National Data Centres watch a global network for signs of a nuclear test. A hard, poorly-supported question keeps recurring: do two anomalous detections — sometimes continents and days apart — share a common source? RaDIA turns that manual, by-eye comparison into a transparent, interactive procedure.
Radioactive particles travel thousands of kilometres, so a single release can surface at stations days and continents apart. Deciding whether two detections share a source means comparing their source–receptor sensitivity (SRS) fields — backward atmospheric maps of where a signal could have come from. The CTBTO tool ecosystem never made that comparison systematic.
In the study, analysts described building an Excel sheet of the links they believe exist, then emailing it round for validation. RaDIA replaces that ritual with an inspectable, repeatable procedure.
An interpretable, parameterized overlap operator that formalises tacit expert practice for cross-sample comparison.
A progressive-disclosure interaction design that reduces cognitive load while preserving analyst agency.
A coordinated-multiple-view design that makes intermediate analytic states inspectable and comparable side by side.
A problem-driven design study with three NDCs, traced from observed breakdowns to requirements to implemented features.
RaDIA's linking rests on one artifact — the SRS field — and one number — the overlap count. Here's what they mean, in plain language.
Every radionuclide detection gets an SRS field (source–receptor sensitivity) — the output of a backward atmospheric-transport model. Starting from the detector's location and collection time, it runs the winds in reverse and asks, cell by cell across the map and back through time (up to 14 days): if material had been released here, then, how much would have reached this detector? The number in each cell is a dilution value — the concentration at the detector per unit emitted at that source. So an SRS field is a 4-D grid (lat, lon, time, sensitivity): a heatmap of candidate source regions, bright where a release could plausibly explain the detection, vanishing where it couldn't.
Dilution values span roughly six orders of magnitude (≈ 10⁻²⁰ to 10⁻¹⁴), and the bottom of that range is essentially noise. The dilution threshold τ = 10⁻¹⁸ is the cutoff that separates cells with meaningful atmospheric sensitivity from that noise. It encodes how analysts already read these maps — they focus on the regions that light up, not on comparing every faint value. Above τ a cell belongs to the field's plausible source region; below τ it's ignored. τ is a domain standard and stays fixed — it isn't a knob the analyst tunes.
To compare two detections, RaDIA binarises each field at τ — a cell is either inside the plausible source region or not — and counts the cells where both fields are inside, at the same place and the same time (within a ±1-day window). That count is the overlap: large when the two detections point at the same release, near-zero when they don't. It's reported as an absolute cell count and as a percentage of the reference field's region.
It's a count of cells above the threshold, not a weighted sum of every value — a deliberate simplification the analysts preferred, because a countable, inspectable number is easier to trust in a high-stakes call.
With those three terms in hand, make it concrete: the reference field is the detection under investigation, each candidate field is another sample tested against it, and the overlap is the cells where both exceed τ. Drag the backward-time slider to step through timestamps: each field's sensitivity sweeps back from its own station toward a source region, and the overlap lights up at the timestamps where the two converge. Switch the candidate to compare a related detection with an unrelated look-alike.
Illustrative, synthetic SRS fields on an abstract map — not the paper's real atmospheric-model output. The mechanism (backward-in-time sensitivity converging on a source region, counted where both fields exceed the fixed threshold) and the related-vs-unrelated outcome mirror the paper's reactor-leak case; the real maps and numbers are in the PDF.
Once the operator scores candidate samples against the reference, RaDIA lays the associations out as a Sankey. The reference sample sits in the centre; candidates detected earlier flow in from the left and those detected later flow out to the right — the temporal window is ±days, so both directions count. Each ribbon's thickness is its spatial-overlap strength, and node colour flags related (violet) versus excluded (amber) samples. Hover a ribbon to inspect its link.
Thickness follows overlap, not time. KZX44 sits closest to the reference in time, yet its ribbon is the thinnest — temporally near, but spatially unrelated.
Schematic of RaDIA's Sankey encoding — reference centred, candidates earlier (left) and later (right), ribbon width = overlap strength. JPX38 is the case-study link RaDIA surfaced; the other station codes and timestamps are illustrative. The real Sankey is generated from computed overlap metrics — see the PDF.
A simple, cached operator does the linking; coordinated views make the result inspectable.
Match each sample to the nearest SRS file within ±24 h, then count grid cells where both the reference and candidate sensitivity fields exceed a dilution threshold (τ = 10⁻¹⁸) inside a temporal window (default ±1 day). Simple, cached and physically interpretable — it formalises what experts already do by eye.
IMS stations show as colour- and letter-coded markers; SRS fields render as logarithmic heatmaps across six orders of magnitude. A 336-hour backward slider animates how sensitivity evolves before collection; the reference station gets a yellow halo, and a selected candidate's overlap region is drawn straight onto the map.
A temporal Sankey places candidates left-to-right by time separation, with link width = overlap strength and node colour = station. The table lists station, time, isotope concentration and absolute/relative overlap — sort by any column to prioritise the strongest, closest or most concentrated candidates.
Progressive disclosure turns a screen full of knobs into a step-by-step path. Each action unlocks the next, so newer analysts are walked through the procedure instead of facing everything at once.
In a domain with few experts and tightly controlled data, the right move is deep engagement with a handful of analysts, not a big lab study.
Pre-design interviews and workflow walkthroughs, iterative prototyping from sketches to a working dashboard, then expert review.
Four semi-structured interviews (60–90 min) plus walkthroughs where analysts demonstrated their current CTBTO-tool practice.
Operational data and procedures are tightly controlled and judgement needs situated expertise — proxy participants would measure the wrong thing.
Analysts used the prototype while narrating; consistent feedback across independent institutions points to shared challenges, not personal preference.
RaDIA wasn't run as a controlled benchmark — domain access doesn't allow one. The signal comes from a real reactor-leak case (north-eastern Asia, March 2024) and consistent expert feedback.
RaDIA flagged Sample 3 (JPX38, 17 March) as related — an association that initial manual review had not picked up as strong.
Sample 4 was temporally close, but its sensitivity extended toward Russia; the multi-view check correctly excluded it as unrelated.
The shared sensitivity region 177 hours before sampling matched the analysts' suspected reactor location.
All three NDCs recognised and valued the systematic-linking need — consistent feedback across independent institutions.
Honest framing: a single real-world case reviewed with participating analysts, plus qualitative feedback from four experts. No controlled study of verification accuracy is claimed.
The design engaged analysts from three smaller NDCs, which limits a broader, stronger evaluation.
The algorithm counts only cells above the dilution threshold rather than weighting all values — chosen for interpretability, but it discards magnitude information.
RaDIA has no multi-user collaboration or real-time data ingestion, so analysts still gather and wrangle the data themselves.
It validates the analysis workflow; it does not measure verification performance, and the evaluation is qualitative rather than a controlled benchmark.
Analysts had no tool to systematically link radionuclide detections that might share a source — so they compared atmospheric sensitivity maps by eye and tracked hunches in spreadsheets passed around for validation. RaDIA folds that into a single dashboard: a transparent overlap operator proposes candidate links, and coordinated map, timeseries, Sankey and table views let analysts confirm or reject them. In a real reactor-leak case it surfaced a link manual review had missed, while correctly filtering an unrelated detection.