Optimisation of importance sampling implementation in rare-event shielding simulations using Geant4

arXiv:2609.03822 · physics.ins-det, hep-ex, physics.comp-ph · Submitted 2026-09-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Optimisation of importance sampling implementation in rare-event shielding simulations using Geant4".

Paper summary: Kai: So to recap what we’ve seen, this paper focuses on optimizing how we use importance splitting and Russian roulette for rare-event shielding simulations, specifically looking at how to handle those computationally expensive background scenarios.

Mira: The thesis here is that by filling the shielding geometry with a whole number of equal-width importance layers and varying the thickness of those layers for both gamma rays and neutrons, you can balance statistical uncertainty against simulation execution time effectively.

Lev: They are developing this to give guidance for the optimal application of biasing in future low background experiments, inspired by things like the DarkSPHERE experiment.

Kai: The key claim is that the optimal layer thickness, when expressed as a multiple of the primary particle’s mean free path, provides a physics-based prediction that works for any shielding design.

Mira: They tested this across different materials and thicknesses, even using lead instead of water shielding to check if the concept held up.

Lev: The result they highlight is an O(twenty) gain in computational efficiency over the unbiased case when using this optimal configuration <ref:2609.03822#pg2>.

Kai: It’s about providing a simplified basis for doing this balancing act, which is usually needed because the simulated geometry can get quite complex.

Mira: So, it simplifies the process by giving researchers a way to get guidance on how many layers and what thickness they should aim for based on physical principles.

Lev: It’s about moving away from just blindly implementing a biasing scheme and toward using deterministic calculations to guide it, which is necessary when you have these low background constraints.

Kai: The paper sets up the context by describing the importance splitting and Russian roulette method first, which is the technique they are optimizing.

Mira: They then detail their methodology for the optimization study, outlining exactly how they calculate relative uncertainty and execution time for different layer configurations.

Lev: And they show how they implemented this within a Geant4 simulation using a relatively simple geometry to decouple geometric effects from the main optimization studies.

Conclusion: Kai: So looking at the full paper, "Optimisation of importance sampling implementation in rare-event shielding simulations using Geant4", it really lays out how to tackle that trade-off between statistical limitations and computational cost in these simulations.

Mira: The authors show that by optimizing the layer thickness based on the primary particle’s mean free path, you get a physics-based prediction applicable to any shielding design they might need.

Lev: This guidance is generalizable because it works across different energies for a given particle type, and it holds up when tested against different materials like lead or water shielding.

Kai: The implication is that researchers can use this information to get a much faster simulation setup without having to run those overly complex unbiased cases all the time.

Mira: It’s about giving them a way to reduce their computation related carbon footprint by achieving significant efficiency gains, as they did with the O(twenty) speed increase <ref:2609.03822#pg2>.

Lev: Ultimately, it gives them a predictable way to set up these biasing schemes that substantially outperforms the under-, over-, and unbiased cases in terms of computational performance.

P. Knightsa, L. J. Milligana, K. Nikolopoulosa

School of Physics and Astronomy, University of Birmingham · Institute for Experimental Physics, University of Hamburg

physics.ins-det, hep-ex, physics.comp-ph

Submitted: 2026-09-03

Updated: 2026-11-17

Comments: 13 pages, 8 figures, 2 tables

Journal ref: Astropart. Phys. 185 (2027) 103316

DOI: 10.1016/j.astropartphys.2026.103316

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 85/100

The gist: Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without…

Key concepts

Importance Splitting
This biasing technique splits simulated particles into new tracks at specific geometry boundaries based on their importance values. This creates new particles with adjusted weights, increasing the population of interest near the detector while preserving statistical information for later analysis.
Russian Roulette
This technique probabilistically kills simulated tracks that cross boundaries into regions of lower importance. Tracks are killed based on a probability calculated from their post- and pre-step importance values, effectively reducing computational load by eliminating irrelevant paths.
Relative Uncertainty (R)
Relative uncertainty is calculated as the ratio of the standard deviation ($\sigma P$) to the mean probability ($\mu P$) of a primary particle interacting with the detector. Minimizing this value is crucial for rare-event searches, indicating a better statistical representation of the event probability.
Layer Thickness Generalization
The optimal layer thickness can be expressed as a multiple of the primary particle's mean free path in water shielding. This provides a physics-based prediction that works across different energies and particle types, making it applicable to any shielding design.

Terminology

Summary

Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without event biasing. p. 0

Where the paper says it

Milligana,∗ , K. Nikolopoulosa,b aSchool of Physics and Astronomy, University of Birmingham, B15 2TT, Birmingham, United Kingdom b Institute for Experimental Physics, University of Hamburg, 22761, Hamburg, Germany Abstract Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without event biasing.

The method presented optimises the widely used importance splitting and Russian roulette technique for rare-event searches by filling the shielding geometry with a whole number of equal-width importance layers. p. 0
Where the paper says it

Nikolopoulosa,b aSchool of Physics and Astronomy, University of Birmingham, B15 2TT, Birmingham, United Kingdom b Institute for Experimental Physics, University of Hamburg, 22761, Hamburg, Germany Abstract Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without event biasing. We present a method for optimising the widely used importance splitting and Russian roulette technique for rare-event searches, in which a whole number of equal-width importance layers fill the shielding geometry.

This optimisation balances relative uncertainty against execution time by varying the layer thickness for both γ-rays and neutrons across different shielding materials and thicknesses. p. 0
Where the paper says it

We present a method for optimising the widely used importance splitting and Russian roulette technique for rare-event searches, in which a whole number of equal-width importance layers fill the shielding geometry. Balancing relative uncertainty against execution time, we optimise the layer thickness for both γ-rays and neutrons, and across different shielding materials and thicknesses.

The optimal thickness generalises across energies for a given particle type and yields a physics-based prediction applicable to any shielding design. p. 0
Where the paper says it

Balancing relative uncertainty against execution time, we optimise the layer thickness for both γ-rays and neutrons, and across different shielding materials and thicknesses. Expressed as a multiple of the primary particle’s mean free path, the optimal thickness generalises across energies for a given particle type and yields a physics-based prediction applicable to any shielding design.

The optimal configuration gives an O(20) gain in computational efficiency over the unbiased case. p. 0
Where the paper says it

The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time. The optimal configuration gives an O(20) gain in computational efficiency over the unbiased case.

p. 0
Where the paper says it

The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time. The optimal configuration gives an O(20) gain in computational efficiency over the unbiased case.

Biasing Technique Description

The importance splitting and Russian roulette technique is a widely used biasing technique that splits simulated particles, referred to as tracks, at user-defined geometry boundaries based on importance values. This process generates new simulated particles with altered weights and probabilistically kills tracks crossing in the opposite direction. p. 1

Where the paper says it

Applied to a single particle type of interest, this technique ‘splits’ simulated particles, referred to in the following as tracks. That is, it generates new simulated particles with altered weights, at userdefined geometry boundaries, and probabilistically kills tracks crossing in the opposite direction.

The number of new tracks generated in the splitting Nsplit and their weights w are given by specific expressions involving post- and pre-step importance values. p. 1
Where the paper says it

The number of tracks that split off of a single track is given by the ratio of the post- and pre-step importance values. The resulting ‘split’ tracks are then weighted by 1/Ipost where Ipost is the importance value of the new layer.

If a track moves into a layer with a lower importance value it is assigned a probability of being killed according to the expression pkill = 1 − Ipost / Ipre. p. 1
Where the paper says it

The number of new tracks generated in the splitting Nsplit and their weights w are given by Nsplit = Ipost Ipre w = 1 Ipost . If a track moves into a layer with a lower importance value it is assigned a probability of being killed according to the expression pkill = 1 − Ipost Ipre , and its weight assigned accordingly if it survives.

This technique increases the population of particles of interest near the sensitive detector and conserves parameters of interest by accounting for each track’s statistical weight in later analysis. p. 1
Where the paper says it

If a track moves into a layer with a lower importance value it is assigned a probability of being killed according to the expression pkill = 1 − Ipost Ipre , and its weight assigned accordingly if it survives. This increases the population of particles of interest near the sensitive detector, and hence the number of tracks interacting with it, while parameters of interest are conserved by accounting for each track’s statistical weight in later analysis.

Performance Metrics and Optimization

The relative uncertainty (R) is calculated as R = σP / µP, where σP is the standard deviation and µP is the mean of the probability of a primary particle interacting with the SPC. The execution time for each group of combined simulations is taken as the group average execution time. p. 3

Where the paper says it

By taking the mean (µP) and standard deviation (σP ) of each collection of probabilities, the relative uncertainty (R) for that group of combined simulations is: R = σP µP . The execution time for this group is, similarly, taken as the groups average execution time (µt).

The study compares these metrics for neutrons and photons at two different energies each. p. 4
Where the paper says it

to execute the core Geant4 simulation, and is what is referred to as ‘execution time’ in this study. To understand the biasing performance over a range of scenarios, we perform a study on both neutrons and photons, at two different energies for each particle.

The performance trends show that R improves with increasing amounts of layers until a broad minima is reached, after which it starts to increase due to increased computational resources needed. p. 4
Where the paper says it

The performance trends derived from these fits are shown in Fig. A pattern is immediately noticeable, in which R improves with increasing amounts of layers until a broad minima is reached, 10 2 10 3 10 4 Execution Time (s) 10 −2 10 −1 R ela tiv e U n c e r t ain t y ( σ p r o b / μ p r o b ) 5 Layers 12 Layers 18 Layers 22 Layers Figure 3: Relative Uncertainty versus execution time for different number of layers, using 2750 keV γ-rays as the primary particle.

Energy and Particle Dependence

The optimal layer thickness can be expressed as a multiple of the primary particle’s mean free path, generalising across energies for a given particle type and yielding a physics-based prediction. For γ-rays, the turning points occur at layers 15 and 18 with corresponding ranges of [9, 21] and [14, 23] for the two simulated energies. p. 5

Where the paper says it

14 cm for a 1461 keV γ-ray [31]. The turning points can be observed to occur at layers 15 and layers 18 with corresponding ranges of [9, 21] and [14, 23] (after rounding the extracted minima and ranges to the nearest integer), for the simulations of 2750 keV and 1461 keV respectively.

Similarly, for neutrons at 10 MeV and 5 MeV, the turning points are found at layer counts of 19 and 22 respectively with corresponding ranges of [16, 22] and [20, 24]. p. 5
Where the paper says it

In addition, noticing that the minima tend to be broad, two layer counts either side of the turning point corresponding to a 50% increase in the optimal R are found so to quantify the range about the minimum. To predict these trends in terms of the physics of the primary particle, we can express the turning points/ranges as a layer thickness that is equal to a multiple of the particle’s mean free path.

Generalisation to Different Scenarios

The results can be generalised to different shielding materials and thicknesses by testing lead as an alternative material and varying the water shielding thickness. The optimal layer thickness when expressed in terms of λγ/nH2O is compatible across different particle types and energies. p. 7

Where the paper says it

We can see that an optimal layer thickness can be recommended for each particle type, independent of energy, when expressed as a function of the mean free path of the primary particle in the water shielding at a given energy, assuming equally thick layers. We can see that this layer thickness appears to be constant across different energies for the same particle, but is not consistent across different types of particles.

The most optimal scenario achieved a R of 0. p. 9
Where the paper says it

45 R ela tiv e U n c e r t ain t y ( σ p r o b / μ p r o b ) (b) Figure 8: The R performance for the (a) lead and (b) water shielding simulation using a thickness of 120 cm. that the most optimal scenario achieved a R of 0.

074 in 300. p. 9
Where the paper says it

that the most optimal scenario achieved a R of 0. 074 in 300.

p. 9
Where the paper says it

that the most optimal scenario achieved a R of 0. 074 in 300.

p. 9
Where the paper says it

that the most optimal scenario achieved a R of 0. 074 in 300.

00 s, which is O(20) times faster than the unbiased case. p. 9
Where the paper says it

The execution times for the other scenarios are summarised in Table 2, along with the factor increase in execution time compared to the optimal scenario. Notably, compared to the unbiased case, there is a factor 22.

This efficiency gain translates directly into an improved simulation-related carbon footprint. p. 9
Where the paper says it

These are significant improvements in efficiency, and demonstrate the utility of this study for those concerned with minimising their usage of computational resources. This also has clear benefits as a method of decreasing one’s computation-related carbon-footprint, which in certain areas of particle physics is a key component of a researcher’s annual professional CO2e output [35].

Conclusion

The optimisation of importance splitting and Russian roulette provides a method to relieve both statistical limitations and computational cost in rare-event shielding simulations. The resulting optimal configuration substantially outperforms the under-, over-, and unbiased cases, achieving an O(20) gain in computational efficiency. p. 0

Where the paper says it

The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time. The optimal configuration gives an O(20) gain in computational efficiency over the unbiased case.

This guidance is readily generalisable for many rare-event searches with similar shielding geometries. p. 9
Where the paper says it

Event biasing schemes in Geant4 can relieve both, and we have optimised one of the most common of these, importance splitting and Russian roulette, for the rare-event search field, a setting in which it had not previously been studied in detail. The optimisation was performed within a version of the DarkSPHERE shielding geometry, filled with equalwidth biasing layers; as this design is common to many rare-event searches, the results are readily generalisable.

The optimal layer thickness becomes predictable from physics and generalises within particle type and across shielding thicknesses and materials. p. 9
Where the paper says it

Using the relative uncertainty (R) and execution time to understand how R varies with layer count, we determined the optimal layer thickness for both γ-ray and neutron simulations, at two energies each. Expressed in terms of the primary particle’s mean free path, this thickness becomes predictable from physics and generalises within particle type and across shielding thicknesses and materials.

The trend is parabolic, demonstrating that over-biasing is possible. p. 9
Where the paper says it

Expressed in terms of the primary particle’s mean free path, this thickness becomes predictable from physics and generalises within particle type and across shielding thicknesses and materials. In all cases the trend is parabolic, with R degrading at high layer counts for fixed execution time, demonstrating that over-biasing is possible.

The optimal layer thickness when expressed in terms of λγ/nH2O is equal across neutron energies. p. 6
Where the paper says it

Table 1 summarises these results when expressed as a layer thicknesses in terms of λ n H2O. Both thicknesses are compatible, and so we can conclude that the optimal layer thickness when expressed in terms of λ n H2O, are equal across neutron energies.

The results can also be extended to other rare-event shielding simulations. p. 7
Where the paper says it

predictive behaviour can also be extended for other rare-event shielding simulations. Summary of Observed Performance Table 1 summarises the results discussed above.

The optimal configuration substantially outperforms the under-, over-, and unbiased cases. p. 0
Where the paper says it

The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time. The optimal configuration gives an O(20) gain in computational efficiency over the unbiased case.

--- Page 1 ---

Optimisation of importance sampling implementation in rare-event shielding simulations using Geant4 Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without event biasing. We present a method for optimising the widely used importance splitting and Russian roulette technique for rare-event searches, in which a whole number of equal-width importance layers fill the shielding geometry. p. 0

Where the paper says it

Nikolopoulosa,b aSchool of Physics and Astronomy, University of Birmingham, B15 2TT, Birmingham, United Kingdom b Institute for Experimental Physics, University of Hamburg, 22761, Hamburg, Germany Abstract Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without event biasing. We present a method for optimising the widely used importance splitting and Russian roulette technique for rare-event searches, in which a whole number of equal-width importance layers fill the shielding geometry.

Balancing relative uncertainty against execution time, we optimise the layer thickness for both γ-rays and neutrons, and across different shielding materials and thicknesses. p. 0
Where the paper says it

We present a method for optimising the widely used importance splitting and Russian roulette technique for rare-event searches, in which a whole number of equal-width importance layers fill the shielding geometry. Balancing relative uncertainty against execution time, we optimise the layer thickness for both γ-rays and neutrons, and across different shielding materials and thicknesses.

Expressed as a multiple of the primary particle’s mean free path, the optimal thickness generalises across energies for a given particle type and yields a physics-based prediction applicable to any shielding design. p. 0
Where the paper says it

Balancing relative uncertainty against execution time, we optimise the layer thickness for both γ-rays and neutrons, and across different shielding materials and thicknesses. Expressed as a multiple of the primary particle’s mean free path, the optimal thickness generalises across energies for a given particle type and yields a physics-based prediction applicable to any shielding design.

The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time. p. 0
Where the paper says it

Expressed as a multiple of the primary particle’s mean free path, the optimal thickness generalises across energies for a given particle type and yields a physics-based prediction applicable to any shielding design. The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time.

The optimal configuration gives an O(20) gain in computational efficiency over the unbiased case.

--- Page 4 ---

The key parameters governing performance are the layer number and thickness. We link them directly by filling the shielding with a given number, M, of equal-width layers, tshield = M × tlayer. p. 3

Where the paper says it

Optimising Importance Splitting: Computational Effort Versus Relative Uncertainty The key parameters governing performance are the layer number and thickness. We link them directly by filling the shielding with a given number, M, of equal-width layers, tshield = M × tlayer, so that the only parameter varied is the number of layers.

The relative uncertainty and simulation execution time are compared to assess the performance for different values for M. p. 3
Where the paper says it

We link them directly by filling the shielding with a given number, M, of equal-width layers, tshield = M × tlayer, so that the only parameter varied is the number of layers. The relative uncertainty and simulation execution time are compared to assess the performance for different values for M.

Implementing variable layer thicknesses is likely to be non-optimal for the case of mono-energetic simulations, such as those used in this and many other shielding studies, as each primary particle is equally as likely to be stopped anywhere in the shielding. p. 3
Where the paper says it

The relative uncertainty and simulation execution time are compared to assess the performance for different values for M. Implementing variable layer thicknesses is likely to be non-optimal for the case of mono-energetic simulations, such as those used in this and many other shielding studies, as each primary particle is equally as likely to be stopped anywhere in the shielding.

--- Page 5 ---

By taking the mean (µP) and standard deviation (σP) of each collection of probabilities, the relative uncertainty (R) for that group of combined simulations is: R = σP / µP. To execute the core Geant4 simulation, and is what is referred to as ‘execution time’ in this study. p. 4

Where the paper says it

eIoni Compton Figure 2: Visualisations of the track genealogy for two example events: one where a primary neutron results in a hit in the SPC due to a photon, and one where there are two unique hits. to execute the core Geant4 simulation, and is what is referred to as ‘execution time’ in this study.

To understand the biasing performance over a range of scenarios, we perform a study on both neutrons and photons, at two different energies for each particle. p. 4
Where the paper says it

to execute the core Geant4 simulation, and is what is referred to as ‘execution time’ in this study. To understand the biasing performance over a range of scenarios, we perform a study on both neutrons and photons, at two different energies for each particle.

--- Page 7 ---

The turning points can be observed to occur at layers 15 and layers 18 with corresponding ranges of [9, 21] and [14, 23] (after rounding the extracted minima and ranges to the nearest integer), for the simulations of 2750 keV and 1461 keV respectively. It is also worth noting that, as the fixed execution time increase, all values of R shift downwards and the overall trend appears to flatten (indicative by the flattening of the parabola). p. 5

Where the paper says it

The turning points can be observed to occur at layers 15 and layers 18 with corresponding ranges of [9, 21] and [14, 23] (after rounding the extracted minima and ranges to the nearest integer), for the simulations of 2750 keV and 1461 keV respectively. It is also worth noting that, as the fixed execution time increase, all values of R shift downwards and the overall trend appears to flatten (indicative by the flattening of the parabola).

--- Page 8 ---

Table 1 summarises these results when expressed as a layer thicknesses in terms of λγ/nH2O. Both energies are seen to give compatible results. p. 5

Where the paper says it

Table 1 summarises the results for γ-rays. Both energies are seen to give compatible results.

It can be concluded that an amount of layers with equal thicknesses approximate to the above minima, will yield the optimal biasing result for these shielding simulations. p. 5
Where the paper says it

Both energies are seen to give compatible results. It can be concluded that an amount of layers with equal thicknesses approximate to the above minima, will yield the optimal biasing result for these shielding simulations.

--- Page 9 ---

We can see that an optimal layer thickness can be recommended for each particle type, independent of energy, when expressed as a function of the mean free path of the primary particle in the water shielding at a given energy. We can see that this layer thickness appears to be constant across different energies for the same particle, but is not consistent across different types of particles. p. 7

Where the paper says it

We can see that an optimal layer thickness can be recommended for each particle type, independent of energy, when expressed as a function of the mean free path of the primary particle in the water shielding at a given energy, assuming equally thick layers. We can see that this layer thickness appears to be constant across different energies for the same particle, but is not consistent across different types of particles.

--- Page 10 ---

We can see that this layer thickness appears to be constant across different energies for the same particle, but is not consistent across different types of particles.

--- Page 12 ---

The optimal number of layers (exact values shown in parentheses), their thickness, and the range of the minima, for each particle type and energy. p. 8

Where the paper says it

Statistics are lower for the 5 MeV case, owing to a higher λ n H2O. Table 1: The optimal number of layers (exact values shown in parentheses), their thickness, andthe range of the minima, for each particle type and energy.

--- Page 11 ---

As rare-event searches push towards ultra-low-backgrounds and energy thresholds, their background and shielding simulations become increasingly limited by statistics and computational cost. Event biasing schemes in Geant4 can relieve both, and we have optimised one of the most common of these, importance splitting and Russian roulette, for the rare-event search field, a setting in which it had not previously been studied in detail. p. 9

Where the paper says it

Conclusion As rare-event searches push towards ultra-low backgrounds and energy thresholds, their background and shielding simulations become increasingly limited by statistics and computational cost. Event biasing schemes in Geant4 can relieve both, and we have optimised one of the most common of these, importance splitting and Russian roulette, for the rare-event search field, a setting in which it had not previously been studied in detail.

--- Page 10 ---

We find that the most optimal scenario achieved a R of 0.074 in 300.00 s. Notably, compared to the unbiased case, there is a factor 22. p. 9

Where the paper says it

The execution times for the other scenarios are summarised in Table 2, along with the factor increase in execution time compared to the optimal scenario. Notably, compared to the unbiased case, there is a factor 22.

p. 9
Where the paper says it

The execution times for the other scenarios are summarised in Table 2, along with the factor increase in execution time compared to the optimal scenario. Notably, compared to the unbiased case, there is a factor 22.

8 increase in computational efficiency. p. 9
Where the paper says it

Notably, compared to the unbiased case, there is a factor 22. 8 increase in computational efficiency.

p. 9
Where the paper says it

Notably, compared to the unbiased case, there is a factor 22. 8 increase in computational efficiency.

These are significant improvements in efficiency, and demonstrate the utility of this study for those concerned with minimising their usage of computational resources. p. 9
Where the paper says it

8 increase in computational efficiency. These are significant improvements in efficiency, and demonstrate the utility of this study for those concerned with minimising their usage of computational resources.

p. 9
Where the paper says it

8 increase in computational efficiency. These are significant improvements in efficiency, and demonstrate the utility of this study for those concerned with minimising their usage of computational resources.

--- Page 10 ---

This also has clear benefits as a method of decreasing one’s computation-related carbon-footprint, which in certain areas of particle physics is a key component of a researcher’s annual professional CO2e output.

--- Page 10 ---

The optimal layer thickness can be recommended for each particle type, independent of energy, when expressed as a function of the mean free path of the primary particle in the water shielding at a given energy.

--- Page 8 ---

We can see that this layer thickness appears to be constant across different energies for the same particle, but is not consistent across different types of particles. The optimal number of layers (exact values shown in parentheses), their thickness, and the range of the minima, for each particle type and energy.

--- Page 5 ---

The optimal layer thickness can be recommended for each particle type, independent of energy, when expressed as a function of the mean free path of the primary particle in the water shielding at a given energy.

--- Page 11 ---

Event biasing schemes in Geant4 can relieve both, and we have optimised one of the most common of these, importance splitting and Russian roulette, for the rare-event search field, a setting in which it had not previously been studied in detail.

--- Page 11 ---

The optimal layer thickness can be recommended for each particle type, independent of energy, when expressed as a function of the mean free path of the primary particle in the water shielding at a given energy.

--- Page 10 ---

We find that the most optimal scenario achieved a R of 0.074 in 300.0 s. Notably, compared to the unbiased case, there is a factor 22.8 increase in computational efficiency. These are significant improvements in efficiency, and demonstrate the utility of this study for those concerned with minimising their usage of computational resources.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Optimised importance splitting implementation for rare-event searches: The system can dynamically adjust layer thickness for both gamma-rays and neutrons across different shielding materials to achieve an O(20) gain in computational efficiency over unbiased cases.

  2. Physics-based prediction for optimal layer thickness: The AI can predict the optimal tlayer by expressing it as a multiple of the primary particle’s mean free path, yielding a physics-based prediction applicable to any shielding design.

  3. Over-biasing detection: The system can identify when over-biasing is possible by observing that relative uncertainty worsening at high layer counts for fixed execution time, allowing for intelligent control over biasing parameters.

  4. Energy and material generalization: The system can generalize results to different shielding materials and thicknesses by comparing performance across scenarios, such as replacing water with lead while scaling thickness using the ratio of mean free paths in lead and water.

Abstract

Rare-event search experiments demand ever-lower backgrounds, and the resulting radio-pure materials and suppressive shielding make related background simulations computationally expensive without event biasing. We present a method for optimising the widely used importance splitting and Russian roulette technique for rare-event searches, in which a whole number of equal-width importance layers fill the shielding geometry. Balancing relative uncertainty against execution time, we optimise the layer thickness for both γ-rays and neutrons, and across different shielding materials and thicknesses. Expressed as a multiple of the primary particle's mean free path, the optimal thickness generalises across energies for a given particle type and yields a physics-based prediction applicable to any shielding design. The same trend reveals that over-biasing is possible, with relative uncertainty worsening at high layer counts for fixed execution time. The optimal configuration gives an O (20) gain in computational efficiency over the unbiased case.

Sources

Related papers