise and defective pixels, showing that retraining on noise-injected data restores performance.

Results

Sensor Geometry: Smaller pixel pitch in the bending direction ( $y$ ) and increased sensor thickness improve $p_T$ discrimination by providing better-resolved cluster features. Increasing the $y$ -pitch degrades performance, but this loss can be partially compensated by increasing sensor thickness. For the CMS HL-LHC sensor geometry ( $100 \times 25 \times 100$ $\mu m^3$ ), the classifier achieved a signal efficiency of $(91.9 \pm 0.7)\%$ and a data reduction of $(29.7 \pm 1.0)\%$ . The thicker variant ( $100 \times 25 \times 150$ $\mu m^3$ ) recovers some performance loss from the larger pixel pitch but does not perform as well as the $50 \times 12.5 \times 100$ $\mu m^3$ baseline.
Lorentz Drift: The absence of Lorentz drift (in endcap regions or 3D sensors) reduces performance. At a $p_T$ boundary of 0.2 GeV, endcap sensors showed reduced performance compared to barrel sensors. However, the specific impact of removing Lorentz drift (simulated by setting the magnetic field to zero in barrel pixels) resulted in an absolute performance drop of less than 6% across all metrics.
Irradiation: Accumulated radiation damage significantly alters cluster shapes, reducing signal efficiency compared to the baseline by approximately 30–60% in absolute terms if the network weights are not updated. However, retraining the network on irradiated data and updating the ASIC weights recovers nearly all performance, keeping signal efficiency and data reduction within 7% and 2% of their original values, respectively, even at 1100 $fb^{-1}$ .
Noise: Increasing noise levels (up to 100 electrons) causes a considerable drop in performance for a network trained on noiseless data. However, retraining on noise-injected datasets maintains performance within 6% of the baseline. The network also demonstrated robustness to defective pixels (<1% rate), with performance remaining within 1% of the nominal case without retraining.

Significance and Outlook
The paper concludes that while non-ideal conditions (noise, radiation, geometry variations) severely degrade the performance of a static classifier, the Smartpixels approach remains viable through the use of reprogrammable logic in the ASICs. The ability to periodically retrain the network and update weights allows the system to adapt to sensor aging and changing operating conditions.

The authors state that these findings inform design considerations for future detectors aiming to implement intelligent, on-sensor data reduction. They emphasize that for use in the first layer of a trigger system, at least two layers of smartpixels are required. The current work focuses on the innermost layer, which has the most stringent bandwidth requirements. Future work will extend these studies to displaced tracks (relevant for long-lived particles), multi-layer tracking systems, and the hybridization of Smartpixels ASICs with pixel sensors for test beam demonstrations. The results suggest that this technology can reduce data throughput while maintaining high physics performance, a necessity for the success of next-generation collider experiments.

More like this