Optimisation of a silicon-tungsten electromagnetic calorimeter energy response to photons

This paper presents a reoptimization of the silicon-tungsten electromagnetic calorimeter (SiW-ECAL) design for future circular colliders by developing machine learning-based reconstruction methods that significantly improve energy resolution and correct energy leakage.

The Gist

Imagine you are trying to take a perfect photograph of a fireworks display. To get a great picture, you need a camera that is incredibly sharp (high resolution) and sensitive enough to catch even the faintest sparks (low energy) without getting overwhelmed by the bright explosions (high energy).

This paper is about upgrading the "camera" used by future particle physics experiments. Specifically, it's about a device called a Silicon-Tungsten Electromagnetic Calorimeter (SiW-ECAL). Think of this device as a giant, ultra-detailed 3D grid made of alternating layers of heavy metal (tungsten) and sensitive silicon sensors. When a particle (like a photon) hits this grid, it creates a "shower" of smaller particles, and the grid measures how much energy is released.

Here is the simple breakdown of what the researchers did and found:

The Problem: The Old Camera Wasn't Perfect

For years, scientists have used this silicon-tungsten grid to measure particle energy. They usually did this in two simple ways:

1. The "Sum" Method: Just adding up all the energy detected.
2. The "Count" Method: Just counting how many times the sensors were triggered.

The problem is that these old methods struggle with low-energy particles (like faint sparks) and sometimes lose track of high-energy ones. Also, the design of the grid itself hasn't changed much in decades, even though our ability to process data has exploded.

The Solution: Teaching the Camera to "Think"

The researchers decided to stop using simple math and start using Machine Learning (ML). Imagine teaching a computer to look at the pattern of the particle shower and guess the energy, rather than just doing a simple sum.

They tested two types of AI:

• The "Smart Calculator" (MLP): A standard, fast, and efficient neural network.
• The "Super-Computer" (DGCNN): A very complex model that looks at the connections between every single sensor hit.

The Result: The "Smart Calculator" (MLP) was the winner. It was almost as good as the "Super-Computer" but much faster and cheaper to run. It improved the energy measurement accuracy by about 20% for low-energy particles and fixed errors where energy "leaked" out of the detector at high energies.

Redesigning the Grid (The Re-Optimization)

Once they had this smart AI, they asked: "If we have this smart AI, do we need to build the grid exactly the same way we always have?"

They tested different designs to see what worked best with their new AI:

1. Thickness (The "Shield"):

   • Old idea: You need a very thick wall of tungsten to catch all the energy.
   • New finding: Because the AI is so good at fixing "leaks," you can make the wall thinner (about 18 layers of tungsten instead of 24) and still get the same great results. This saves a lot of money and material (about 30% less cost).

2. Sampling Layers (The "Frames"):

   • Old idea: More layers of sensors mean a better picture.
   • New finding: Yes, more layers help, but only up to a point. After 40 layers, adding more doesn't help much. They recommend 30 layers as the sweet spot.

3. Sensor Thickness (The "Film"):

   • Finding: Thicker silicon sensors work better. They are planning to use 0.75 mm thick sensors for the next version.

4. Cell Size (The "Pixels"):

   • Surprise: You might think smaller pixels (cells) mean a sharper image. But for this specific setup, smaller cells actually made the picture worse.
   • Why? When cells are tiny, a single particle might hit multiple cells, confusing the count. The AI couldn't fix this confusion. They found that 5 mm cells are the best size for now.

The Bottom Line

By combining a smarter computer program (Machine Learning) with a slightly redesigned physical detector, the researchers found a way to build a particle detector that is:

• More accurate (especially for low-energy particles).
• Cheaper and lighter (because it can be thinner).
• Ready for the future (suitable for upcoming particle colliders like FCC-ee or CEPC).

In short, they didn't just upgrade the software; they used the software to realize they could build a better, cheaper hardware design.

Technical Summary

1. Problem Statement

Future high-energy physics colliders (such as FCC-ee, CEPC, ILC, and CLIC) require exceptional Jet Energy Resolution (JER) to distinguish complex final states (e.g., $ZZ$ vs. $WW$). Achieving a JER of $\sim3.8\%$ or better necessitates the use of Particle Flow Algorithms (PFA), which rely on highly granular calorimeters to individually reconstruct particle showers.

The Silicon-Tungsten Electromagnetic Calorimeter (SiW-ECAL) is a mature technology developed for linear colliders (ILC). However, its design requires re-optimization due to three emerging factors:

1. Low-Energy Dominance: Particles in jets are concentrated in the low-energy region ($<2$ GeV). Traditional energy summation methods degrade in this range, while hit-counting methods perform better but are not fully integrated into current designs.
2. Machine Learning (ML) Potential: Advanced ML techniques (e.g., GNNs, DGCNN) can exploit the high granularity of SiW-ECAL to outperform traditional reconstruction, yet the detector geometry has not been optimized specifically for these algorithms.
3. Circular Collider Constraints: Proposed circular Higgs factories have lower center-of-mass energies than linear colliders, allowing for a reduced material budget. The current design may be over-engineered for these specific energy ranges.

2. Methodology

Simulation and Baseline Geometry

• Framework: Simulations were performed using DD4hep (built on ROOT and Geant4).
• Prototype: A flexible 120-layer prototype was modeled, consisting of 0.7 mm tungsten absorbers ($\sim0.2 X_0$) and 0.75 mm silicon sensors.
• Baseline Configuration: The study used a 30-layer sampling configuration with $5 \times 5 \text{ mm}^2$ silicon cells as the reference geometry.
• Data: Photon datasets were generated with energies ranging from 0.1 GeV to 60 GeV (1.25M events for training/validation; 50k events per energy point for testing).

Reconstruction Approaches

The study compared four reconstruction methods:

1. Sum E: Traditional summation of deposited energy.
2. N Hits: Counting hits above a threshold (0.1 MIP).
3. MLP (Multi-Layer Perceptron): A neural network using 152 physics-motivated features (energy sums, hit counts per layer, ratios, etc.).
4. DGCNN (Dynamic Graph Convolutional Neural Network): A state-of-the-art graph-based model using $k$-nearest neighbors (KNN) on hit positions.

Key Technical Decisions:

• Loss Function: The Huber loss was selected for the MLP. It combines Mean Square Error (MSE) for small residuals and a linear penalty for large residuals ($>5\%$), ensuring the model prioritizes linearity before optimizing resolution.
• Model Selection: While DGCNN showed superior performance at high energies, it was deemed computationally prohibitive for iterative geometry optimization. The MLP was chosen as the primary tool for re-optimization due to its speed, comparable low-energy performance, and efficiency.

Optimization Process

The SiW-ECAL geometry was re-optimized by varying:

• Total Absorber Thickness: Ranging from $16 X_0$ to $24 X_0$.
• Sampling Layers: Varying the number of readout layers (20 to 60).
• Silicon Thickness: Testing different sensor thicknesses.
• Cell Size: Testing cell sizes from $1 \times 1 \text{ mm}^2$ to $10 \times 10 \text{ mm}^2$.

3. Key Contributions

1. ML-Driven Reconstruction: Demonstrated that an MLP-based approach significantly outperforms traditional "Sum E" and "N Hits" methods, particularly in the low-energy regime ($<5$ GeV).
2. Correction of Energy Leakage: Showed that ML can effectively correct for energy leakage in thinner calorimeters ($16 X_0$), recovering performance close to thicker designs.
3. Geometry Re-Optimization: Provided specific, data-driven recommendations for the next generation of SiW-ECAL designs tailored for circular colliders and ML reconstruction.
4. Counter-Intuitive Findings: Identified that smaller cell sizes do not necessarily improve resolution due to the "right-hand tail" effect in hit multiplicity distributions when particles traverse multiple small cells.

4. Key Results

Performance Improvements

• Resolution Gain: The MLP method achieved an approximate 20% improvement in energy resolution in the low-energy range compared to traditional methods. At 100 MeV, improvements reached up to 30%.
• Linearity: The Huber loss function successfully maintained linearity within 5% across the energy spectrum, correcting the non-linearity observed in "N Hits" methods at high energies.

Optimized Geometry Recommendations

Based on the MLP reconstruction, the study proposes the following design parameters:

• Material Thickness: A minimum of $18 X_0$ of tungsten is recommended. This reduces the material budget by approximately one-third compared to CEPC/ILD designs while maintaining performance for 60 GeV photons (after MLP correction for leakage).
• Sampling Layers: Performance saturates beyond 40 layers. 30 sampling layers are recommended as the optimal balance between performance and cost.
• Silicon Thickness: Thicker sensors are preferred. 0.75 mm silicon is planned for the next generation, offering better resolution above 1 GeV.
• Cell Size: Contrary to the assumption that finer granularity is always better, the optimal cell size remains 5 mm. Smaller cells ($1$ mm) degrade resolution because single particles often cross multiple cells, creating a tail in the hit distribution that complicates reconstruction.

5. Significance and Outlook

• Cost and Feasibility: The proposed re-optimized design significantly reduces the material budget and associated costs for future circular Higgs factories (FCC-ee, CEPC) without compromising physics performance.
• Hardware-Software Synergy: The study highlights a critical shift where detector design is no longer static but is co-optimized with advanced reconstruction algorithms.
• Future Directions:
  • The conclusion regarding cell size may evolve as more sophisticated ML models (e.g., those better exploiting transverse granularity) are developed.
  • Future work could explore specialized layer designs (e.g., thinner absorbers in the first 10 layers) to further leverage ML capabilities.
  • The developed ML methods are applicable to other calorimetry studies, including Hadronic Calorimeters (HCAL) and test beam data analysis.

In summary, this paper establishes that integrating Machine Learning into the design loop of the SiW-ECAL allows for a leaner, more cost-effective detector that meets the stringent resolution requirements of future particle physics experiments.