Source-Decomposed Bayesian Optimization of Layered Radiation Shields

Detectors that search for very rare events, such as dark matter or neutrinoless double-beta decay, sit inside nested shells of material. Outer polyethylene slows neutrons, lead stops gamma rays, and inner copper absorbs the X-rays the lead itself emits. Choosing the thickness of each shell today relies on physics intuition checked by Monte Carlo simulation. Each simulation tracks particles one at a time through the geometry and is very expensive, because almost none of the simulated particles reach the detector. We propose to automate this design with Bayesian optimization, a method that fits a probabilistic model to a few costly trials and uses it to pick the next trial. The twist is that we do not model the total background as one noisy black box. We model each background source separately, because each is simulated separately anyway: external gammas, environmental neutrons, cosmic-muon-induced particles, and radioactivity inside the shield materials. The optimizer then decides three things: which source to simulate next, at which design, and with how many simulated particles.

The two-sentence version is this. The total background is a known weighted sum of per-source rates, and each rate depends mostly on a few layers. An optimizer that models the sources as separate tasks and spends simulation effort where the total is most uncertain should reach a good shield design with far fewer processor-hours than one that treats the total as a single objective. A second benefit comes from the same structure. When the source strengths change, for example when a new material assay arrives or the detector moves from sea level to an underground site, the design can be re-optimized without new simulations.

Core Claim

Write $\lambda$ for the shield design (layer thicknesses, discrete material choices per layer, and veto timing and threshold settings). Write $b_k(\lambda)$ for the rate of events that deposit energy in the signal window per simulated primary particle from source $k$, and $w_k$ for the physical strength of source $k$, taken from assay or site flux. The quantity to minimize, subject to a mass or footprint budget, is

$$B(\lambda) = \sum_{k=1}^{K} w_k \, b_k(\lambda).$$

A simulation of source $k$ at design $\lambda$ with $n$ primaries returns a count $c \sim \text{Poisson}(n\, b_k(\lambda))$. Its cost grows roughly linearly in $n$, and the count is often zero. We place an independent Gaussian-process prior on each $\log b_k$. A Gaussian process is a distribution over smooth functions whose uncertainty shrinks near observed points. We fit it with a Poisson likelihood rather than Gaussian noise, and we choose the next triple $(\lambda, k, n)$ to maximize information gained about the minimizer of $B$ per CPU-second.

The claim has three parts, each testable. The first concerns sample efficiency. On a fixed simulation budget, the method finds lower-background designs than optimizers that see only the total. The second concerns cost. With separate per-source surrogates, fitting costs about $\sum_k N_k^3$ rather than $(\sum_k N_k)^3$, where $N_k$ is the number of runs for source $k$. The third concerns robustness. Because the known sum structure encodes how the sources combine, the model does not need the sources to be positively correlated. In shields they often are not: more lead lowers the gamma rate but adds material that muons break apart into neutrons.

Relation to Prior Work

The method builds directly on two papers. Practical Bayesian Optimization of Machine Learning Algorithms supplies the Matérn 5/2 kernel, the averaging of kernel settings by Markov chain Monte Carlo, cost-aware acquisition ("expected improvement per second"), and parallel runs using hallucinated outcomes for pending jobs. It treats the objective as a single function with Gaussian noise, which is badly matched to sparse Poisson counts. Multi-Task Bayesian Optimization is the closest existing work. It models several related objectives jointly with a learned task-correlation matrix and chooses cheap auxiliary tasks by information gain per unit cost. On Online LDA it reached the optimum in 6 days versus 10 for single-task optimization, and on logistic regression in 40 versus 100 minutes.

Our proposal differs from multi-task optimization in three ways. Its tasks are cheaper stand-ins for one target, while ours are additive parts of the target with known weights. Its correlation matrix must be learned and assumes useful positive correlation, while ours needs none. It also acknowledges cubic cost in the total number of observations across tasks, which our factorized model avoids. Neither paper touches physical simulation.

The shielding literature shows that the problem exists and has not been solved. A Low Background Gamma Ray Spectrometer with Anticosmic Shielding layers lead, tin, and copper and adds a scintillator veto against cosmic rays. Background Model for the Majorana Demonstrator builds a per-source background budget from Geant4 simulations combined with material assays, projecting 4.1 counts per region-of-interest-tonne-year. In both, the layers and cuts are set by hand. The Majorana source-by-source budget is exactly the decomposition we exploit. It shows that practitioners already simulate sources separately and combine them with assay weights. Event reconstruction in a liquid xenon Time Projection Chamber with an optically-open field cage shows that veto-like settings, such as using the outer xenon "skin" as an active veto, change background discrimination by measurable amounts. That motivates including veto parameters in $\lambda$.

SOBRA - Shielding Optimization for BRAchytherapy proves that some discrete shield-mask scheduling problems are NP-hard. That result limits our claims rather than supplying a method. Our design space also includes discrete choices, such as material order, but we make no exactness claim. We treat material order as a categorical input and test whether the optimizer rediscovers descending-atomic-number grading.

Multi-task learning enters as a lens on failure. RotoGrad: Gradient Homogenization in Multi-Task Learning shows that shared models suffer when tasks pull in conflicting directions, and that aligning tasks adds computation. Our sources are conflicting tasks of exactly this kind. We avoid forced sharing through the known sum rather than learned alignment, and we test a shared neural surrogate with per-source output heads as the scalable alternative.

Research Program

We will build an optimization loop around Geant4, the standard Monte Carlo toolkit for particle transport. It has three components. The first is a surrogate for each source: a Gaussian process on $\log b_k$ with a Poisson likelihood fit by Laplace approximation. It has a mixed kernel, Matérn over thicknesses and a categorical kernel over material choices, and a mean function taken from simple exponential-attenuation estimates, so the model starts from textbook physics. The second is an acquisition rule that scores candidate triples $(\lambda, k, n)$. It samples functions from the per-source posteriors, sums them with weights $w_k$, and estimates how much a simulation of source $k$ with $n$ primaries would reduce uncertainty about the location of the minimum, divided by the predicted CPU cost. The cost is modeled per source, as Snoek et al. do for runtime. The third is a cost accounting that reports, for every run, surrogate fitting time against simulation time. This checks empirically whether surrogate overhead ever matters here, a question that neither Bayesian optimization paper measures on its own tasks.

The analytical work is limited in scope. We will state conditions on the weights and per-source uncertainties under which spending simulations on the source with the largest weighted posterior variance near the incumbent design is near-optimal. We will then check whether the learned acquisition behaves this way.

Baselines

The naive reference point is the hand design. For the spectrometer testbed, this is the lead/tin/copper-plus-veto configuration described by Hai et al. For the underground testbed, it is a Majorana-like nesting of polyethylene, lead, and copper at published proportions. We simulate both at high statistics in our own geometry. We cite the Majorana background figure only as a scale check. We will not reproduce that figure, since we lack their assay inputs.

The second baseline is quasi-random Sobol search under the same CPU budget. Fair Benchmark for Unsupervised Node Representation Learning found random search comparable to tree-structured Parzen estimators when budgets are tight, so this baseline is a serious one.

The third baseline is single-task Bayesian optimization on the total $B$, following Snoek et al.: one Gaussian process on the log total, expected improvement per second, and every trial simulating all sources at a fixed primary count.

The strongest baseline is multi-task Bayesian optimization in the style of Swersky et al. The sources are treated as tasks with a learned correlation matrix, and cost-aware entropy search selects the source, with the same primary-count choices available. All methods share the same simulator, geometry, parallel worker count, primary-count menu, total CPU-hour budget, and initial Sobol design. Each method's own settings are tuned on the synthetic test problems described below, not on the target geometry.

For shielding optimization, no reported result was found on Lacuna for any of these baselines. The only reported multi-task numbers are the ML benchmark figures quoted above, which do not transfer. "Beating" a baseline therefore means that on at least two of three testbeds, our method reaches a design whose independently verified background lies within the best-known value's confidence interval using no more than half the CPU-hours of the baseline, or reaches a lower verified background at equal budget.

Two ablations isolate the key ingredient. The first keeps the per-source Poisson surrogates but always simulates all sources together, which isolates the benefit of choosing which source to simulate. The second replaces the Poisson likelihood with Gaussian noise on log counts plus a small offset, which isolates the benefit of handling zero counts properly.

Why not just…?

Why not just use multi-task Bayesian optimization as published? It is the real threat, and it may do well when the sources happen to be positively correlated over the design range. We expect it to struggle in two regimes. When sources are anti-correlated, the learned correlation must discover the relation from sparse data. When source weights change, a model trained on weighted totals has to relearn. We build a testbed where lead thickness trades gamma suppression against muon-induced neutrons, and we measure regret as a function of the fitted correlation. If multi-task optimization matches us there, the contribution shrinks to the reweighting benefit, and we will say so.

Why not just run random or Sobol search on a cluster? For two or three thickness parameters this may be enough. The spectrometer grid sweep will show directly whether random search is competitive in low dimensions. The case for a model grows with dimension and with the spread of per-run costs, since neutron and muon simulations are far costlier per useful count than gamma simulations. We report the crossover dimension rather than assume it.

Why not just optimize an analytic attenuation model and confirm with a few simulations? Exponential attenuation captures primary gammas but misses secondary X-rays from lead, neutrons produced when muons break up nuclei, and veto timing effects. These are exactly the effects that set the background floor after good shielding. We use the analytic model as the surrogate's prior mean and report the verified background of its optimum as a baseline. If that optimum is within error of ours, the simulations were not needed for that testbed.

Why not just train a neural surrogate with shared layers and one output per source, which scales linearly? This is the natural multi-task-learning answer to cubic Gaussian-process cost. We include it. We expect it to be poorly calibrated at the few hundred runs typical here, but we will measure the run count at which it overtakes the Gaussian process, alongside our timing of surrogate versus simulation cost.

Proposed Experiments

The first testbed is synthetic: additive test functions with known per-source structure, controllable anti-correlation, and Poisson observations. It gives exact regret curves and a direct test of the anti-correlation hypothesis.

The second is a sea-level gamma spectrometer modeled on Hai et al. It has lead, tin, and copper thicknesses, an optional polyethylene layer, a material-order choice, and veto threshold and time gate, with gamma, neutron, and muon sources. This space is small enough for a dense high-statistics grid that serves as ground truth.

The third is a Majorana-like underground germanium geometry with a mass budget and an intrinsic-radioactivity source whose rate scales with copper and lead volumes. For this testbed we report verified background and mass.

Metrics are the verified background at the recommended design, measured by a fresh high-statistics run, against cumulative CPU-hours. We also report calibration, meaning how often the true background falls inside the model's stated interval, and the split of wall-clock time between surrogate and simulator.

The decisive checks are four. Our method must beat multi-task optimization on the anti-correlated synthetic and spectrometer settings. The ablation without source selection must lose most of the gain, confirming that choosing which source to simulate carries the benefit. After the source weights change to model moving the detector underground, the re-optimized design without new simulations must match a full rerun within error. And the optimizer should recover descending-atomic-number ordering unprompted. Failure on the first check would falsify the core claim.