Abstract
Precision measurements of neutron properties, like its permanent electric dipole moment, rely on understanding complex experimental setups in detail. We show how the properties of stored and transported ultracold neutron ensembles can be simulated reliably. In a second step, we illustrate how they can be used for simulation-based inference of the parameters associated with underlying physics processes such as neutron capture or beta decay. Our proof of principle for simulation-based inference confronts a longstanding challenge with ultracold neutrons: low measurement statistics coupled with a complex apparatus.
Introduction
Ultracold neutrons (UCNs) are a vital tool for testing the Standard Model, particularly in measuring the neutron's Electric Dipole Moment (EDM). Finding a non-zero EDM would be a major clue in explaining why the universe contains more matter than antimatter. However, UCN experiments are notoriously difficult. Neutrons are "ultracold" when they move slower than a human sprinter, allowing them to be trapped in bottles. But they are rare, and the apparatuses used to store and transport them are incredibly complex.
Currently, researchers often only get a "final count" of neutrons at the end of an experiment. If counts are low, it’s hard to tell if the neutrons were lost because the trap walls were slightly absorbent, the helium buffer gas was "dirty" with impurities like $^3$He, or if the neutrons simply decayed. This "black box" nature of the apparatus makes it hard to optimize experiments and extract fundamental physics constants with high precision.
Methodology
This research introduces a framework for analyzing UCN experiments by combining high-precision GEANT4 simulations with modern Simulation-Based Inference (SBI). The approach consists of two primary components:
- High-Fidelity Simulation: The researchers upgraded GEANT4UCN, a specialized version of the GEANT4 toolkit, to perform ray-tracing simulations of UCNs. Unlike standard particle simulations, this accounts for gravity (which significantly curves UCN paths) and complex quantum-mechanical interactions at the trap walls.
- Simulation-Based Inference (SBI): Because the "likelihood function" (the mathematical probability of seeing a specific result given a set of parameters) is too complex to write down for a UCN trap, the framework uses machine learning. Specifically, a Conditional Flow Matching (CFM) generator—a modern generative model—is employed to learn the relationship between experimental parameters (like loss rates) and the resulting data (time-of-flight spectra).
Datasets and Training
The study utilized a custom dataset generated for training and evaluation, consisting of approximately 9,000 simulations of the SuperSUN UCN source. The researchers varied several key parameters:
- Accumulation times ($t_a$)
- Holding times ($t_h$)
- Wall loss factors ($f$)
- Energy-independent loss rates ($\Gamma'$)
The model was trained on binned 2D-histograms of observed particles, which count neutrons versus time and velocity.
Key Insights and Results
A primary finding of the study was the strong anti-correlation between the wall loss factor ($f$) and energy-independent losses ($\Gamma'$). In typical experiments, a reduction in neutron counts could be attributed to either "bad walls" or gas impurities. The SBI posterior distributions clearly show how these two factors trade off against each other, providing experimentalists with a clearer picture of their systematic uncertainties.
The model also correctly identified physical saturation limits. For instance, after a certain point, increasing the accumulation time does not change the neutron spectrum because the trap reaches a steady-state density. This explains why some parameters are more challenging to infer than others—a useful insight for designing future experimental runs. In benchmark tests, the model successfully recovered experimental timing parameters with high precision and performed a joint inference of critical loss parameters where the "ground truth" values fell squarely within the predicted high-density regions.
Conclusion
While the SBI framework offers a powerful new tool for UCN physics, it remains dependent on the fidelity of the forward simulator. If the GEANT4 model fails to account for a specific physical effect, the inference will be biased. Future work will focus on rigorous coverage studies to ensure statistical calibration and the application of this framework to real experimental data from the SuperSUN and PanEDM collaborations. This ML-enhanced approach represents a significant step toward extracting precise physics results from complex, low-statistics UCN experiments.
Summary updated