When Muon Tomography Meets Unknown Cargo: A Falsification Study of High-Z Detection

Abstract

Can cosmic-ray muons detect hidden high-Z material in unknown cargo? A plain-language summary of a one-day computational study. Quantum Clarity LLC · September 2026 Repository: muon-tomography-lab The short version We built a computer simulation to test whether a particular statistical method could detect a block of dense, heavy metal hidden inside cluttered cargo, using nothing but naturally occurring cosmic-ray particles. The simulated volume was a one-metre cube of surrogate cargo, not a full shipping container, and the hidden object in the decisive experiments was a 10 cm cube of tungsten. It worked well when the method was given an exact description of the surrounding cargo. It stopped working entirely when the cargo was unknown — even though the unknown cargo came from the same random generator as the cargo descriptions already in the method's library of possibilities. That is a negative result, and it is the point of this report. It rules out one plausible approach and identifies where the real difficulty sits. The idea Cosmic rays constantly strike the Earth's atmosphere and produce muons — heavy, electron-like particles that pass through most matter. Near sea level, integrated over all arrival directions, of order one muon lands on every square centimetre of ground every minute, day and night. Muons bend slightly as they pass through matter, and they bend much more sharply when passing through dense, heavy elements — uranium, tungsten, lead — than through steel, water, or air. So if you place particle detectors above and below a shipping container and measure how much each muon bends, you can in principle infer what is inside. This is real and not speculative. Los Alamos National Laboratory pioneered the technique in the early 2000s and working systems have been fielded. It has two appealing properties: it needs no radiation source, and dense shielding is itself strongly scattering, so the material a smuggler would add to conceal a threat can become part of the detectable signature rather than hiding it. The catch is that cosmic muons arrive at their own pace. You cannot turn up the beam. Scanning a container properly takes a long time, and port throughput is the operational bottleneck. The question we set out to answer Our original hypothesis: containers differ in how hard they are to inspect, so a system that decides for itself when it has seen enough should be able to clear easy containers quickly and spend its time on difficult ones. To test that, we first had to establish a baseline: how well does conventional detection work, and at what exposure? That baseline question consumed the entire study, because the answer kept turning out to be more complicated than expected. What we built A four-layer simulation, deliberately separated so that each layer could be tested independently: The world — cargo scenes described in metres: dozens of randomly placed blocks of steel and water, sometimes containing a hidden block of lead, tungsten, or uranium, sometimes containing a harmless dense steel block as a decoy, sometimes containing nothing unusual. The physics — muons generated with a realistic angular and energy distribution, traced through the scene, scattered according to the standard physics of multiple Coulomb scattering. The measurements — an immutable record of what a detector would see for each muon: where it entered, where it left, and how much it bent. The analysis — several different methods for turning those measurements into a verdict: threat, or no threat. The separation matters. An early version accidentally let the analysis settings change the physical scene, which invalidated a full day of results. The rebuilt version verifies by cryptographic hash that the world and the measurements are identical no matter how they are later analysed. Everything ran on a single desktop computer with one consumer graphics card. Simulating several hours of real cosmic-ray exposure takes under a second. What we tested, and what happened We worked through nine explanations in sequence. Each was killed by a test designed specifically to kill it. Hypothesis Result Maybe we just need to scan for longer No. Performance stopped improving. Maybe where muons cluster is itself informative No. Chance-level once instrument bias was removed. Maybe the heavy-metal signal is inherently weak No — the signal is strong. Maybe the problem is searching the whole container Part of it, but not all. Maybe an exact statistical model would solve it Yes, spectacularly — when the cargo was known. Maybe the cargo can be one of a known set of possibilities Appeared to work — but the method simply identified which one. Not a real test. Maybe it works on genuinely unseen cargo No. Performance fell to chance. Maybe that collapse was a numerical artefact No. It survived the correction. Maybe a second viewing angle would rescue it No measurable benefit. Two findings deserve emphasis. We measure performance with AUC, which has a direct meaning: given one volume that contains a hidden object and one that does not, AUC is the fraction of the time the method ranks them the right way round. An AUC of 1.0 is perfect, 0.5 is a coin flip. The information is really there. When we told the analysis exactly where the hidden object was and asked only "is this dense or not," it reached AUC 0.97. The physics carries the signal. Finding it is the hard part. When the same analysis had to search the volume without being told where to look, it fell to AUC 0.72. And when the surrounding cargo was unknown as well, it fell to AUC 0.45–0.52 — no better than guessing. We then tried to break our own conclusion, and could not. The statistical model contained no allowance for detector measurement error, which can make such a method wildly overconfident and could by itself have produced the collapse. We added a realistic allowance — 1 milliradian of angular error, 1 millimetre of position error — declared the correction and its pass/fail threshold in advance, and reran. The overconfidence disappeared as intended: the score scale fell by a factor of six. The collapse did not. Blind performance stayed at AUC 0.47–0.50. That is what makes the conclusion below a finding rather than a numerical artefact. What this means Within this model, the decisive unresolved bottleneck is neither the strength of the physical signal nor the length of the scan. It is that the method must work out what the ordinary cargo looks like and whether something dangerous is hidden inside it at the same time — and those two problems interfere with each other. (That is a statement about this study. It does not establish that physics or exposure could never be limiting in a higher-fidelity system.) An analysis given a perfect description of the benign cargo performs superbly. The same analysis, facing cargo it has never seen, cannot separate "this is an unusual but harmless load" from "this contains something dense." That reframes the research question. The interesting problem is not "how do we scan more cleverly" or "where should we point the detector." It is: How can a detector reason about essentially unlimited variation in ordinary cargo without needing an exact description of it? That is a substantially harder statistical problem, and we did not solve it. What this study does not show These limits are as important as the findings. It does not show that muon tomography does not work. Working systems exist and are fielded. This study examined one specific analysis method inside one specific simulation. It does not show that no method can handle unknown cargo. It shows that this method required the true cargo to be present in its library of possibilities, and failed when it was not. The method was not trained in the machine-learning sense; it averaged over sixteen explicit cargo hypotheses. Better approaches may well exist; we did not test them. The physics was never checked against an independent simulator. We used a fast approximation throughout. Validating it against Geant4, the standard particle-physics simulation package, is the obvious next step and was not done. Every number here is conditional on that approximation being adequate. The final test used a very short exposure — about 500 usable muon tracks, roughly nine seconds of real-world scanning through a one-square-metre aperture. That budget was chosen for reasons that later proved partly artefactual. A longer-exposure test was authorised in advance only if intermediate results cleared a threshold; they did not, so it was not run. The simulated cargo was simple — rectangular blocks of four materials, from a single random generator, inside a one-metre cube rather than a full-scale container. Real cargo is not. The detector was idealised: no efficiency losses, no absorption modelling, and scattering treated as the only measurable effect. Why publish a negative result Because the specific thing that failed here is something people generally assume works. The step from "this method works when the background is known" to "this method works in the field" looks small and is routinely taken without checking. We checked, and the step does not hold. Anyone planning to build on this kind of approach can now spend their effort on the part that is actually hard. The study also documents its own errors. Several conclusions were retracted mid-study when a check failed: a bug that let analysis settings alter the simulated world; a scoring routine that silently discarded unmeasurable cases and reported perfect accuracy as a result; a comparison that unknowingly handicapped one side; a statistical test that was numerically meaningless. Each was found by a test built to find it, and each is documented in the repository alongside the correction. The falsification trail is the contribution. The negative answer is what it produced. Reproducing this All