Unfolding Recoil Spectra With the Light Pattern, Not Just the Light Total

When a neutron or a dark matter particle knocks a nucleus inside a scintillator (a material that flashes when struck), the flash at low energy is tiny. Standard analysis keeps only one number per event: the total light seen by a single photomultiplier, called the pulse height. It then mathematically "unfolds" the histogram of pulse heights into the true recoil-energy spectrum. That inversion fails near the detection threshold. One reason is that the same recoil energy produces different amounts of light depending on where in the cell it happens, so the detector's response is smeared. Below a few hundred keV of neutron energy, that smearing makes the inversion unstable.

We propose to read the scintillator with a segmented, coded photosensor array that collects essentially the same total light as a single tube but also records how the light is shared across channels. We then unfold the recoil spectrum with a forward model that includes interaction position as a hidden variable. In plain terms, we do not correct for light-collection variation after the fact. We put it inside the inversion, so the model knows which events were poorly collected and how uncertain their positions are. The claim to test is whether this lowers the energy at which the recoil spectrum can still be reconstructed faithfully. The test uses the standard ²⁵²Cf neutron reference, at matched photon budget, against the best published single-tube unfolding.

Core Claim

Recoil-spectrum unfolding is a linear inverse problem, $y = Ax$. Here $x$ is the true energy spectrum, $y$ is the measured pulse-height histogram, and $A$ is the detector response matrix, each column giving the pulse-height distribution produced by one true energy. The current reference method is High-fidelity unfolding of 252Cf neutron spectra in EJ-301 scintillators. It builds $A$ from Geant4 simulation, calibrates the energy broadening against ¹³⁷Cs and ²²Na Compton edges, and inverts with two standard iterative unfolding algorithms, MLEM and GRAVEL. It reaches a 0.35 MeV threshold, an average deviation below 4.28%, and a best agreement of 1.42% (2-inch cell, MLEM) relative to the ISO 8529-1 reference. Prior work stalled at 1–2 MeV. That paper reports that going from a 1-inch to a 2-inch cell improved resolution but did not move the threshold. This is our motivating evidence that collecting more light alone is not the remedy.

Our claim concerns how the response is represented. Light-collection nonuniformity is one of the effects that broadens the columns of $A$, and it is a systematic, not purely random, effect. If each event also yields a coarse light-sharing pattern $c$ across $M$ channels, the measurement becomes a joint histogram $y(h, c)$, where $h$ is pulse height. The response then lifts to

$$ y(h,c) = \sum_{E}\sum_{r} A(h, c \mid E, r)\, p(r)\, x(E) + n , $$

where $r$ indexes interaction-position cells, $p(r)$ is the known position distribution set by irradiation geometry, and $n$ is noise. The lifted matrix has narrower blocks, since each block is the response at a fixed position. Its columns for neighbouring energies should therefore be less correlated than the position-averaged columns of the single-tube $A$. Better separation between neighbouring columns is exactly what governs stability near threshold.

The coding is where resolution-throughput decoupling enters. Optical work by Cull et al. and Fernández et al. established that in coded-aperture spectrometers, the size of an individual feature sets resolution while the total open area sets throughput. Here the "feature" is a position cell and the "open area" is total photosensor coverage. A coded light-sharing guide can map $N$ position cells onto $M < N$ readout channels without blocking any photons.

What Is New, and What Is Not

None of the work in the research file combines the three ingredients. The closest work is the Zhang et al. unfolding itself, which is purely zero-dimensional: one pulse height per event, with no spatial information.

The coded-optics line is a different setting. It includes coded dispersive spectrometers, the single-snapshot coded spectral imager of Wagadarikar et al., and the lensless Spectral DiffuserCam. All of these supply the encoding principle we reuse, but they work on bright, static optical scenes with many photons per pixel.

The one coded-aperture dark matter instrument in the file is GECCO. It masks incoming gamma rays to obtain 0.5-arcminute angular resolution and never encodes light produced inside a target.

Liquid xenon time projection chambers such as PandaX-4T already recover event position, using drift time and the ionization signal. This is supporting evidence that position knowledge matters at sub-keV energies. A single-volume scintillator has no such channel, which is the gap we fill.

We deliberately do not propose placing a blocking mask inside a low-energy detector. At a few detected photons per event, discarding even half the light would raise the threshold. The design must be photon-conserving: coding by light sharing, not by occlusion.

Research Program

The program has three parts, in increasing cost.

The first is a simulation study in Geant4. We model an EJ-301 cell read by a pixelated silicon photomultiplier array, both with and without a coded light-guide that multiplexes position cells into fewer channels. The simulation includes optical photon transport and proton quenching. From it we build the single-tube response $A$ and the lifted response $A(h,c \mid E,r)$. We then measure:

  • how strongly neighbouring energy columns overlap at low energy, in both representations;
  • the condition number of each response matrix, as a function of energy cut and photon budget.

This tells us, before building anything, whether lifting helps at the photon counts actually present near 0.35 MeV.

The second part is a hardware prototype. We build a 2-inch EJ-301 cell matching the Zhang et al. geometry, read by a segmented array. The array can be operated in three modes:

  • fully pixelated;
  • coded-multiplexed;
  • summed, where all channels are added in software to emulate a single tube with identical photon detection efficiency.

We calibrate the lifted response per position cell using collimated ¹³⁷Cs and ²²Na Compton-edge scans. This extends the Zhang et al. calibration procedure from one broadening parameter set to a position-indexed set.

The third part is inversion. We use MLEM, extended to the lifted model by treating position as a latent variable, together with a nonnegative penalized least-squares solver as a cross-check. The lifted operator has block structure: $A^T A$ is a sum of per-position blocks. Its spectral norm, which sets the solver step size, is cheap to compute here because the energy dimension is only hundreds of bins. So the Lipschitz-constant machinery of Holst and Overgaard is not needed. We note this so that nobody expects acceleration to be part of the contribution.

Baselines

The naive reference is the summed mode of our own hardware: the same cell and the same sensors, with the channel pattern thrown away, unfolded with standard MLEM and GRAVEL on a single-tube response. This is the cleanest ablation, because photon budget, scintillator, sensor efficiency, source exposure and calibration time are identical. Only the use of the light pattern differs.

The strongest existing method is the Zhang et al. pipeline. The comparison numbers are a 0.35 MeV threshold, an average deviation below 4.28%, and a best agreement of 1.42% for the 2-inch cell with MLEM, all against ISO 8529-1. We will reproduce it on a conventional photomultiplier-read 2-inch cell. That reproduction tests whether our sensor change alone moves the numbers, and anchors us to the published figures.

The key ablation separates joint unfolding from correct-then-unfold. In correct-then-unfold, each event's pulse height is first divided by the light-collection efficiency at its estimated position, and the corrected histogram is then unfolded in one dimension.

A second ablation compares fully pixelated with coded-multiplexed readout at fixed channel count. It tests whether the coding, rather than segmentation per se, earns its place.

Every method gets the same data, the same Geant4 response-generation budget, and the same tuning budget, which covers iteration count or regularization strength. Tuning is done on a held-out simulated spectrum, never on the ISO reference.

We count the proposal as winning if the lifted joint unfolding keeps average deviation from ISO 8529-1 at or below 4.28% down to an energy measurably below 0.35 MeV. The shift must exceed the bin-to-bin spread seen across repeated bootstrap resamplings. It must also beat summed mode and correct-then-unfold at the same threshold.

Why not just…?

Why not just use a bigger cell or a more efficient photomultiplier? The Zhang et al. comparison is direct evidence against relying on this. Their 2-inch cell outperformed the 1-inch cell in resolution and efficiency, yet the reported threshold remained 0.35 MeV. Our summed-mode ablation also controls for photon budget, so any gain we report cannot be attributed to collecting more light.

Why not just regularize the existing one-dimensional unfolding better? Better priors, Tikhonov smoothing or a stopping rule tuned per energy are real competitors. Tikhonov smoothing is not demonstrated in the file. Software improvements of this kind could move the threshold without new hardware. We will therefore give the summed-mode baseline the full regularization menu under the same tuning budget. If a well-regularized one-dimensional unfolding matches lifted unfolding, then the light pattern adds nothing, and we will report that.

Why not just correct each event for position and then unfold? This is the most serious threat, and it is how time projection chambers handle position dependence. At high energy, with many photons, it should work almost as well as joint unfolding. Our hypothesis is that it degrades near threshold. With few photons, the position estimate is noisy, and dividing by a low collection efficiency inflates noise in exactly the bins that set the threshold. Joint unfolding propagates that uncertainty instead of committing to a point estimate. The correct-then-unfold ablation settles this directly, using deviation versus energy in the bottom bins. If the two match within errors, the contribution reduces to the readout hardware plus a known correction, and we will say so.

Why not just use a plain pixelated array instead of coding? If channel count is not a constraint, plain pixelation may be equally good. The argument for coding is the resolution-throughput principle applied to channels: in larger or cryogenic detectors, each readout channel costs cables, heat load and noise. The pixelated-versus-coded comparison at matched channel count tells us whether coding preserves the conditioning gain. If it does not, the transferable lesson is just "segment the readout."

Proposed Experiments and Decisive Checks

The first decisive check is in simulation and is cheap. At the photon counts expected near 0.35 MeV, does the lifted response reduce the overlap between neighbouring energy columns relative to the single-tube response? If position cells cannot be distinguished at those photon counts, the program stops here, because the hardware cannot help.

The second check is a toy unfolding. We fold known spectra, including a ²⁵²Cf-shaped spectrum and spectra with injected low-energy features, through each response. We then compare recovered spectra by bias and spread over many pseudo-experiments, across all baselines.

The third and main check is the measured ²⁵²Cf spectrum, compared with ISO 8529-1 bin by bin. The headline metric is the lowest energy at which average deviation stays below the 4.28% level reported by Zhang et al. Secondary metrics are the best-bin agreement, to compare with their 1.42%, and stability of the result under ±10% perturbation of the calibration. The perturbation test addresses calibration drift, the known weakness of computational methods with calibrated response matrices. A lifted response has more parameters to drift, so robustness under this test is a condition of success, not an afterthought.

For low-energy recoil searches more broadly, the result is a methodological test, not a physics reach. The file reports no result on Lacuna for sub-keV spectrum unfolding in single-volume scintillators. We therefore make no claim about dark matter limits such as the PandaX-4T spin-independent bound of 1.1 × 10⁻⁴³ cm² at 3 GeV/c². A clear positive result at the MeV scale would justify a simulation study of photon-conserving coded readout for small low-threshold scintillating targets. A clear negative result would tell the field that position-aware unfolding is not where the threshold is being lost.