A Neural-Network Framework for Tracking and Identification of Cosmic-Ray Nuclei in the RadMap Telescope

This paper presents a neural-network framework utilizing Geant4 simulations to reconstruct cosmic-ray nuclei properties in the RadMap Telescope, achieving high angular resolution, charge separation accuracy (up to 99.8% for hydrogen), and energy resolution (<20% below 1 GeV/n) to enable precise determination of astronaut radiation doses.

The Gist

The Big Picture: Why Do We Need This?

Imagine you are planning a long road trip to Mars. You know the road is dangerous, but you don't have a map. You also don't know exactly what kind of "weather" (radiation) you'll encounter. Will it be a light drizzle of harmless particles, or a hurricane of heavy, bone-damaging cosmic rays?

Currently, our spacesuits and ships are like umbrellas that work okay in the rain but might fail in a hurricane. To build better protection, we need to know exactly what the "weather" is made of. We need to count the particles, identify if they are light (like hydrogen) or heavy (like iron), and measure how much energy they have.

The problem? Most current space detectors are like blurry security cameras. They can tell you something hit the wall, but they can't clearly see if it was a pebble or a bowling ball, or how fast it was moving.

The Solution: The RadMap Telescope

The authors built a new detector called the RadMap Telescope. Think of it not as a camera, but as a giant, 3D grid of glowing spaghetti.

• The Hardware: Inside the detector, there are 1,024 plastic fibers (the "spaghetti") stacked in layers. When a cosmic ray particle zips through them, it makes the plastic glow (scintillate).
• The Challenge: When a particle hits, it leaves a trail of light. But because the particles move so fast and interact in complex ways, the trail looks messy. A heavy iron nucleus might look like a fuzzy, wide smear, while a light proton looks like a thin, sharp line.
• The Old Way: In the past, scientists tried to use complex math formulas to guess what the particle was. It was like trying to solve a Rubik's cube by hand while running a marathon. It took 15 minutes to analyze just one particle. That's too slow for real-time space monitoring.

The New Trick: The "AI Detective"

This paper introduces a Neural Network (a type of Artificial Intelligence) that acts like a super-smart detective. Instead of doing math by hand, the AI looks at the "glow patterns" left by the particles and instantly figures out what they are.

Here is how the AI solves the three main mysteries:

1. The "Where and Which Way?" (Track Reconstruction)

The Analogy: Imagine you see a car's headlights on a foggy night. You need to figure out if it's coming toward you, going away, or driving sideways.
The AI's Job: The detector sees the light in two different "shadows" (projections). The AI looks at these shadows and calculates the exact angle of the particle's path.
The Result: It can tell the direction of the particle with incredible precision (better than 1.4 degrees). It's like being able to tell if a car is driving North-North-East just by looking at its taillights from a mile away.

2. The "Who Are You?" (Charge Identification)

The Analogy: Imagine you are at a party. You can't see the guests' faces, but you can hear how loud they are shouting. A whisper might be a child (Hydrogen), a normal voice might be an adult (Carbon), and a roar might be a giant (Iron).
The AI's Job: Heavier particles make the plastic fibers glow much brighter and wider than light ones. The AI analyzes the "brightness and width" of the trail to guess the particle's identity (its atomic number, Z).
The Result:

• For light particles (Hydrogen and Helium), the AI is 99.8% accurate. It's basically perfect.
• For medium particles (up to Oxygen), it's over 95% accurate.
• For heavy particles (like Iron), it gets a bit trickier because the "roars" start to sound similar. However, it can still tell you, "This is definitely an Iron-like particle," even if it's not 100% sure of the exact number.

3. The "How Fast?" (Energy Measurement)

The Analogy: If you see a car skid, you can guess how fast it was going by how long the skid marks are.
The AI's Job: The AI looks at how much energy the particle dumped into the fibers. If the particle stops inside the detector, the AI adds up all the light to get the total energy. If it flies right through, the AI looks at the shape of the light trail to guess the speed.
The Result: For the most dangerous and common particles, the AI can measure their energy with an error margin of less than 20%. That is good enough to calculate exactly how much radiation dose an astronaut would receive.

Why Is This a Big Deal?

1. Speed: The AI does in milliseconds what used to take 15 minutes. This means the telescope can monitor radiation in real-time, warning astronauts if a dangerous solar storm is hitting right now.
2. Accuracy: It can distinguish between different types of particles much better than current tools. This helps scientists calculate the "biological dose"—essentially, how much damage the radiation will actually do to human DNA.
3. Simplicity: The detector itself is small and lightweight (perfect for space), but the "brain" (the AI) makes it act like a massive, complex laboratory.

The Catch (Limitations)

The authors are honest about the flaws:

• Heavy Particles are Hard: The AI is great at spotting light particles but struggles a bit with the heaviest ones because their trails get messy and overlap.
• Simulation vs. Reality: The AI was trained on computer simulations. In the real world, the fibers might be slightly crooked, or the electronics might act up. The AI needs to be tested on the actual hardware to see if it can handle those "glitches."
• Shielding: The study assumed the detector was floating in open space. In reality, it's inside a spaceship with walls. Those walls break up heavy particles before they reach the detector, which actually helps the AI because it mostly sees the lighter, easier-to-identify particles.

The Bottom Line

This paper presents a smart, fast, and compact way to "see" the invisible radiation in space. By using a neural network (AI) to read the glowing trails of cosmic rays, the RadMap Telescope can tell us exactly what kind of radiation astronauts are facing. This is a crucial step toward keeping future explorers safe on their journeys to the Moon and Mars.

Technical Summary

1. Problem Statement

Future deep-space missions (Moon, Mars) face significant health risks from exposure to cosmic radiation, including cancer, cardiovascular disease, and cognitive impairment. Accurate radiation protection requires detailed knowledge of the space radiation environment, specifically the spectral, temporal, and spatial variations of cosmic-ray nuclei.

Current operational radiation monitors often lack spectroscopic capabilities. They typically measure Linear Energy Transfer (LET) or energy deposition in planar detectors, which introduces large uncertainties in determining the particle's nuclear charge ($Z$) and kinetic energy ($E_{kin}$). Without precise $Z$ and $E_{kin}$, calculating the biologically relevant dose is imprecise. The RadMap Telescope aims to address this by using a compact, low-power scintillating-fiber tracking calorimeter, but interpreting its complex data requires advanced reconstruction methods beyond traditional algorithms (which are too slow for real-time processing).

2. Methodology

The authors developed a neural-network (NN) framework to reconstruct particle properties from the RadMap Telescope's Active Detection Unit (ADU).

• Detector Geometry: The ADU consists of 1024 scintillating plastic fibers ($2 \times 2$ mm cross-section, 80 mm length) arranged in 32 alternating layers. Light is detected by Silicon Photomultipliers (SiPMs). The data is projected onto two 2D images ($yx$ and $yz$ planes), creating "event signatures" that resemble images of particle tracks.
• Challenges: The reconstruction must handle:
  • Straggling: Statistical fluctuations in energy loss.
  • Fragmentation: Nuclei breaking into lighter fragments within the detector.
  • Quenching: Reduced light yield at high energy deposition densities (Bethe-Bloch saturation).
  • Direction Ambiguity: Difficulty distinguishing the travel direction for minimum-ionizing particles (MIPs) where energy deposition is uniform.
• Simulation Data: Training and testing data were generated using Geant4 (version 11.2). The simulation included ionization quenching but deliberately excluded external shielding to assess the detector's intrinsic performance. The dataset covered nuclei from Hydrogen ($Z=1$) to Iron ($Z=26$) with a log-uniform energy distribution to avoid bias toward low energies.
• Neural Network Architecture: The framework uses three sequentially applied networks based on Convolutional Neural Networks (CNNs) with Inception modules (parallel convolutions with varying filter sizes) to handle tracks of different scales and orientations.
  1. Track Reconstruction: A classification network determines the track angles ($\theta, \phi$) by outputting probability distributions over discrete bins.
  2. Charge Determination: A two-step approach. The first network identifies light nuclei ($Z=1$ to $8$), passing "overflow" events to a second network for heavy nuclei ($Z=9$ to $26$). This mitigates the difficulty of distinguishing heavy nuclei with similar energy loss profiles.
  3. Energy Measurement: 26 element-specific networks perform regression to estimate kinetic energy per nucleon. These networks are trained on data including neighboring charges ($Z \pm 1$ or $Z \pm 2$) to robustly handle charge misidentification.

3. Key Contributions

• Real-Time Feasibility: Replaces computationally expensive methods (Bayesian filters, MCMC) that took ~15 minutes per event with a neural network framework capable of real-time processing.
• Two-Stage Charge Identification: Introduces a split-network strategy to overcome the performance drop-off for heavy nuclei, significantly improving accuracy for the full range of cosmic-ray elements.
• Robustness to Charge Confusion: Demonstrates that training energy networks on data with slight charge variations ($|\Delta Z| \le 1$ or $2$) effectively mitigates the impact of charge misidentification on energy resolution.
• Comprehensive Performance Benchmarking: Provides a detailed analysis of angular resolution, charge separation, and energy resolution across different particle types and energies using simulated data.

4. Key Results

The framework was tested on simulated data for protons and iron nuclei across various energy regimes (stopping, monoenergetic, and MIPs).

• Track Reconstruction:

  • Achieved an angular resolution of better than $1.4^\circ$ for both protons and iron nuclei.
  • Mean bias was negligible ($< 0.12^\circ$).
  • Performance was slightly lower for low-energy protons due to fewer traversed fibers but remained within radiation monitoring requirements.

• Charge Determination:

  • Hydrogen ($Z=1$): 99.8% accuracy.
  • Helium ($Z=2$): 99.3% accuracy.
  • Light Nuclei ($Z \le 8$): >95% accuracy for exact identification; >99% if allowing $|\Delta Z| \le 1$.
  • Heavy Nuclei ($Z > 8$): Accuracy drops for exact identification (e.g., ~30-40% for Iron) due to quenching and fragmentation. However, allowing $|\Delta Z| \le 2$ raises accuracy to >83% across the entire range.
  • The framework reliably identifies the "element family" (e.g., iron-like) even if the exact charge is ambiguous.

• Energy Measurement:

  • Hydrogen: Resolution $\sigma_E/E_{kin} < 5\%$ for $E < 100$ MeV and $< 16\%$ for $E < 1$ GeV.
  • Iron: Resolution $\sigma_E/E_{kin} < 10\%$ for $E < 400$ MeV/n and $< 20\%$ for $E < 2$ GeV/n.
  • The framework successfully handles through-going particles by learning the relationship between energy deposition profiles and velocity.

5. Significance

• Biologically Relevant Dose: By accurately determining $Z$ and $E_{kin}$ for the most abundant and biologically damaging cosmic rays (H, He, and heavier nuclei), the RadMap Telescope can calculate the biologically relevant dose with significantly higher precision than current LET-based monitors.
• Mission Safety: The ability to perform spectroscopic measurements in a compact, low-power package is critical for long-duration missions where shielding mass is limited.
• Feasibility of Deep Learning in Space: The study validates that deep learning can effectively solve complex inverse problems in particle physics within the constraints of space instrumentation, provided the training data is carefully curated to handle physical limitations like fragmentation and quenching.
• Future Outlook: While the current study uses an idealized detector model (no external shielding), the authors note that shielding actually improves the relative dominance of light nuclei (H/He) in the dose, which are the elements the framework identifies with the highest accuracy. Future work will focus on integrating realistic detector models (crosstalk, misalignment) and potentially more complex architectures (Transformers, Graph NNs).