An investigation of fast simulation techniques for pion showers using kernel density estimators with the CALICE AHCAL Technological Prototype

This paper presents a data-driven fast simulation algorithm for pion showers in the CALICE AHCAL Technological Prototype, developed using kernel density estimators on 2018 test beam data, which achieves excellent agreement with measured observables and includes a method for interpolating showers at arbitrary energies.

The Gist

Imagine you are trying to predict how a specific type of rainstorm (a "pion shower") will splash when it hits a giant, complex sponge (a particle detector called the AHCAL).

In the world of high-energy physics, scientists usually try to predict these splashes using a super-computer simulation called Geant4. Think of Geant4 as a master chef trying to recreate a dish from scratch by understanding every single chemical reaction of the ingredients. It's incredibly accurate, but it takes a long time to cook—sometimes days to simulate just a few storms.

This paper introduces a new, much faster way to predict these splashes. Instead of cooking from scratch, the researchers decided to learn from the actual rainstorms that already happened.

Here is how they did it, broken down into simple steps:

1. The Problem: Too Much Cooking Time

The standard method (Geant4) is like trying to simulate the physics of every single water droplet hitting the sponge. It's precise, but it's slow. For massive experiments like those at CERN, they need to simulate millions of storms, and waiting days for each one isn't practical. They needed a "fast food" version that still tasted like the real thing.

2. The Solution: The "Cheat Sheet" (Kernel Density Estimators)

The researchers looked at real data collected in 2018 at CERN. They had recorded exactly how 10,000 real pion storms hit the detector.

Instead of trying to calculate the physics, they used a mathematical tool called a Kernel Density Estimator (KDE).

• The Analogy: Imagine you have a photo of a crowd of people. You want to guess where a new person will stand in the crowd. Instead of calculating the wind, the gravity, and the social anxiety of every single person, you just look at the photo and say, "Most people stand here, so the new person will probably stand here too."
• How it works: The KDE takes the real data points (the actual hits on the detector tiles) and creates a smooth "map" of probability. It says, "Based on what we saw before, there is a 90% chance a hit will happen in this specific spot with this specific energy."
• The Result: They can now generate a brand new, fake storm by simply "sampling" from this map. It's like rolling a die that is weighted to match the real world perfectly.

3. The Test: Does the Fake Rain Look Real?

They ran their new "fast simulation" and compared it to two things:

1. The Real Data: The actual storms recorded in 2018.
2. The Slow Simulation: The traditional Geant4 method.

The Verdict: The fast simulation was a huge success.

• It matched the real data almost perfectly.
• In some cases, it was actually better than the slow simulation (Geant4), which sometimes had tiny errors.
• It captured complex details, like how the energy spreads out or how the "center of gravity" of the storm shifts.
• Speed: It was roughly 1,000 times faster than the traditional method. Simulating 10,000 storms took a few minutes instead of several days.

4. The Magic Trick: Predicting Storms They Never Saw

There was one catch: The fast simulation only worked for the specific energy levels they had recorded (e.g., 40 GeV, 80 GeV, 120 GeV). What if they needed to simulate a 60 GeV storm, which they didn't record?

They developed an Interpolation method.

• The Analogy: Imagine you know exactly how a 40-year-old and an 80-year-old walk. You want to know how a 60-year-old walks. You don't need to measure a 60-year-old; you can just take a step from the 40-year-old and a step from the 80-year-old, and blend them together.
• How it works: To simulate a 60 GeV storm, the algorithm takes a "snapshot" of a 40 GeV storm and an 80 GeV storm. It mathematically blends them together, giving more weight to the one closer to 60.
• The Result: This worked beautifully for almost everything. The simulated 60 GeV storms looked just like real data. The only thing that didn't match perfectly was the exact number of hits (the "count" of splashes), which showed a double-peak instead of a single smooth curve. But for everything else—energy, shape, and spread—it was spot on.

Summary

The paper presents a "fast forward" button for particle physics simulations.

• Old Way: Calculate every physical law from scratch (Slow, accurate, but expensive).
• New Way: Learn from real photos of the event and generate new ones based on patterns (Fast, highly accurate, and data-driven).

They proved that by using real data and smart math (KDEs), they can simulate how particles hit a detector thousands of times faster than before, while still getting the physics right. They even figured out how to guess what happens at energy levels they haven't tested yet, by blending the results of the levels they have tested.

What they didn't do: They did not test this on other types of particles (like electrons or muons) in this specific study, nor did they try to predict energies outside the range of their data (extrapolation). They stuck strictly to pion showers within the 10 to 200 GeV range.

Technical Summary

Problem Statement
High-energy physics experiments rely heavily on detailed simulations of particle interactions with detector material to validate designs and interpret experimental data. While Monte-Carlo frameworks like Geant4 provide high accuracy, they are computationally expensive and time-consuming. To address this, "fast simulations" are employed to capture essential interaction information with reduced resource requirements. However, existing fast simulation approaches often rely on neural networks trained on simulated data, requiring complex architecture optimization and training procedures. Furthermore, data-driven methods are typically confined to the specific energy ranges available in the training dataset, lacking robust mechanisms to predict shower behavior at arbitrary, unmeasured energies.

Methodology
This work presents a data-driven fast simulation algorithm for pion showers in the CALICE AHCAL Technological Prototype, utilizing Kernel Density Estimators (KDEs). The methodology is grounded in a test beam dataset recorded at CERN in 2018, covering negatively charged pion beams with initial energies ranging from 10 GeV to 200 GeV.

1. KDE Implementation: The authors model the probability density functions (PDFs) of hit energy distributions for individual calorimeter tiles. Instead of parametric assumptions, they use non-parametric KDEs with Gaussian kernels. The input consists of $n=10,000$ real pion shower events, where each event is represented as a high-dimensional vector ($d=36,518$) containing hit energies relative to the shower start layer and the lateral center of gravity (CoG). A bandwidth of $h=0.01$ MIP was optimized to balance smoothing and structural preservation.
2. Event Generation: Simulated events are generated by sampling from the estimated PDFs. To preserve correlations between tiles, the simulation generates all 36,518 hit energies for an event simultaneously.
3. Energy Interpolation: To extend the simulation to energies not present in the dataset, an interpolation algorithm was developed. For a target energy $E_{int}$, the algorithm selects two reference energies ($E_{small} < E_{int} < E_{large}$). It generates events for both references, maps the cumulative distribution function (CDF) of the lower energy to the higher energy to establish a correspondence, and then linearly weights the resulting hit energies based on the proximity of the reference energies to the target. This process is performed bidirectionally to mitigate systematic biases.

Key Contributions

• Data-Driven KDE Framework: The paper introduces a model-independent approach using KDEs to simulate hadronic showers directly from experimental test beam data, avoiding the training complexities associated with neural network-based methods.
• High-Dimensional Correlation Preservation: The method successfully simulates the full kinematic structure of pion showers, including complex correlations between hit energies across the entire detector volume, rather than just marginal distributions.
• Interpolation Algorithm: A specific algorithmic approach is introduced to interpolate hit energy distributions between known beam energies, allowing for the prediction of shower behavior at arbitrary energies within the dataset's bounds.
• Computational Efficiency: The study quantifies the computational speed-up of the fast simulation compared to full Geant4-based simulation.

Results
The performance of the fast simulation was validated against the 2018 test beam data and full simulation (Geant4 via CaliceSoft) across various kinematic variables:

• Kinematic Variables: The KDE-based simulation demonstrated excellent agreement with data for total energy, shower radius, longitudinal CoG, energy fractions, and shower moments (variance, skewness, kurtosis). In several distributions (e.g., hit energy and total energy peaks), the fast simulation performed slightly better than the full simulation, which exhibited minor shifts or peak height discrepancies.
• Correlations: Two-dimensional correlation plots and correlation matrices confirmed that the fast simulation accurately reproduces the linear correlations (and anti-correlations) between variables such as total energy, CoG, and shower radius. The correlation factors differed by at most 0.11 compared to data.
• Interpolation Performance: The interpolation algorithm successfully predicted kinematic distributions for intermediate energies (e.g., 60 GeV and 120 GeV) with high fidelity. However, the authors noted significant deviations in the "number of hits" distribution, where the interpolated results showed double peaks rather than the single maximum observed in data.
• Computational Speed: The fast simulation achieved a speed-up factor of approximately $O(1000)$ compared to full simulation. Generating 10,000 events took roughly 227 seconds (approx. 23 ms/event) for the fast simulation, compared to 4.81 days (approx. 42 s/event) for full simulation.

Significance and Claims
The paper claims that this data-driven approach offers a reliable and efficient alternative to traditional full simulations and neural network-based methods. By leveraging real experimental data, the simulation captures realistic detector responses and imperfections inherent to the setup. The authors emphasize that the method is particularly effective at reproducing the kinematic behavior and correlation structures of pion showers.

The significance of the work lies in its ability to provide accurate predictions for pion showers at arbitrary energies within the measured range without the computational cost of full simulation. While the interpolation method shows excellent agreement for most variables, the authors modestly acknowledge its limitations regarding the "number of hits" distribution and the challenges of extrapolation beyond the dataset's energy boundaries. The work establishes a foundation for future improvements, such as dimensionality reduction, the inclusion of timing information, and the potential extension to other particle types, aiming toward a comprehensive data-driven simulation prototype for highly granular calorimeters.