Six-second helicopter flyover identifies mixed radioactive sources

A helicopter crossing a contaminated site at roughly 10 m/s spends only a few seconds near any given source. A team from the Paul Scherrer Institute (PSI) in Switzerland and the University of Michigan reports that this is enough. In a paper published on 26 October 2026 in Nature Communications, they use 6 s of single-pass data benchmarked against independent in-situ and laboratory assays. They report decisive identification of source mixtures (>5σ), meter-scale localization and percent-level activity recovery. The university's summary puts the numbers more concretely: sources pinpointed within two meters and activities estimated within two percentage points of their true values.

The advance is mainly computational. The detector is a commercial NaI(Tl) array. What changed is how the team builds, and how quickly it can evaluate, the physical model against which sparse counts are compared.

01Why a flyover spectrum is hard to read

Each radionuclide emits gamma rays at characteristic energies. In principle, the measured spectrum identifies the isotope and its amount. In practice, every material has a fingerprint made of its gamma-ray energy spectrum, and that fingerprint degrades in flight. Photons scatter in the air and in the aircraft's structure, losing energy before they reach the crystals. Counts are also scarce. PSI says that of around 10,000 photons emitted by a ground source, only one reaches the helicopter, and even that one is scattered by metal surfaces inside it.

The authors describe the resulting inversion as ill-posed. Mobile gamma-ray spectrometry is the gold standard for source identification, but current approaches are inaccurate because data are scarce and the inversion is ill-posed. The standard remedy is full-spectrum analysis. Instead of integrating a few energy windows, it fits the whole measured spectrum to simulated templates. The problem is cost. Faithful templates need Monte Carlo transport through air and airframe, and for each flight geometry, atmosphere and source position that can take hours or days. No one can run that during an emergency.

02Splitting the transport problem in two

The central trick is factorization. The team separates a photon's journey into atmospheric transport and the response of the helicopter and detector. According to the Michigan summary, atmospheric calculations account for air temperature and humidity, while the aircraft portion uses precomputed functions specific to the helicopter model. The paper's methods describe a "gamma-ray flux bank" generated with the FLUKA code's radioactive-decay module. They also describe separate platform simulations that record energy deposition in each crystal and reduce it to expected pulse-height spectra.

The heavy Monte Carlo work therefore happens once, ahead of time. The preprint reports about 10² core-hours per source, with 2 × 10⁹ primary particles spread over 10² independent runs. A given flight geometry then needs only a lookup and a combination step. Breitenmoser says this reorganization, rather than extra computing power, is what cut simulations that took hours or days to microseconds.

The platform model builds on earlier PSI work. The Swiss system has four prismatic NaI(Tl) crystals with a total volume of 16.8 l, mounted in the cargo bay of a Super Puma. In an earlier validation study, the team reported a median relative spectral error below 10% for most radionuclides in the detector-response model.

03Bayesian inference over thousands of templates

With templates this cheap, the inference can afford to be thorough. The framework compares thousands of predicted spectra against the measured one and continuously updates its statistical confidence about what lies below. The keywords of the preprint name Markov chain Monte Carlo as the sampler. The unknowns are estimated jointly: which nuclides are present, where the source is, and how active it is. Natural background from potassium, thorium and uranium in the soil is fitted as well.

The output is a posterior distribution rather than a single best fit. That matters operationally. A responder gets an activity estimate with a credible interval, and a quantified preference for "Cs-137 plus Ba-133" over "Cs-137 alone". The >5σ mixture result is a statement of that kind.

04What was actually tested

The field data come from the ARM22 exercise at the Thun military training ground on 16 June 2022. The preprint says both configurations used the same two sealed sources, Cs-137 and Ba-133, with acquisition at 1 Hz. The configurations were:

• Hover: about 5 min at roughly 90 m, used for controlled benchmarking of activities and background.
• Single pass: about 30 m altitude and about 10 m/s, over three scenarios: Cs-137 alone, Ba-133 alone, and a co-located dual source with under 20 cm separation.

Truth values were independent. Source activities came from the manufacturer's laboratory assays, and natural background was constrained by in-situ spectroscopy at 23 positions on the training ground. At 10 m/s, six seconds corresponds to a track of about 60 m (my arithmetic). The inference is therefore drawn from a short stretch of flight around the source.

For context, earlier Bayesian work on road-vehicle surveys is not directly comparable. One study driving at 50 km/h with 1 s sampling found that predicted deviations in activity and position stayed around 30%, because of uncertainty in the data. That study used different detectors and geometry. Still, it shows the gap between an approximate point-source model and a platform-specific full-spectrum model.

05Why it matters, and what remains unproven

In a reactor accident or a nuclear detonation, aircraft may be the only way to reach the affected area. The team argues that in-flight data can inform decisions on evacuation, sheltering and potassium iodide distribution. For nuclear security, the useful property is mixture resolution. Telling a medical Cs-137 source from a mixed or industrial inventory, and doing it in one pass, changes what a search team does next. The authors also list routine facility monitoring, geochemical mapping and even extraterrestrial radionuclide abundance as applications. The press materials add environmental remediation and mineral exploration.

The limits are visible in the paper's own design:

• Few nuclides: The tests used two sealed point sources. A real release involves dozens of isotopes, extended plumes and ground deposition, so the template library and the inference problem both grow.
• One platform and site: The aircraft functions are specific to one helicopter model. Another airframe needs its own validated Monte Carlo model and precomputation.
• Favourable geometry: Flight tracks passed directly over the sources, which were known to be there. Searching blind over an unknown area is a harder test.
• Percentages: The "two percentage points" and "two meters" figures come from the press releases. The paper itself says "percent-level" and "meter-scale". Readers should check the per-scenario tables in the paper before citing a single figure.

The next steps are not spelled out in the sources I reviewed. Reasonable ones would be more nuclides, extended and distributed sources, and runs under operational conditions. The data are archived on Zenodo, so independent groups can test the approach on the same spectra. The main result is that a physics-aware factorization made a rigorous Bayesian treatment fast enough to apply to data from a flight lasting a few seconds.