space-rha.comShort courseChapter 2 · Shielding and transport

Chapter 2 · Lesson 5 of 6 · about 10 minutes

Reviewing a transport result

A missing cover, a filled cavity or a wrong material can change the answer before any of the physics comes into play. Inspect the model and its mass first. Then check whether the calculation represents the hardware and the environment you intended.

A believable model

Visible mistakes

A missing lid, a cavity modeled as solid, a wrong density or two overlapping solids are common ways a mass model goes wrong. Each of them shows up in the geometry diagnostics, with no dose calculation needed.

A useful cross-check

Compare the modeled mass, the dimensions and the material assignments with the controlled design data. Confirm that the openings and covers in the model are the ones the flight configuration will have.

Precision

Running more histories makes the statistical error bar smaller. It does nothing for a wrong source, a wrong mass model or the wrong physics.

Convergence and discrepancy checks

Treat the numerical questions (geometry sampling, score statistics, solver settings) separately from the uncertainty in the source and the physics. When two tools disagree, bring their inputs into line and cut the problem down until the difference can be located. Two tools can agree on a total dose while the particle components behind that total differ.

Matched case. Make the two cases the same: species and energy bounds, angular definitions, duration, materials and densities, target size and material, and output units. Compare the physics models and the data releases, not only the settings on the user interface.

Convergence. Increase the angular or history sampling that matters for your score and see whether repeated runs give a stable answer. Vary the energy bins, the scorer dimensions and any numerical setting that could matter. What has to converge is the score you will base the decision on. The total number of histories says little about that on its own.

Discrepancy. Before explaining a difference in the totals, compare the intermediate spectra and the contribution of each particle type. Reproduce an analytic limit or a published simple geometry with each tool. Do not average two tools that disagree until you understand why they disagree.

Independent evidence. A second front end on the same transport kernel checks the workflow around the kernel, and only that. Validating the physics needs measurements or benchmarks whose scope and uncertainty are known.

3D scene

Reviewing the model

Four classic model errors to find: a missing lid, a filled cavity, a wrong density and overlapping solids. The diagnostics are geometric; they show what a transport result built on the error would inherit.

  1. The mass model as it should be: walls, board, package and target, with the declared materials and densities. Compare its mass, its dimensions and its material assignments with the controlled design data, and confirm that the openings and covers are the ones the flight configuration will have.

  2. The lid is absent from the model. Every ray toward that face now meets far less material than the hardware will present, and the diagnostics show it directly: the path histogram and the total mass change before any physics is run.

  3. A cavity modeled as solid adds mass that is not there and shields the target from directions that are in fact open. The error is invisible in a dose total and obvious in the geometry: the modeled mass exceeds the design mass.

  4. The geometry is right and the density is wrong, so every ray through that solid carries the wrong areal mass. A comparison of modeled mass against the controlled design data catches it; running more histories does nothing for it, because precision is not accuracy.

  5. Two solids occupy the same space, and the material along a ray through the overlap is counted twice or resolved by whichever rule the tool applies. The geometry diagnostics flag the overlap; the dose calculation would silently inherit it.

The original space-rha.com environment and mission renderer (NASA imagery and spacecraft models, credits in the view), with the course scenes drawn by the same renderer. Trails, clocks and counts are illustrative; the scene computes no dose or rate.

Exercise

Write a verification plan that keeps three things apart: sampling precision, checks on the implementation and the validity of the physics.

Spend the effort on the differences that could change the engineering conclusion.

Going deeper

Numerical validity and physical evidence

The uncertainty in a transport result comes from the source, the geometry, the interaction data and the numerical approximations. Weighted or rare-event estimators can converge unevenly. Production thresholds and stepping rules change how an implementation behaves. A credible result needs diagnostics on the estimator and, separately, evidence that the physics being modeled applies to the case.

Estimator diagnostics. Look at which histories contribute to the score, at the tail of the weight distribution, at how independent runs compare and at the uncertainty in each bin or response that feeds the decision. A nominal history count and a single relative-error figure can hide a contribution that was barely sampled.

Threshold semantics. In Geant4, the secondary-production range threshold decides which secondaries are created; it is not a cutoff on tracking. Check the processes selected, the production settings, the tracking policy and how energy is accounted for. Test the sensitivity at material interfaces and in small scoring volumes.

Verification and validation. Geometry and normalization tests tell you whether the implementation is correct. Analytic or code-to-code benchmarks tell you whether a defined calculation comes out right. Comparisons with accelerator or flight data tell you whether the physics applies, within the configuration and measurement uncertainty of those data. Each kind of evidence answers its own question.

Decision sensitivity. Vary the source, material and geometry assumptions that control the endpoint over their plausible ranges. Keep numerical precision, model discrepancy and source uncertainty as separate items. If the result sits near an acceptance limit, better evidence on the inputs is generally worth more than more histories.

Takeaway

Check the physical inputs before you read anything into a precise result.

Check yourself

A Monte Carlo dose comes with a 2 % statistical uncertainty. Is that the uncertainty of the dose?

No. It is the uncertainty of the sample given the model. The model carries its own uncertainty from the source spectrum, the geometry as built, the material compositions and the physics options, which is generally far larger and is bounded by sensitivity cases rather than by running longer.

What does a control calculation show that the main calculation cannot?

That the setup reproduces a known answer: a slab or sphere case with a published value, or a measurement. If the control fails, the main result is not trusted whatever its statistics; if it passes, the setup is credited for the cases it resembles.

Two tools agree on the total dose at a part to within a few percent, and the result is accepted on that basis. What can the agreement hide?

Two tools can agree on a total dose while the particle components behind that total differ, so the intermediate spectra and the contribution of each particle type should be compared before the totals are read. If the two tools are two front ends on the same transport kernel, the agreement checks the workflow around the kernel and nothing else; validating the physics needs measurements or benchmarks whose scope and uncertainty are known. And when two tools disagree, they should not be averaged until the reason is understood.

The modeled mass of an enclosure comes out above the controlled design mass. Which of the classic model errors does that point to, and why would a dose total not have shown it?

A cavity modeled as solid adds mass that is not there and shields the target from directions that are in fact open; overlapping solids, whose material a tool may count twice, can do the same. The error is invisible in a dose total because the calculation silently inherits the geometry it is given, while the geometry diagnostics (the modeled mass, the path histogram and the overlap flags) show it before any physics is run. Running more histories does nothing for it, because precision is not accuracy.

A Geant4 setup lists a secondary-production range threshold, and a reviewer reads it as a tracking cutoff below which particles are dropped. What is the distinction, and where does it matter most?

The production threshold decides which secondaries are created; it is not a cutoff on tracking. The review should look at the processes selected, the production settings, the tracking policy and how energy is accounted for, since each changes how the implementation behaves. The sensitivity should be tested at material interfaces and in small scoring volumes, where the local electron fluence and dose can change and the usual bulk-dose assumptions can break down (lesson 2.3).

Further reading