Radioactive Decays as a Moisture Gauge in Liquid Argon Detectors

Large liquid argon detectors record particle tracks by letting ionization electrons drift for several meters to readout wires. Trace impurities, oxygen and especially water, capture those electrons along the way, so the recorded charge shrinks with drift distance by an amount that changes over time and differs from place to place. Today this attenuation is calibrated offline from cosmic-ray muons and dedicated purity monitors. That approach will get weak exactly where it matters most: deep underground, where muons are rare, and during transients such as pump stops or slow moisture outgassing. The proposal is to use the detector's own intrinsic radioactivity as the calibration source instead. Radon daughters and argon-39 are the dense, uniform, always-present background that rare-event physicists normally try to reject. Many thousands of these small, low-energy deposits ("blips") are spread through the whole volume at every moment.

The core idea, in two sentences: treat the radioactive background as a large population of simultaneous probe sources, and feed their charge-versus-drift-time statistics into a slowly updated estimate of the electron-capture rate on a spatial grid. That estimate is guarded against the compounding errors that affect any recursive tracker, and it is tied to a physical impurity model extended with the missing water term. If this works, one measurement stream would give underground-viable, spatially resolved, near-real-time purity tracking. It would also give the first data-driven constraint on cryogenic water dynamics, which current impurity models explicitly cannot close.

Core Claim

The electron lifetime $\tau(\mathbf{x}, t)$ at position $\mathbf{x}$ and time $t$ sets the collected charge $Q = Q_0 \exp(-t_d / \tau)$, where $Q_0$ is the charge at creation and $t_d$ the drift time. To first order the capture rate is a sum over impurity species,

$$ \frac{1}{\tau(\mathbf{x},t)} = k_{O}\,c_{O}(\mathbf{x},t) + k_{W}\,c_{W}(\mathbf{x},t) + \dots, $$

where $c_O$ and $c_W$ are the oxygen and water concentrations and $k_O$, $k_W$ their capture coefficients.

We make three claims. First, the drift-time dependence of blip charge spectra can estimate $1/\tau$ on a coarse spatial grid, over time windows short enough to resolve purity transients. Second, a recursive estimator of this field can accumulate drift, much as video memory trackers do. It needs an explicit arbitration step that compares it against an independent single-window estimate and resets it when the two diverge. Third, fitting the resulting $1/\tau(\mathbf{x},t)$ history with an impurity model that adds a surface-sorbed water reservoir will predict future lifetime better than an oxygen-only model. That gain would count as indirect evidence about the water term.

What is new is using intrinsic radioactivity as the calibration signal rather than as clutter. A second new element is borrowing closed-loop memory correction from video tracking for a physical state estimate. Nothing in the retrieved literature does either. The ProtoDUNE-SP purity study (Spatial and temporal evaluations of the liquid argon purity in ProtoDUNE-SP) lumps all impurities into one effective lifetime measured from muons and purity monitors, and applies the result offline. The MicroBooNE radon study (Measurement of ambient radon progeny decay rates and energy spectra in liquid argon using the MicroBooNE detector) reconstructs radon-daughter decays to measure their rates, not to track purity.

Research Program

The work builds on three pieces of prior art, each playing a different role.

The motivation comes from two results. ProtoDUNE-SP shows that purity is spatially non-uniform, differing between the inlet and outlet sides. It also shows that purity dips when recirculation pauses, so a single time-binned scalar lifetime is biased (Spatial and temporal evaluations of the liquid argon purity in ProtoDUNE-SP). The impurity-concentration model of Modeling impurity concentrations in liquid argon detectors tracks oxygen across four phases with coupled ordinary differential equations. It matches data to a 2.9% root-mean-square fractional deviation, but the authors state that water cannot yet be modeled. Its cryogenic solubility is unknown, it sticks strongly to cold walls, and it may form an ice-like fifth phase. Sorption also dominates below 10 ppb, which is the regime a clean detector lives in.

The method comes from two sources. The first is the differentiable detector simulator of Differentiable simulation of a liquid argon time projection chamber. It fits entangled lifetime, diffusion and field-distortion parameters jointly, to within 1% in simulation. The authors flag a circularity on real data: the true energy deposits are unknown. Radioactive decays reduce that circularity because their energy distributions are fixed by nuclear physics. The radon-chain alpha lines and the argon-39 beta spectrum act as standard candles, which muon tracks do not provide. The second method source is the closed-loop correction scheme of Detector-in-the-Loop Tracking: Active Memory Rectification for Stable Glottic Opening Localization, which supplies the arbitration logic. In that work, a recursive tracker (SAM2) and an independent single-frame detector are normalized onto a common confidence scale. A small state machine then decides whether to blend them, trust the tracker, or hard-reset the tracker's memory.

The program has four parts.

The first part is blip selection. A lightweight classifier separates isolated low-energy deposits from track and shower hits, using the sparse-convolution segmentation approach of Semantic segmentation with a sparse convolutional neural network for event reconstruction in MicroBooNE. That network was trained on primary interactions and, as the file notes, risks confusing radioactive deposits with delta rays or small showers. We retrain it with an explicit radioactive-decay class. This is where scaling to many targets enters: in a surface detector, blips must be separated from hundreds of crossing cosmic tracks, and association errors bias the charge statistics. Blips near tracks are therefore down-weighted using the reconstruction hierarchy from Pandora (Reconstruction of interactions in the ProtoDUNE-SP detector with Pandora).

The second part is the per-window estimate, which plays the role of the "detector" in the closed loop. Within each spatial cell and time window, we fit the drift-time dependence of the blip charge distribution. The fit is a forward-model match, running the differentiable simulator on the known decay spectra, rather than a peak fit. This avoids the heavy charge quenching of alpha deposits, which a peak fit would struggle with. The fit also yields a width, which serves as the confidence of the estimate.

The third part is the recursive estimate, which plays the role of the "tracker." A state-space filter over the $1/\tau$ grid uses the impurity ODE as its dynamics, so it smooths the noisy per-window estimates. Its failure mode matches the recursive memory corruption in Video Object Segmentation Using Space-Time Memory Networks. A biased window, such as one with an unmodeled field-distortion change, is absorbed into the state and propagates forward. An arbitration step modeled on the four-state controller of Detector-in-the-Loop Tracking: Active Memory Rectification for Stable Glottic Opening Localization compares the filter against the per-window estimate and the sparse muon and purity-monitor readings. When disagreement persists beyond calibrated bounds, it resets the affected cells.

The fourth part is the water term. We extend the ODE with a latent surface-bound water reservoir and a sorption/desorption exchange with the liquid. The reservoir's parameters are fit only through their imprint on $1/\tau(\mathbf{x},t)$: the slow recovery tails after transients and the spatial asymmetry between inlet and outlet. We do not claim to measure water directly. The claim is falsifiable. The water-augmented model must forecast held-out lifetime trajectories better than the oxygen-only model with an equal number of free parameters, and the fitted exchange rates must be stable across independent transients.

Baselines

The naive reference point is a single lifetime per run period, taken from the purity monitors and applied uniformly to every hit. This is the "do nothing beyond standard practice" level. It establishes how much bias in charge per unit length (dQ/dx) is left on the table.

The strongest existing method is the ProtoDUNE-SP cosmic-muon procedure of Spatial and temporal evaluations of the liquid argon purity in ProtoDUNE-SP. It uses cathode-crossing and cosmic-ray-tagger-matched muon tracks in time bins, cross-checked against purity monitors. The paper reports lifetimes above 30 ms, often above 100 ms, and chi-square per degree of freedom around 1 between the two muon methods on the BL anode assemblies. No reported per-bin lifetime uncertainty or transient-detection latency was found on Lacuna for this method. We will therefore rerun it ourselves on the same data periods, with the same spatial binning and time windows, and use its output as the reference.

Fairness hinges on the muon rate. At the surface, muons are abundant and the baseline will be strong. We will evaluate at the full surface rate and at muon rates subsampled to emulate underground depth. Beating the baseline means two things. At underground-emulated muon rates, the blip method gives lower lifetime error per spatial cell and time window, and detects injected or real purity dips sooner. At full surface rates, it is statistically consistent with the muon method, with chi-square per degree of freedom near 1.

Two further baselines isolate the key ingredients. A plain Kalman filter on muon-only estimates tests whether blips add information. The blip estimator without the arbitration and reset step tests whether closed-loop correction matters. Downstream, we will also report Pandora's test-beam reconstruction efficiency, published as 86.1 ± 0.6% for 1 GeV/c pions and 84.1 ± 0.6% for 1 GeV/c protons in Reconstruction of interactions in the ProtoDUNE-SP detector with Pandora. These numbers must not degrade when our corrections replace the standard ones.

Why not just…?

Why not just keep offline muon calibration in finer time bins? At a surface prototype this is a genuine threat. Muons are plentiful there, and the ProtoDUNE-SP methods already agree well with each other. The case for blips rests on two regimes: sparse muons (underground) and fine spatial cells, where muons give few crossings per cell per window. The muon-subsampling experiment settles this directly. If blips offer no gain at underground-emulated rates, the idea fails and we will say so.

Why not just use the purity monitors? Purity monitors sample the liquid at a few fixed locations. ProtoDUNE-SP shows that purity varies with position and between the inlet and outlet sides (Spatial and temporal evaluations of the liquid argon purity in ProtoDUNE-SP), so the monitors cannot supply a spatial map. We keep them as an independent check of the reconstructed map at their positions, rather than as a competitor.

Why not just fit the differentiable simulator to real data directly, with no filter? Its authors note that real energy deposits are unknown, which makes the fit circular. The simulator is also about 25 times slower than the non-differentiable version (Differentiable simulation of a liquid argon time projection chamber). We use it only per window, on standard-candle decays, where both problems shrink. The no-filter ablation measures what the recursive layer adds.

Why not just a simple exponential smoother with no physical model? The smoother may track lifetime as well in steady periods. The impurity model earns its place only if it forecasts recovery after transients better. That is exactly the held-out forecasting test, so a tie there would count against the water-term claim.

Proposed Experiments

The first stage is closure tests in simulation. We generate events with radon-chain and argon-39 decays at realistic rates, overlaid with cosmic tracks. The injected lifetime fields have known spatial gradients and scripted dips that mimic pump stops. We then measure lifetime error per cell and dip-detection latency against the muon baseline across muon rates.

The second stage is real data. We apply the method to the ProtoDUNE-SP October–November 2018 cosmic run, subject to data access, and check agreement with the published muon and purity-monitor lifetimes. The main target is the transient periods, where static tables are known to fail.

The third stage is the water test. Using the BNL 20-liter stand's oxygen data as the anchor for the oxygen-only model (Modeling impurity concentrations in liquid argon detectors), we compare oxygen-only and water-augmented ODEs on held-out ProtoDUNE-SP lifetime trajectories. No reported quantitative water result was found on Lacuna, so any stable fitted water exchange rate would be a first estimate. If collaborators can run a controlled moisture injection on the test stand, that would be the decisive physical check.

Three outcomes are decisive together:

  • Underground-rate advantage: a clear accuracy or latency gain at emulated underground muon rates.
  • Surface-rate consistency: statistical agreement with the muon method at the surface rate.
  • Water-term evidence: a forecasting improvement from the water term that holds across independent transients.

The main risk is quenching and recombination of alpha charge, together with argon-39 spectrum systematics, which may make the per-window estimate too noisy. We will quantify this first in simulation, before any claims on data.