RadMap Telescope: Cosmic-Ray Spectroscopy

• RadMap Telescope is a compact, low-power spaceborne instrument that reconstructs cosmic-ray nuclei events to determine charge, energy per nucleon, and arrival direction.
• It uses a scintillating-fiber tracking calorimeter with advanced neural network reconstructions, achieving near geometric limits in angular resolution.
• The evolution from a mobile terrestrial mapping platform to a space-based spectroscopic sensor underscores its role in precise radiation-field characterization for dosimetric applications.

Searching arXiv for RadMap-related papers to support the article. RadMap Telescope is a compact, low-power radiation telescope developed for spectroscopic space-radiation monitoring, with the explicit aim of reconstructing cosmic-ray nuclei event-by-event in terms of charge $Z$, energy per nucleon, and arrival direction in the MeV–GeV/n regime (Meyer-Hetling et al., 18 Aug 2025). The same name, “RadMAP,” also appears in an earlier and distinct terrestrial context: the Radiological Multi-sensor Analysis Platform, a mobile radiation-background characterization system built on a 20-ft box truck and used to measure fast neutron backgrounds in the San Francisco Bay Area (Davis et al., 2016). The overlap in nomenclature has generated an obvious potential confusion. In the 2016 study, RadMAP is not an astronomical telescope but a mobile multi-sensor survey platform for neutron-background characterization; in the 2025 study, RadMap Telescope denotes a scintillating-fiber tracking calorimeter intended for operational space-radiation measurements. Taken together, these works show a transition from mobile environmental neutron-background sensing to compact event-resolved cosmic-ray spectroscopy, while preserving a common emphasis on radiation-field characterization under realistic operational constraints (Davis et al., 2016, Meyer-Hetling et al., 18 Aug 2025).

1. Nomenclature and historical context

The term “RadMAP” originally referred to the Radiological Multi-sensor Analysis Platform, a mobile, multi-sensor radiation detection and background characterization system with onboard power, built on a 20-ft box truck (Davis et al., 2016). It originated as the Naval Research Laboratory’s MISTI gamma-ray platform and was transferred to Lawrence Berkeley National Laboratory in late 2011, where it was expanded to include fast-neutron detection and environmental and geospatial sensing for mobile background studies (Davis et al., 2016). In that setting, the platform’s role was to characterize fast neutron backgrounds and their dependence on atmospheric and urban structural conditions rather than to serve as a telescope in the astronomical sense (Davis et al., 2016).

The later RadMap Telescope is a different instrument class. It is described as a compact, low-power radiation telescope developed to bring spectroscopic capability into operational space-radiation monitoring, specifically event-by-event recovery of species, energy per nucleon, and arrival direction (Meyer-Hetling et al., 18 Aug 2025). This instrument has already collected data inside the International Space Station from April 2023 to January 2024, although the cited feasibility study is based on Geant4-simulated data and neural-network reconstruction rather than on a full end-to-end analysis of flight measurements (Meyer-Hetling et al., 18 Aug 2025).

This naming continuity suggests a broader programmatic lineage centered on radiation-field mapping. A plausible implication is that “RadMap” evolved from denoting a mobile terrestrial mapping platform to denoting a compact spaceborne spectroscopic sensor, but the two instruments should not be conflated: their detector architectures, operational environments, and measurement objectives are substantially different (Davis et al., 2016, Meyer-Hetling et al., 18 Aug 2025).

2. Instrument architectures

The terrestrial RadMAP platform employed 16 EJ-309 organic liquid scintillator cells supplied by Sandia National Laboratories for fast-neutron detection (Davis et al., 2016). Each cell was a 5-inch diameter by 5-inch long aluminum cylinder, with a total active volume of approximately 25 L across the array (Davis et al., 2016). The cells were oriented horizontally and stacked vertically in two columns of eight within the truck cargo space (Davis et al., 2016). EJ-309 was selected for pulse-shape discrimination capability, enabling neutron/gamma separation, and the experiment used the tail-to-total method for all PSD calculations (Davis et al., 2016). Seven detectors were coupled to 5-inch Hamamatsu photomultipliers and nine to 5-inch Photonis photomultipliers (Davis et al., 2016). Data acquisition used two Struck SIS3320 digitizers, nominally 250 MHz and 12-bit, operated at 200 MHz, corresponding to 5 ns per sample (Davis et al., 2016).

That system also integrated non-radiological sensors. Positioning used a NovAtel SPAN GNSS/INS installed in January 2012, with centimeter-level accuracy and a 100 Hz data rate, designed to remain robust under intermittent satellite reception (Davis et al., 2016). Weather measurements were obtained from a Davis Vantage Vue Wireless Weather Station recording atmospheric pressure, temperature, and absolute humidity (Davis et al., 2016). Each identified neutron event was associated with GPS coordinates and the contemporaneous weather metrics (Davis et al., 2016).

By contrast, the RadMap Telescope’s Active Detection Unit is a scintillating-fiber tracking calorimeter comprising 1024 plastic scintillating fibers (Meyer-Hetling et al., 18 Aug 2025). Each fiber has a square cross-section of $2\times2$ mm and a length of 80 mm (Meyer-Hetling et al., 18 Aug 2025). The fibers are arranged in 32 layers of alternating orientation, forming an imaging tracking calorimeter whose signals are projected into two 2D views, $yx$ and $yz$, each represented as a $16\times32$ pixel image (Meyer-Hetling et al., 18 Aug 2025). The stack depth in the $y$ direction is $32\times2$ mm, giving 64 mm of active plastic scintillator (Meyer-Hetling et al., 18 Aug 2025). Each fiber is read out by a SiPM on one end, and the instrument records per-fiber light intensities that, to leading order, scale with energy lost in the fiber (Meyer-Hetling et al., 18 Aug 2025).

The two architectures reflect different physical measurement problems. The truck-based system optimized neutron-background surveying with environmental co-registration and PSD-capable bulk scintillators (Davis et al., 2016). The space instrument optimized event imaging and spectroscopic reconstruction of charged cosmic-ray nuclei with fine segmentation and per-fiber readout (Meyer-Hetling et al., 18 Aug 2025).

3. Measurement principles and observables

In the 2016 RadMAP platform study, the measured quantity of interest was the fast-neutron background at ground level in the 500 keV to 8 MeV range (Davis et al., 2016). Event selection relied on tail-to-total PSD in EJ-309 to distinguish neutron interactions from gamma-ray events, and neutron events passing PSD classification within the analysis window were retained (Davis et al., 2016). The physical interpretation of the measured neutrons was explicitly tied to cosmic-ray secondaries: below the Pfotzer maximum, fast neutrons at ground level are largely products of cosmic-ray showers whose flux attenuates exponentially with atmospheric depth (Davis et al., 2016).

In the spaceborne RadMap Telescope, the principal observables are track images and longitudinal energy-deposition patterns across the fiber stack (Meyer-Hetling et al., 18 Aug 2025). Tracks are parameterized by spherical angles $\phi\in[-180^\circ,180^\circ)$ and $\theta\in[0^\circ,180^\circ]$ referenced to detector axes (Meyer-Hetling et al., 18 Aug 2025). For stopping particles, the measured Bragg curve encodes $Z$, $2\times2$0, and hence $2\times2$1; for through-going particles, the longitudinal energy-loss profile still carries information on $2\times2$2 and $2\times2$3 (Meyer-Hetling et al., 18 Aug 2025). The visible response departs from proportionality to $2\times2$4 at high ionization density because of quenching, which is modeled in Geant4 and described as a saturation-like compression of high-$2\times2$5 signals (Meyer-Hetling et al., 18 Aug 2025).

Three complicating effects are emphasized for the telescope: energy-loss straggling, nuclear fragmentation, and ionization quenching (Meyer-Hetling et al., 18 Aug 2025). Straggling introduces substantial event-by-event fluctuations relative to mean Bethe–Bloch behavior. Fragmentation generates lower-$2\times2$6 secondaries and hadronic debris and becomes more probable with increasing $2\times2$7 and traversed material. Quenching reduces separation power at high $2\times2$8 because stopping power scales approximately as $2\times2$9 at fixed $yx$0, while the visible signal is compressed (Meyer-Hetling et al., 18 Aug 2025). These are intrinsic obstacles to species and energy reconstruction in thin segmented scintillator systems.

The contrast is therefore between environmental background metrology and per-particle spectroscopy. The ground system measured aggregate neutron count-rate modulation by environmental variables (Davis et al., 2016). The space instrument reconstructs individual charged-particle properties from detector images and energy-loss topology (Meyer-Hetling et al., 18 Aug 2025).

4. Environmental neutron-background characterization in the mobile RadMAP platform

The mobile RadMAP study assembled 37 usable runs with the scintillators between May 2012 and December 2013 across the San Francisco Bay Area, including urban cores, bridges, tunnels, and rural or sparsely built areas spanning sea level to above 3800 ft elevation (Davis et al., 2016). In 2014, while the truck was stationary for maintenance, more than 100 long stationary datasets of 12–15 hours each were collected at LBNL Building 88, providing high-statistics data for weather and geomagnetic analyses (Davis et al., 2016).

The principal empirical result was the expected exponential dependence of fast-neutron count rate on atmospheric pressure (Davis et al., 2016). Combining mobile and stationary data, the count rate $yx$1 as a function of pressure $yx$2 in mbar was fit as

$yx$3

with fit parameter uncertainties $yx$4 and $yx$5 (Davis et al., 2016). Over the measured pressure range, the rate decreased by about 32%; explicitly, the paper notes about 32% suppression from 970 mbar to 1030 mbar, with rates ranging from roughly 3 CPS at low pressure to about 2.1 CPS at high pressure (Davis et al., 2016).

This dependence was then used for event-level normalization to standard atmospheric pressure, 1013.25 mbar (Davis et al., 2016). Each detected neutron at pressure $yx$6 was assigned the weight

$yx$7

After this correction, the rate-versus-pressure distribution became flat, with a best-fit near the sea-level mean, $yx$8 CPS (Davis et al., 2016). The linear fit to the pressure-adjusted histogram was

$yx$9

which was statistically consistent with no residual pressure dependence (Davis et al., 2016).

The operational value of this normalization was quantified through 60 s count-rate distributions across all runs (Davis et al., 2016). The systematic width, defined as the RMS after subtracting Poisson statistics in quadrature, decreased from 0.113 CPS in unadjusted data to 0.078 CPS after pressure adjustment, corresponding to a 31% reduction in systematic uncertainty (Davis et al., 2016). The paper explicitly connects this reduction to improved background predictability and improved source detectability (Davis et al., 2016).

Residual environmental dependencies were weaker. After pressure normalization, no significant residual correlation was found with temperature in the 500 keV–8 MeV fast-neutron band; the reported linear fit was

$yz$0

(Davis et al., 2016). Absolute humidity showed a weak residual positive correlation,

$yz$1

although the authors cautioned that more data and analysis were needed and noted that any slight increase could reflect hydrogen-rich air downscattering high-energy neutrons into the detectable band (Davis et al., 2016). Increased geomagnetic activity, parameterized by Kp, produced a small suppression after pressure adjustment, with the linear fit implying about 2.3% suppression at $yz$2; this was judged insufficient for routine correction but relevant during major Forbush decreases or ground level enhancements (Davis et al., 2016).

A further consistency check related altitude-dependent rates to Pfotzer’s atmospheric depth model using an absorption length of $yz$3 and found good agreement when pinning the predicted curve to the measured sea-level mean of about 2.2 CPS (Davis et al., 2016). This supports the interpretation that the dominant modulation is the changing atmospheric overburden experienced by cosmic-ray secondaries.

5. Urban shielding, sky-view factor, and structural suppression

A distinctive feature of the mobile RadMAP analysis was the use of sky-view factor (SVF) to quantify the degree of urban structural shielding (Davis et al., 2016). SVF was defined as the fraction of unobstructed sky visible from a point out of $yz$4 steradians (Davis et al., 2016). In urban environments, lower SVF corresponds to increased shielding by surrounding buildings and overhead structures, which suppresses the background of cosmic-ray secondaries, including neutrons (Davis et al., 2016).

To estimate SVF along the truck route, the analysis used NOAA/USGS coastal lidar data providing latitude, longitude, and elevation per point (Davis et al., 2016). RadMAP positions were discretized in 3 m bins along the route (Davis et al., 2016). A simple 2D open-sky angle in the transverse plane was found inadequate in dense urban intersections because it could remain $yz$5 in 2D even when overhead structures substantially reduced the true visible sky (Davis et al., 2016). The adopted method followed Zakšek et al.: for $yz$6 angular slices around the point, one finds the highest obstructed elevation angle $yz$7 above the local horizon within a fixed radius $yz$8, and computes

$yz$9

The implementation used $16\times32$0 slices at $16\times32$1 intervals and a search radius $16\times32$2 m around the truck (Davis et al., 2016). Elevation angles were obtained by subtracting each slice’s open-sky angle from $16\times32$3, with slice orientations set by truck bearing and all lidar points at or above truck elevation within each slice included (Davis et al., 2016).

The method had explicit limitations. Structures beyond 35 m could still occlude the sky, producing SVF overestimation in dense high-rise environments; the paper notes likely overestimation in San Francisco because of tall buildings outside the 35 m radius (Davis et al., 2016). There was also temporal mismatch between lidar acquisition and neutron acquisition. At one San Francisco location, lidar indicated low SVF from dense buildings that had actually been demolished before the RadMAP run, yielding an anomalously high neutron rate in the lowest SVF bin, 0.10–0.15 (Davis et al., 2016). For that reason, the authors excluded $16\times32$4 from fits and recommended onboard upward-looking lidar to remove the mismatch (Davis et al., 2016).

Despite these limitations, the observed suppression was strong. Pressure-adjusted neutron rates versus SVF in Berkeley, Downtown Oakland, and Downtown San Francisco all showed consistent suppression at low SVF, with differences among areas plausibly linked to building density (Davis et al., 2016). A combined quadratic fit, excluding $16\times32$5, gave

$16\times32$6

where $16\times32$7 is count rate in CPS and $16\times32$8 is SVF (Davis et al., 2016). Between $16\times32$9 and $y$0, the pressure-adjusted rate dropped from 2.25 CPS to 0.95 CPS, corresponding to 58% suppression (Davis et al., 2016). Each of the three urban areas exhibited more than 50% suppression across the measured SVF range (Davis et al., 2016).

This result is physically significant because it indicates that structural shielding dominates over any additional neutron production in building materials via spallation (Davis et al., 2016). It also provides an operational lesson: in urban neutron-background modeling, structural context can modulate rates at least as strongly as meteorological variables, and in the measured Bay Area datasets the SVF effect exceeded the pressure effect, which alone produced about 32% suppression across 970–1030 mbar (Davis et al., 2016).

6. Neural-network reconstruction in the RadMap Telescope

The 2025 feasibility study formulated the RadMap Telescope reconstruction problem as direct inference from detector images (Meyer-Hetling et al., 18 Aug 2025). The two raw grayscale projections, each $y$1, were used as network inputs without explicit hit clustering, segment finding, or handcrafted $y$2 features (Meyer-Hetling et al., 18 Aug 2025). The approach therefore relied on CNNs to learn latent representations of track geometry and ionization signatures from the images themselves.

Track-angle reconstruction used one CNN with inception-style feature extraction to learn $y$3 from the $y$4 view and $y$5 from the $y$6 view, with the true polar angle recovered through

$y$7

The task was treated as dual classification with 0.2° bins, giving 1800 classes for $y$8 and 900 for $y$9 (Meyer-Hetling et al., 18 Aug 2025). Outputs were pseudo-probability distributions over angle classes, and the highest-probability class was selected (Meyer-Hetling et al., 18 Aug 2025). Training used AdamW and early stopping, and the track network had approximately 2.8 million trainable parameters (Meyer-Hetling et al., 18 Aug 2025). Training employed 0.7 million events with 0.2 million validation events per epoch (Meyer-Hetling et al., 18 Aug 2025).

Charge reconstruction used two consecutive CNNs, each with multiple flat inception layers and about 2.1 million parameters (Meyer-Hetling et al., 18 Aug 2025). The first, or low-$32\times2$0 network, classified $32\times2$1–8 with an overflow bin for charges above 8. The second, or high-$32\times2$2 network, classified $32\times2$3–26 and was trained with $32\times2$4 included to mitigate boundary effects (Meyer-Hetling et al., 18 Aug 2025). Training used 9 million events and testing used 1 million events, with energies spanning 20 MeV to 5 TeV (Meyer-Hetling et al., 18 Aug 2025).

Energy-per-nucleon reconstruction used 26 element-specific CNN regressors returning continuous $32\times2$5 outputs (Meyer-Hetling et al., 18 Aug 2025). For $32\times2$6, each network was trained on mixed charges within either $32\times2$7 or $32\times2$8 to improve robustness against charge confusion (Meyer-Hetling et al., 18 Aug 2025). Typical energy ranges were 20 MeV/n to 1 GeV/n, extended to 10 GeV/n for the heaviest elements (Meyer-Hetling et al., 18 Aug 2025). Each element-specific network was trained on about 1.8 million events, and evaluation used a combined 8 million-event dataset (Meyer-Hetling et al., 18 Aug 2025). A branched variant added a filter network routing events to stopping versus through-going energy regressors (Meyer-Hetling et al., 18 Aug 2025).

The simulation and dataset generation were deliberately simplified to expose intrinsic detector capability (Meyer-Hetling et al., 18 Aug 2025). The detector model contained only the 1024-fiber stack and omitted surrounding support, housing, electronics, and ISS structure, thus removing external scattering and fragmentation (Meyer-Hetling et al., 18 Aug 2025). Ionization quenching was included, but optical and electrical cross-talk, SiPM saturation, channel-to-channel gain variations, and fiber misalignment were not (Meyer-Hetling et al., 18 Aug 2025). The source model used the most abundant isotopes from hydrogen to iron, excluded electrons and gamma rays, imposed equal elemental abundances to avoid severe class imbalance, and sampled energies from log-uniform distributions over task-specific ranges (Meyer-Hetling et al., 18 Aug 2025). Unless otherwise stated, events had to contain at least three hit fibers in each projection, while stricter criteria were used for angle training and benchmarking (Meyer-Hetling et al., 18 Aug 2025).

7. Performance, dosimetric significance, and unresolved limitations

The feasibility study reported angular resolutions defined through the Gaussian $32\times2$9 of the direction-independent residual distributions, using the central 68% interval and a correction for the intrinsic direction ambiguity of near-MIP cases (Meyer-Hetling et al., 18 Aug 2025). For protons, the results were $\phi\in[-180^\circ,180^\circ)$0 and $\phi\in[-180^\circ,180^\circ)$1 for MIP 3 GeV events; approximately $\phi\in[-180^\circ,180^\circ)$2 for 120 MeV monoenergetic events in both angles; and $\phi\in[-180^\circ,180^\circ)$3 and $\phi\in[-180^\circ,180^\circ)$4 respectively for stopping events (Meyer-Hetling et al., 18 Aug 2025). Iron showed weak energy dependence with $\phi\in[-180^\circ,180^\circ)$5 and $\phi\in[-180^\circ,180^\circ)$6 (Meyer-Hetling et al., 18 Aug 2025). Biases were negligible, with $\phi\in[-180^\circ,180^\circ)$7 smaller than the 0.2° bin width (Meyer-Hetling et al., 18 Aug 2025). The reported resolutions were described as close to the geometric limit set by the effective $\phi\in[-180^\circ,180^\circ)$8 mm pixel size (Meyer-Hetling et al., 18 Aug 2025).

Charge separation performance was highly species dependent (Meyer-Hetling et al., 18 Aug 2025). Exact-charge assignment across all $\phi\in[-180^\circ,180^\circ)$9 was 59% overall (Meyer-Hetling et al., 18 Aug 2025). For hydrogen, the classifier achieved 99.8% accuracy and 99.6% purity; for helium, 99.3% accuracy and 98.8% purity (Meyer-Hetling et al., 18 Aug 2025). Light nuclei with $\theta\in[0^\circ,180^\circ]$0 were reconstructed with accuracy well over 95% and purity at least 84% (Meyer-Hetling et al., 18 Aug 2025). Exact-charge accuracy for the largest $\theta\in[0^\circ,180^\circ]$1 values fell to 30–40%, but misassignments typically went to neighboring charges. When one allowed $\theta\in[0^\circ,180^\circ]$2, heavy-element accuracy remained at least 70%; allowing $\theta\in[0^\circ,180^\circ]$3 increased it to at least 83%, with mean purity rising to 83% and 91% respectively (Meyer-Hetling et al., 18 Aug 2025). The performance discontinuity at the $\theta\in[0^\circ,180^\circ]$4 handover was attributed to the intrinsic difficulty of high-$\theta\in[0^\circ,180^\circ]$5 separation under quenching, straggling, and fragmentation, as well as inter-network boundary effects (Meyer-Hetling et al., 18 Aug 2025).

Energy resolution was reported as $\theta\in[0^\circ,180^\circ]$6 in logarithmic bins (Meyer-Hetling et al., 18 Aug 2025). Hydrogen achieved $\theta\in[0^\circ,180^\circ]$7 below 100 MeV, $\theta\in[0^\circ,180^\circ]$8 below 300 MeV, and $\theta\in[0^\circ,180^\circ]$9 below 1 GeV (Meyer-Hetling et al., 18 Aug 2025). Helium achieved $Z$0 below 300 MeV/n and $Z$1 below 800 MeV/n (Meyer-Hetling et al., 18 Aug 2025). Carbon showed realistic energy resolution in the 10–25% range, though low-purity subsets affected by charge confusion could show a local maximum of about 34% near 100 MeV/n, while the ideal pure-charge limit approached about 10% (Meyer-Hetling et al., 18 Aug 2025). Iron achieved $Z$2 near 150 MeV/n, $Z$3 below 400 MeV/n, and $Z$4 below 2 GeV/n (Meyer-Hetling et al., 18 Aug 2025). The global claim was that the framework achieves less than 20% energy resolution for energies below 1 GeV/n up to iron, with substantially better performance for H and He across the most dosimetrically relevant energies (Meyer-Hetling et al., 18 Aug 2025).

These performance levels matter because the telescope is intended to supply the spectroscopic inputs needed for absorbed dose and dose-equivalent estimation (Meyer-Hetling et al., 18 Aug 2025). The paper gives two standard formalisms. In an energy-deposition view,

$Z$5

In an LET-based view,

$Z$6

with one example quality factor relation given as the ICRP-60 piecewise form (Meyer-Hetling et al., 18 Aug 2025). The study does not implement a full dosimetric pipeline, but it argues that the demonstrated reconstruction of $Z$7 is sufficient to feed standard NASA and ICRP workflows (Meyer-Hetling et al., 18 Aug 2025).

Several limitations are explicit and materially important. The simulated geometry excluded housing, electronics, and ISS structure, so shielding-dependent fragmentation, scattering, and anisotropic acceptance were not represented (Meyer-Hetling et al., 18 Aug 2025). Detector nonidealities such as cross-talk, SiPM saturation and gain variations, fiber-placement tolerances, and nonuniform light yield were also omitted and are expected to degrade the quoted best-case resolutions (Meyer-Hetling et al., 18 Aug 2025). High-$Z$8 separation remains intrinsically difficult because quenching compresses high $Z$9, while straggling and fragmentation broaden class overlap (Meyer-Hetling et al., 18 Aug 2025). Events with few hit fibers are also intrinsically harder, so selection cuts improve performance at the cost of sensitivity to low-energy or corner-crossing tracks (Meyer-Hetling et al., 18 Aug 2025). Finally, the training distributions used equal elemental abundances and log-uniform spectra to avoid class imbalance, whereas real GCR abundances and energy spectra differ markedly; future work is therefore directed toward realistic priors, domain adaptation, physics-informed training, and potentially more expressive architectures such as transformers or GNNs (Meyer-Hetling et al., 18 Aug 2025).

The resulting picture is technically coherent. The older RadMAP platform established a methodology for high-fidelity environmental radiation characterization through event-level contextualization, normalization, and structural modeling (Davis et al., 2016). The newer RadMap Telescope extends the name into space-radiation instrumentation, where the emphasis shifts from environmental modulation of aggregate count rates to event-level reconstruction of particle identity, kinematics, and dosimetric relevance (Meyer-Hetling et al., 18 Aug 2025). The shared conceptual core is radiation mapping under operational conditions, but the instruments themselves belong to different detector lineages and should be distinguished accordingly.