Un-Baking Space Charge: Laser-Free Electric Field Maps from Cosmic Muons

When a 3D vision system reconstructs an object from photographs, it often paints shadows and highlights straight into the object's color. The light was shared by the whole scene, but the model stored it as a property of each surface point. A liquid-argon particle detector has the same failure in a different medium. Positive ions left by cosmic rays pile up inside the detector and bend its electric field. Standard reconstruction then assumes the field is uniform, so the bend gets stored as crooked particle tracks. MicroBooNE measured these apparent shifts at up to about 15 cm (UV laser field map). The fix that works in vision carries over directly. Model the shared effect (lighting, or the field) as one physically constrained object common to many observations. Model each observation's own content (surface color, or the true track) with a simple prior. Then rely on variety across observations, meaning many lighting directions or many track angles, to pull the two apart.

We propose to recover a detector's full 3D space-charge field without a dedicated laser, using only the stream of cosmic muons that already crosses it. The model is an electrostatic potential shared by all tracks in a time window. It is constrained by Gauss's law, by the fact that the charge density is positive (the accumulated charge is positive ions), and by fixed voltages on the detector walls. Each muon is modeled as an unknown straight line. The two are fitted jointly by pushing lines through a differentiable model of how electrons drift. The same field also sets how much charge each muon deposits, through the field dependence of electron–ion recombination, so charge measurements add an independent check on the field strength. The central scientific question is identifiability: when is the field uniquely determined by cosmic tracks alone, and how many tracks and angles does that require? We treat that question with the same care inverse rendering gives to the "static illumination" degeneracy.

Core Claim

A single straight track cannot reveal distortion along its own direction. Sliding points along a line leaves the line unchanged. This is the detector version of the vision problem where, under one fixed light, a dim white surface looks the same as a bright dark one. We claim three things. First, with enough spread in cosmic-muon angles, the distortion becomes identifiable. Second, the constraints (the field is the gradient of a potential, the charge density implied by Gauss's law is non-negative, and the potential is fixed at the anode, cathode, and field cage) remove the leftover freedom that angle spread alone cannot. Third, the resulting map should match the MicroBooNE laser map to within that map's own sub-centimeter uncertainty, using data every surface detector already records.

What is new is not "use cosmics for calibration". MicroBooNE already equalizes charge response with crossing muons (charge calibration). The new elements are these:

  • the field is treated as a shared hidden cause, fitted with a physics-constrained potential, not interpolated;
  • per-track geometry and field are factorized jointly;
  • a stated condition says when that factorization is unique;
  • recombination couples the geometric and charge-based evidence inside the same fit.

What the Two Literatures Supply

The vision side supplies the method template.

  • Joint geometry, material and lighting. Extracting Triangular 3D Models, Materials, and Lighting From Images recovers meshes, surface materials and one environment light jointly from many views. The light is kept low-capacity and shared, so it cannot absorb per-surface detail. Our potential plays the environment light's role. That work assumes straight light rays to a surface; our "rays" are curved drift paths through a volume, set by a differential equation.
  • Staged fitting. UniDream shows that fitting all material channels at once smears highlights into base color, and that freezing some channels before fitting others prevents it. We test the analogous schedule: first fit geometry and field, then fit charge-related parameters. We run it as an ablation, not an assumption.
  • The underlying ambiguity. The ill-posedness analysis in Cremers et al. motivates our identifiability question directly.

The detector side supplies motivation and constraints.

  • Motivation. Wire-Cell converts drift time to position assuming a constant drift speed, so distortion is baked into its 3D charge images (Wire-Cell). Pandora's ProtoDUNE-SP analysis attributes part of its data–simulation disagreement to underestimated space charge (Pandora ProtoDUNE-SP). Both show the problem matters downstream.
  • Constraint on design. The differentiable detector simulator (Differentiable simulation of a LArTPC) fits entangled detector parameters by gradient descent, but its forward pass is about 25 times slower than the ordinary simulator, and it has only been tested on synthetic data. It also fits a handful of global parameters, not a continuous 3D field. We therefore do not differentiate through wire waveforms. We fit reconstructed 3D points, which reduces the forward model to integrating drift paths.
  • Reference. The laser method (UV laser field map) is both our reference measurement and our strongest competitor.

Research Program

The forward model. We represent the potential $\Phi(\mathbf{x})$ as the known nominal potential plus a correction. The correction is stored on a coarse 3D B-spline grid, with a small coordinate-network variant for comparison. The electric field is $\mathbf{E} = -\nabla\Phi$, and the implied charge density is $\rho = -\epsilon\nabla^2\Phi$, where $\epsilon$ is the permittivity of liquid argon.

An electron released at true position $\mathbf{x}$ drifts along

$$\dot{\mathbf{x}}(t) = \mu(|\mathbf{E}|)\,\mathbf{E}(\mathbf{x}(t)),$$

where $\mu$ is the field-dependent electron mobility. We integrate this path with a fixed-step differentiable solver until it reaches the anode. The arrival position and drift time, converted with the nominal drift speed, give the apparent point $F_\Phi(\mathbf{x})$ that reconstruction would report.

The fit. Each cosmic muon $k$ with known start time is a line $\ell_k(s)$, where $s$ is position along the line. The start time comes from tracks that cross the cathode or anode, or from matching the track to a scintillation-light flash. We minimize

$$\mathcal{L} = \sum_k \sum_i \min_s \big\| \mathbf{h}_{ki} - F_\Phi(\ell_k(s)) \big\|^2 + \lambda_Q \sum_k \sum_i \big(Q_{ki} - \hat Q(|\mathbf{E}(\ell_k)|)\big)^2 + \lambda_\rho \int \max(0,-\rho)^2\, d\mathbf{x}.$$

The terms are defined as follows. $\mathbf{h}_{ki}$ are the reconstructed hit points of track $k$. $Q_{ki}$ is the measured charge per unit length at each hit. $\hat Q$ is the charge a minimum-ionizing muon would deposit after recombination (Modified Box model) at the local field strength, and after electron-lifetime attenuation. The last term penalizes negative charge density. The potential is pinned to its nominal value on the anode, cathode and field-cage surfaces.

Staging and time windows. Hit-to-line correspondence is solved by alternating projection, in the spirit of the laser paper's iterative closest-point step. Fits are made per time window, so space-charge changes over time appear as a sequence of maps.

Identifiability analysis. We linearize $F_\Phi$ around the nominal field and compute how well-conditioned the problem is for a given set of track angles and positions. We then report which field patterns remain unconstrained, with and without the Gauss-law and positivity constraints. This yields a concrete prediction to check against simulation: the number of tracks, and the angular coverage, needed to reach 1 cm displacement accuracy in each detector region.

Baselines

Do nothing. The naive reference is uniform-field reconstruction, which leaves the full distortion in place. On MicroBooNE this means apparent shifts of up to about 15 cm (UV laser field map). We report every error metric relative to this, on the same held-out tracks.

Laser map. The strongest existing method is the MicroBooNE UV-laser map (UV laser field map). It reports:

  • maximum displacement of about 15 cm, with typical uncertainty under 1 cm;
  • field distortions up to about 30 V/cm on a 273.9 V/cm nominal field, with uncertainties below a few V/cm;
  • validation biases at the sub-centimeter level for displacements.

The comparison is fair only if the laser map is not an input. We therefore fit using cosmic tracks alone and compare displacement fields voxel by voxel, in the same grid and fiducial volume. "Beating" it is not the realistic bar where lasers reach. The claim succeeds if our map agrees with the laser map within the laser's own sub-centimeter uncertainty in laser-covered regions. It succeeds more strongly if, in regions the lasers cannot reach, our map predicts held-out track straightness and stopping-muon range better than the laser map's interpolation does.

Standard-candle charge calibration. For charge measurements, the baseline is MicroBooNE's staged calibration (charge calibration). It reduced the discrepancy between range-based and charge-based energy for stopping muons from −3.2% to −1.2%, with a 2–3% absolute energy-scale uncertainty. We rerun it on the same data with and without our field map substituted for its space-charge step. We count an improvement if the range-versus-charge discrepancy shrinks further, with the fit budget held equal.

Two naive baselines on ProtoDUNE-SP. We also compare against:

  • Pandora's reconstruction as-is, whose 1 GeV/c end-point resolution is 2 cm for 68% of events (Pandora ProtoDUNE-SP);
  • a simple displacement-interpolation method that places corrections only at track endpoints crossing known surfaces and smoothly interpolates between them.

Ablations. Four ablations isolate the key ingredients:

  • Free displacement field. Replace the potential with an unconstrained displacement field (no Gauss law, no positivity). This isolates the physics constraint.
  • No charge term. Set $\lambda_Q = 0$. This isolates the recombination coupling.
  • Narrow angles. Restrict training tracks to a narrow angular band. This directly tests the identifiability claim.
  • Staged versus joint. Compare UniDream-style staged fitting with fully joint fitting.

All variants share the same tracks, grid resolution, and optimizer iteration budget.

Why not just…?

Why not just install lasers everywhere? Where a laser exists it is the gold standard, and we use it as ground truth rather than compete with it. Its coverage, however, is limited by line of sight and entry ports. It also measures in dedicated runs, not continuously (UV laser field map). Many detectors have no laser at all. The time-window experiment below tests whether cosmic maps can follow changes between laser runs.

Why not just interpolate displacements measured at cathode- and anode-crossing track endpoints? This is the real threat. It is simpler, uses the same data, and may capture most of the distortion near the walls. The difference lies in the volume interior and along track directions, where endpoints say nothing. If the interpolation baseline matches our method in the interior against the laser map, the potential-based factorization adds nothing and we will report that. The narrow-angle and free-displacement ablations show where each method's information actually comes from.

Why not just run the full differentiable detector simulator? It is the principled end-to-end route. It is also about 25 times slower than the ordinary simulator, needs memory workarounds, and has only been validated on synthetic closure tests fitting a few global parameters (Differentiable simulation of a LArTPC). Fitting a continuous 3D field through wire waveforms, while also inferring every unknown track, is the unconstrained inverse problem that work leaves open. We accept the cost of starting from reconstructed points. Our experiments bound that cost by checking whether errors in upstream point reconstruction bias the map in simulation.

Why not just train a network on simulation to predict corrections? The ProtoDUNE-SP disagreement is attributed partly to simulation underestimating space charge (Pandora ProtoDUNE-SP). A model trained on that simulation would inherit exactly that bias. Our fit uses no field-shape prior learned from simulation, only electrostatics.

Decisive Experiments

Simulated closure. In a simulated MicroBooNE-geometry detector, we inject known charge densities: a smooth analytic profile, a profile distorted by fluid flow, and a localized pocket of charge. We then report displacement-field error as a function of track count and angular coverage, and compare it with the identifiability analysis's prediction. A mismatch here would falsify our understanding of the method before any real-data claim.

Real data against the laser map. On MicroBooNE data, we fit cosmic-only maps and compare them with the laser map as described in Baselines. We also test whether a cosmic-derived map from one period predicts held-out cosmic tracks, and stopping-muon range versus charge, in a later period. This is the detector analog of relighting a reconstructed object under a new light.

Transfer to ProtoDUNE-SP. With no laser available, we judge success indirectly. We check whether applying our map moves test-beam end-point resolution below Pandora's reported 2 cm (68%) at 1 GeV/c. We also check whether it narrows the up-to-5% efficiency gap between data and simulation.

The test that decides the idea is the interior comparison against endpoint interpolation. If the constrained factorization does not beat it in laser-covered interior regions, the vision analogy was decorative. If it does, the claim holds.