Identifying Neutron Sources using Recoil and Time-of-Flight Spectroscopy

This paper introduces a Bayesian protocol that combines full-spectrum template matching with probabilistic evidence evaluation to successfully identify single- and two-source neutron configurations from recoil and time-of-flight spectroscopy data with high statistical significance, even at low event counts.

The Gist

Imagine you are standing in a dark room with your back turned to a group of people throwing different types of balls at a wall. You can't see the throwers, but you can hear the thud of the balls hitting the wall and feel the vibration. Your goal is to figure out: Who is throwing the balls? Are there one or two people? And how hard are they throwing?

This is essentially the problem scientists face when trying to identify neutron sources. Neutrons are tiny, invisible particles that fly out of radioactive materials. They are crucial for everything from finding water on Mars to ensuring nuclear safety, but they are notoriously difficult to "see" or identify because they don't carry an electric charge and often look very similar to each other when they hit a detector.

Here is a simple breakdown of what this paper achieves, using everyday analogies:

1. The Problem: The "Noisy Crowd"

Currently, trying to identify a neutron source is like trying to identify a specific singer in a crowded stadium by listening to a muffled recording.

• The Issue: Most neutron sources emit a continuous stream of particles that look almost identical.
• The Noise: The detector itself blurs the signal (like a bad microphone), and the air or walls the particles pass through change their path (like an echo).
• The Old Way: Scientists usually looked for "clues" like secondary gamma rays (like looking for a specific color of confetti thrown with the balls) or used rough guesses. If the confetti was missing or the room was too dark, they couldn't tell who was throwing the balls.

2. The Solution: The "Bayesian Detective"

The authors (David Breitenmoser and his team at the University of Michigan) created a new method using Bayesian statistics. Think of this as a super-smart detective who doesn't just guess; they calculate the probability of every possible scenario.

• The Template Matching: Imagine you have a library of "sound profiles" for different throwers. One profile is for a "Cf-252" thrower (a specific type of nuclear material), and another is for a "PuBe" thrower.
• The Comparison: When the detector hears a new "thud," the detective compares the sound against the library. Instead of just saying "It sounds like Thrower A," the detective calculates: "There is a 99.9% chance this is Thrower A, and only a 0.01% chance it's Thrower B."
• The "Evidence": The paper introduces a way to mathematically prove how sure they are. They can say, "We are 4-sigma confident," which is a fancy way of saying, "We are so sure that if we were betting, we'd bet our life on it."

3. The Two "Ears": Recoil vs. Time-of-Flight

The team tested their detective method using two different ways of listening to the balls, which they call Recoil and Time-of-Flight (TOF) spectroscopy.

• Recoil Spectroscopy (The "Heavy Hit"): This measures how hard the ball hits the wall and bounces back. It's like feeling the vibration in the floor. It's fast and catches a lot of balls (high event rate).
• Time-of-Flight (The "Speed Trap"): This measures how long it takes the ball to travel between two points. It's like timing a runner. It's more precise about the speed but catches fewer balls because it requires two sensors to "agree" on the same ball.

The Finding: The "Heavy Hit" method (Recoil) was the better detective. It gathered information much faster and with more clarity than the "Speed Trap" (TOF), even when there were very few balls thrown.

4. The Results: Finding the Needle in the Haystack

The team tested their method in a lab with two types of neutron sources (Cf-252 and PuBe).

• Single Source: They could easily identify which one was active.
• Mixed Sources: They could even tell when both were throwing balls at the same time, even if one was throwing much harder than the other.
• Low Counts: They achieved this with as few as 1,000 particles. In the world of nuclear physics, that's a tiny amount of data, making the method incredibly efficient.

5. Why This Matters

This isn't just a lab trick. This new "Bayesian Detective" approach opens a new window for:

• Planetary Science: Helping rovers on Mars or the Moon figure out what the soil is made of just by listening to the neutrons bouncing off the ground.
• Nuclear Security: If a truck passes a checkpoint, this method could instantly tell security if it's carrying harmless medical isotopes or dangerous special nuclear material, even if the material is hidden behind lead shielding.
• Emergency Response: If there's a nuclear accident, first responders can quickly identify the source of the radiation to know how to protect themselves.

The Bottom Line

The authors have built a mathematical "super-sense" that allows us to look at a messy, blurry cloud of invisible particles and say with extreme confidence: "That's not just random noise; that's a specific type of nuclear source, and we know exactly what it is." They turned a blurry, confusing signal into a clear, high-definition picture of the source.

Technical Summary

1. Problem Statement

Neutron source identification is critical for applications ranging from planetary science (e.g., determining surface composition) to nuclear security (e.g., distinguishing special nuclear material). However, direct discrimination of neutron sources from measured energy spectra has remained fundamentally elusive due to several challenges:

• Spectral Similarity: Most neutron sources (e.g., spontaneous fission vs. $(\alpha,n)$ reactions) emit continuous energy spectra with strong similarities.
• Ill-Conditioned Inverse Problems: Detector response broadening and modulation by intervening materials (shielding, scattering) obscure the underlying source characteristics.
• Limitations of Current Methods: Existing approaches rely on indirect signatures (secondary gamma/alpha emissions) or qualitative integral metrics. These methods often lack quantitative statistical significance, struggle with multi-source scenarios, and fail when secondary emissions are attenuated or absent.

2. Methodology

The authors propose a Bayesian protocol that directly infers source ensembles from measured neutron spectra by combining full-spectrum template matching with probabilistic evidence evaluation.

A. Bayesian Framework

• Model Comparison: The problem is formulated as a Bayesian model comparison. The goal is to determine which source model $M$ (e.g., single source vs. mixed sources) best explains the observed data $D$.
• Model Evidence: The support for a model is quantified by the marginal likelihood (Bayesian evidence), $Z = \int L(\theta; D, M) p(\theta | M) d\theta$, calculated using nested sampling (via the dynesty package).
• Likelihood Function: The authors employ a probabilistic negative binomial model to account for overdispersion beyond intrinsic Poisson statistics. The likelihood $L$ is parameterized by the neutron emission rates $\xi$ and a dispersion parameter $\alpha_{NB}$.
• Priors: Weakly informative priors are used for emission rates (truncated normal) and dispersion (exponential) to avoid restrictive assumptions.

B. Spectral Templates

• Data-Driven Generation: Instead of relying solely on theoretical simulations, the authors use Generalized Additive Models (GAMs) trained on experimental single-source data to generate spectral templates ($\psi_s$).
• Forward Mapping: The expected spectral response is modeled as a linear superposition of these templates: $M(\theta) = \sum \xi_s \psi_s$.
• Robustness: Templates were generated under varying conditions (different azimuthal angles, lead shielding) to ensure generalization to the identification experiments.

C. Experimental Setup

• Detector: A 12-bar array of Organic Glass Scintillator (OGS) coupled to dual-ended Silicon Photomultipliers (SiPMs).
• Modalities: The system simultaneously performs:
  1. Recoil Spectroscopy: Analyzing single-bar events to measure light output (proton recoil).
  2. Time-of-Flight (TOF) Spectroscopy: Analyzing coincidence events between bars to reconstruct incident neutron energy.
• Sources: Experiments utilized $^{252}\text{Cf}$ (spontaneous fission) and $^{239}\text{Pu-Be}$ ($( \alpha, n)$) sources in single and mixed configurations.

3. Key Contributions

1. Direct Quantitative Identification: The first demonstration of a method that directly infers source ensembles from neutron spectra with high statistical confidence, bypassing the need for secondary emission proxies.
2. Bayesian Evidence Evaluation: Introduction of a rigorous statistical framework (Bayes factors and posterior odds) to discriminate between single- and multi-source models, providing a quantitative measure of significance (e.g., $>4\sigma$).
3. Event Scaling Analysis: A comprehensive study on how identification fidelity scales with the number of detected events, revealing that reliable identification is possible with as few as $\sim 10^3$ events.
4. Modal Comparison: A systematic comparison between Recoil and TOF spectroscopy, demonstrating that Recoil spectroscopy offers superior information gain per event due to lower statistical dispersion.

4. Key Results

• High-Confidence Discrimination: The protocol successfully recovered single- and two-source configurations with decisive statistical confidence ($\Delta \log Z > 8.0$, corresponding to $>4\sigma$) even at low event counts ($\sim 10^3$).
• Robustness to Imbalance: The method remains robust even when source emission rates are highly imbalanced (ratios up to $10^2$). While higher event counts ($\sim 10^6$) are required for extreme imbalances, the correct source set is still recovered with decisive confidence.
• Recoil vs. TOF Performance:
  • Recoil Spectroscopy: Achieved higher information gain and faster convergence to the true model. This is attributed to a significantly lower overdispersion parameter ($\alpha_{NB}$) compared to TOF, meaning each recoil event carries more statistical weight.
  • TOF Spectroscopy: While effective, it required more events to reach the same level of confidence due to higher statistical dispersion inherent in the coincidence reconstruction process.
• Prior Independence: The model selection outcome was found to be dictated by the observed data rather than the specific choice of model priors (non-committal vs. weakly informative), confirming the robustness of the evidence-based approach.

5. Significance and Future Outlook

• Operational Impact: This work opens a new observational window for nuclear forensics, emergency response, and non-proliferation, enabling robust source attribution without relying on secondary gamma signatures which may be shielded.
• Scientific Application: The methodology is applicable to planetary science (e.g., analyzing neutron spectra from Mars or the Moon) to infer surface composition and moisture content.
• Scalability: The authors propose that this framework can be extended to complex, realistic scenarios by integrating high-fidelity neutron transport simulations (e.g., MCNP, Geant4) to generate physics-informed template banks. This would allow for the joint inference of source ensembles and modulation processes (shielding, geometry) in uncontrolled environments.

In conclusion, the paper establishes a rigorous, quantitative foundation for neutron source identification, proving that full-spectrum Bayesian analysis can overcome the historical limitations of spectral similarity and low-count statistics.