Dual-Ray Analysis: Techniques & Applications

• Dual-ray analysis is a measurement strategy that combines two radiative channels to reduce ambiguities in radiographic interpretation.
• It spans modalities such as simultaneous petawatt laser X-ray/neutron generation, dual-view X-ray inspections, and dual-energy radiography for material discrimination.
• The approach improves detection accuracy by leveraging complementary information and advanced calibration and fusion techniques across channels.

Dual-ray analysis denotes a family of measurement and inference strategies in which two radiative channels, views, or energies are combined to improve radiographic interpretation. In the literature considered here, the term spans at least three technically distinct settings: simultaneous petawatt-laser generation of MeV X-rays and neutrons for dual radiography of dense materials (Cohen et al., 14 Apr 2026), dual-view X-ray inspection in which vertical and side images are jointly processed for prohibited-item detection (Tao et al., 2024), and dual-energy MeV transmission analysis for atomic-number discrimination in cargo radiography (Lalor et al., 2023). Across these settings, the common objective is to reduce the inferential incompleteness of a single measurement by introducing a second, correlated source of information, whether that second source is a different particle species, a second viewing geometry, or a second beam spectrum.

1. Modalities and formal definitions

In simultaneous neutron–X-ray radiography, the dual-ray aspect arises because high-brilliance secondary sources are generated simultaneously in a single shot using a single laser beam, enabling multiplexed probing of high-speed events (Cohen et al., 14 Apr 2026). In dual-view X-ray inspection, the duality is geometric: paired vertical and side X-ray images are used to emulate the way human inspectors rely on two distinct views when detecting prohibited items (Tao et al., 2024). In dual-energy radiography, the duality is spectral: two transmission measurements are acquired at different endpoint energies, typically denoted $T_1=T(E_1;Z,\lambda)$ and $T_2=T(E_2;Z,\lambda)$, in order to infer material properties such as $Z$ and thickness or areal density (Lalor et al., 2023).

The transmission formalism is explicit in the dual-energy setting. For a monoenergetic beam of energy $E$ passing through a uniform slab of thickness $t$ and atomic number $Z$,

$$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$

For a bremsstrahlung source with detector response,

$$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$

where $E_0$ is the endpoint energy and $\lambda=\rho\,t$ is the areal density (Lalor et al., 2023). In simultaneous neutron radiography, the relevant transmitted quantity for resonance transmission analysis is instead

$T_2=T(E_2;Z,\lambda)$0

so that neutron resonances yield sharp dips in transmission at specific energies (Cohen et al., 14 Apr 2026). In dual-view learning systems, the relation is encoded not by attenuation physics alone but by cross-view correspondence constraints such as

$T_2=T(E_2;Z,\lambda)$1

which link object locations and class identities across views (Tao et al., 2024).

A plausible implication is that “dual-ray analysis” is best understood not as a single instrument class but as a general inferential pattern: paired measurements are exploited to recover structure, composition, or semantics that remain ambiguous under a single acquisition channel.

2. Petawatt simultaneous X-ray and neutron radiography

The petawatt experiment reported quantitative measurements of photon spectra from $T_2=T(E_2;Z,\lambda)$2 to $T_2=T(E_2;Z,\lambda)$3 MeV and angular distributions in the petawatt interaction regime, using an ultra-intense $T_2=T(E_2;Z,\lambda)$4, ultra-short ($T_2=T(E_2;Z,\lambda)$5) laser pulse (Cohen et al., 14 Apr 2026). The laser parameters were $T_2=T(E_2;Z,\lambda)$6, energy on target $T_2=T(E_2;Z,\lambda)$7, central wavelength $T_2=T(E_2;Z,\lambda)$8, a focused spot $T_2=T(E_2;Z,\lambda)$9 using an $Z$0 OAP, and peak intensity

$Z$1

A series of solid foils tilted $Z$2 with respect to the laser axis was irradiated. The targets included low-$Z$3 SiN $Z$4, mid-$Z$5 Al $Z$6, and high-$Z$7 Au $Z$8 and W $Z$9. The neutron “catcher” converter was a $E$0 LiF disk placed $E$1 behind the primary target (Cohen et al., 14 Apr 2026).

The diagnostics were explicitly multiplexed. They comprised a $E$2 electron spectrometer covering MeV–$E$3, proton imaging with a stack of RCF films on target normal recording up to $E$4, an X-ray/$E$5 spectrometer at $E$6 covering $E$7–$E$8, X-ray imaging with a PerkinElmer 1621 amorphous-Si panel at $E$9, angular X-ray mapping with radiophotoluminescence dosimeters on a $t$0 array, and neutron activation using a SPAC stack of activation foils $t$1 immediately behind the LiF converter, followed by $t$2-ray counting with BEGe / HPGe detectors (Cohen et al., 14 Apr 2026).

The measured X-ray emission from a typical thick Au target $t$3 followed

$t$4

At $t$5, the conversion efficiency per steradian was $t$6 for Au$t$7, $t$8 for Al$t$9, and $Z$0 for SiN$Z$1. The forward photon yield integrated over $Z$2–$Z$3 was $Z$4–$Z$5. The X-ray source size, inferred from the PSF of the X-ray MTF, was $Z$6–$Z$7, shrinking slightly for thicker foils. The angular pattern exhibited two lobes along the laser axis $Z$8 and along the target surface $Z$9, consistent with surface-emission simulations (Cohen et al., 14 Apr 2026).

For neutrons, the total yield from Li activation was inferred from $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$0, implying $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$1, with $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$2 from $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$3. The energy-resolved fluence from $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$4 activation over $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$5–$$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$6 was

$$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$7

After moderation with a $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$8 HDPE moderator, the epithermal flux at $$T(E;Z,t)=\frac{I_{\rm transmitted}(E)}{I_{\rm incident}(E)}=\exp[-\mu(E,Z)\,t].$$9 from the source was $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$0, with energy resolution $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$1 above $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$2 (Cohen et al., 14 Apr 2026).

These measurements establish a concrete dual-radiography configuration in which the same shot provides MeV-photon structural information and neutron-based compositional sensitivity.

3. Measurement pipelines, calibration, and inverse processing

The petawatt study coupled broad angular and spectral characterization with explicit detector-response modeling. X-ray dose at $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$3 locations was recorded on RPL dosimeters and mapped over $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$4, while neutron angular sampling was obtained implicitly from the SPAC foils, with the forward spectrum corroborated by MCNP6 transport (Cohen et al., 14 Apr 2026). X-ray spectrometer response functions were computed via Geant4, and the spectra were unfolded with the CMA-Unfold algorithm to yield absolute photon distributions $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$5; the typical unfolding uncertainty near cut-off was $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$6. RPL dosimeter response was characterized up to kGy doses. Activation $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$7-efficiencies were calibrated with $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$8, $$T(E_0;Z,\lambda)=\frac{\int_0^\infty D(E)\,\phi(E;E_0)\,\exp[-\mu(E,Z)\,\lambda]\,dE}{\int_0^\infty D(E)\,\phi(E;E_0)\,dE},$$9, $E_0$0, $E_0$1, and $E_0$2 sources, with detector efficiencies extrapolated via Geant4. The reported uncertainties were $E_0$3–$E_0$4 for X-ray conversion efficiencies and $E_0$5–$E_0$6 for activation activities (Cohen et al., 14 Apr 2026).

In dual-view detection, the measurement pipeline is algorithmic rather than spectrometric. AENet consists of two parallel pipelines—Main View and Auxiliary View—with an expert-model mechanism for challenging categories and a final fusion step (Tao et al., 2024). The main-view pipeline takes input $E_0$7, extracts features $E_0$8, and predicts boxes and class scores with a detection head,

$E_0$9

Its loss is

$\lambda=\rho\,t$0

During training, hard-category “expert models” are learned by cropping each ground-truth box and training specialized detectors $\lambda=\rho\,t$1 on those patches (Tao et al., 2024).

The auxiliary-view pipeline takes $\lambda=\rho\,t$2, localizes salient regions,

$\lambda=\rho\,t$3

transfers class labels via one-to-one correspondence,

$\lambda=\rho\,t$4

and predicts auxiliary detections through

$\lambda=\rho\,t$5

Its loss is

$\lambda=\rho\,t$6

After merging overlapping predictions, the corresponding region in $\lambda=\rho\,t$7 is cropped using known epipolar shift $\lambda=\rho\,t$8 and refined through the expert $\lambda=\rho\,t$9, producing final outputs

$T_2=T(E_2;Z,\lambda)$00

A general cross-view fusion can also be written as

$T_2=T(E_2;Z,\lambda)$01

with an optional consistency term

$T_2=T(E_2;Z,\lambda)$02

giving the overall objective

$T_2=T(E_2;Z,\lambda)$03

The joint prediction function is

$T_2=T(E_2;Z,\lambda)$04

subject to cross-view correspondence constraints (Tao et al., 2024).

In dual-energy analysis, the inverse problem is analytically compact. Defining $T_2=T(E_2;Z,\lambda)$05 and $T_2=T(E_2;Z,\lambda)$06 for monoenergetic beams yields

$T_2=T(E_2;Z,\lambda)$07

and therefore

$T_2=T(E_2;Z,\lambda)$08

If the ratio function is invertible, $T_2=T(E_2;Z,\lambda)$09 follows from $T_2=T(E_2;Z,\lambda)$10, and then $T_2=T(E_2;Z,\lambda)$11. The same section also supplies the attenuation decomposition

$T_2=T(E_2;Z,\lambda)$12

with approximate scalings

$T_2=T(E_2;Z,\lambda)$13

This provides the physical basis for both inversion and its failure modes (Lalor et al., 2023).

4. Material identification, structure recovery, and semantic detection

The dual-radiography analysis in the petawatt study explicitly separates structural and compositional roles. For X-ray imaging, the reported spatial resolution was $T_2=T(E_2;Z,\lambda)$14, established using a grid IQI and an MTF at $T_2=T(E_2;Z,\lambda)$15, with source size $T_2=T(E_2;Z,\lambda)$16. For the epithermal neutron channel, $T_2=T(E_2;Z,\lambda)$17 implies capability to resolve $T_2=T(E_2;Z,\lambda)$18-wide resonances and thus isotopic discrimination. The stated dual-mode interpretation is direct: X-ray CT for density/structure, and NRTA for elemental/isotopic tagging (Cohen et al., 14 Apr 2026).

The material-identification example was a PHITS simulation of moderated neutrons through a layered sample consisting of $T_2=T(E_2;Z,\lambda)$19 concrete / $T_2=T(E_2;Z,\lambda)$20 Fe / $T_2=T(E_2;Z,\lambda)$21 $T_2=T(E_2;Z,\lambda)$22. The transmitted spectrum exhibited concrete absorption features associated with Al, Si, and O resonances, an Fe-56 notch near $T_2=T(E_2;Z,\lambda)$23, and a $T_2=T(E_2;Z,\lambda)$24 giant resonance at $T_2=T(E_2;Z,\lambda)$25 (Cohen et al., 14 Apr 2026). Within the constraints of the supplied evidence, this is the most explicit demonstration of dual-ray analysis as joint structural and elemental interrogation.

In dual-view X-ray inspection, the recovered variable is neither attenuation coefficient nor isotope signature but detection accuracy under occlusion and projection ambiguity. The LDXray dataset comprises $T_2=T(E_2;Z,\lambda)$26 paired vertical/side X-ray images, with $T_2=T(E_2;Z,\lambda)$27 annotated instances in 12 prohibited categories: MP, OL, PC1, PC2, LA, GL, TA, BL, CO, NL, UM, and CG. Each image has exactly 2 paired views, the average number of instances per image is $T_2=T(E_2;Z,\lambda)$28, and the image resolution is typically $T_2=T(E_2;Z,\lambda)$29 px, up to $T_2=T(E_2;Z,\lambda)$30 (Tao et al., 2024). The dual-view mechanism improved the challenging umbrella category from $T_2=T(E_2;Z,\lambda)$31 to $T_2=T(E_2;Z,\lambda)$32, an increase of $T_2=T(E_2;Z,\lambda)$33 in $T_2=T(E_2;Z,\lambda)$34, and the “Nonmetallic Lighter” category from $T_2=T(E_2;Z,\lambda)$35 to $T_2=T(E_2;Z,\lambda)$36, a gain of $T_2=T(E_2;Z,\lambda)$37 (Tao et al., 2024).

A concise comparison of the three dual-ray modalities is given below.

Modality	Paired information	Primary inferred quantity
Simultaneous PW radiography	MeV X-rays + neutrons	Density/structure; elemental/isotopic tagging
Dual-view X-ray inspection	Vertical + side images	Prohibited-item detection
Dual-energy radiography	Low + high endpoint energies	$T_2=T(E_2;Z,\lambda)$38 and thickness/areal density, subject to degeneracy

This suggests that dual-ray analysis is unified less by sensor hardware than by the use of complementary constraints to separate variables that are entangled in single-channel measurements.

5. Non-uniqueness, ambiguity, and limitations

The most explicit statement of a fundamental limitation appears in the dual-energy cargo literature. Two distinct pure materials $T_2=T(E_2;Z,\lambda)$39 and $T_2=T(E_2;Z,\lambda)$40 can yield the same dual-ray readings if

$T_2=T(E_2;Z,\lambda)$41

which implies the degeneracy condition

$T_2=T(E_2;Z,\lambda)$42

If the ratio function $T_2=T(E_2;Z,\lambda)$43 is non-monotonic, a single measured value can correspond to two different atomic numbers (Lalor et al., 2023).

The physical explanation given is competition between $T_2=T(E_2;Z,\lambda)$44-pair production in the high-energy beam and $T_2=T(E_2;Z,\lambda)$45-photoelectric absorption in the low-energy beam, which makes $T_2=T(E_2;Z,\lambda)$46 non-monotonic. The study states that this non-uniqueness is present even in systems with perfect resolution and zero statistical noise, and that currently deployed commercial radiographic systems are fundamentally incapable of distinguishing between high-$T_2=T(E_2;Z,\lambda)$47 nuclear materials and miscellaneous mid-$T_2=T(E_2;Z,\lambda)$48 cargo contents (Lalor et al., 2023).

These conclusions were validated through Monte Carlo transparency simulations using Geant4 with the QGSP_BIC physics list. The simulated system used 4 MeV, 6 MeV, and 10 MeV bremsstrahlung endpoint spectra generated by firing electrons at a tungsten radiator with copper/steel filtering, and a CdWO$T_2=T(E_2;Z,\lambda)$49 scintillator crystal $T_2=T(E_2;Z,\lambda)$50 in energy-integrating mode. The transmitted photon counts formed $T_2=T(E_2;Z,\lambda)$51 and $T_2=T(E_2;Z,\lambda)$52, and no additional noise model was imposed beyond Geant4’s statistical uncertainties. The simulated $T_2=T(E_2;Z,\lambda)$53 points lay exactly on the theoretical “$T_2=T(E_2;Z,\lambda)$54-curves,” and at low areal densities and high $T_2=T(E_2;Z,\lambda)$55, multiple $T_2=T(E_2;Z,\lambda)$56-lines intersected. An equivalent-line example showed that for the $T_2=T(E_2;Z,\lambda)$57 system at $T_2=T(E_2;Z,\lambda)$58, an iron slab $T_2=T(E_2;Z,\lambda)$59 is indistinguishable from a tungsten slab $T_2=T(E_2;Z,\lambda)$60 (Lalor et al., 2023).

By contrast, the limitations in dual-view detection are primarily category- and geometry-dependent rather than imposed by a closed-form physical degeneracy. The reported improvements are largest for challenging categories such as UM, CO, CG, and NL, which indicates that the second view is especially useful when single-view appearance is intrinsically weak or confounded (Tao et al., 2024). In the petawatt neutron–X-ray setting, the limiting factors are instead source yield, moderation losses, spectrometer unfolding uncertainty, and detector calibration uncertainty rather than an explicitly proven non-uniqueness theorem (Cohen et al., 14 Apr 2026).

6. Performance, scaling, and research directions

The dual-view inspection results show consistent gains across seven detection models. Reported $T_2=T(E_2;Z,\lambda)$61 values improved from $T_2=T(E_2;Z,\lambda)$62 to $T_2=T(E_2;Z,\lambda)$63 for Faster R-CNN, $T_2=T(E_2;Z,\lambda)$64 to $T_2=T(E_2;Z,\lambda)$65 for Cascade R-CNN, $T_2=T(E_2;Z,\lambda)$66 to $T_2=T(E_2;Z,\lambda)$67 for Sparse R-CNN, $T_2=T(E_2;Z,\lambda)$68 to $T_2=T(E_2;Z,\lambda)$69 for RetinaNet, $T_2=T(E_2;Z,\lambda)$70 to $T_2=T(E_2;Z,\lambda)$71 for CenterNet, $T_2=T(E_2;Z,\lambda)$72 to $T_2=T(E_2;Z,\lambda)$73 for RepPoints, and $T_2=T(E_2;Z,\lambda)$74 to $T_2=T(E_2;Z,\lambda)$75 for ATSS. The cross-model generalization study states that AENet yields consistent mAP improvements of $T_2=T(E_2;Z,\lambda)$76–$T_2=T(E_2;Z,\lambda)$77 across 7 leading detectors. In ablations, the best expert on cropped patches was ATSS with mAP $T_2=T(E_2;Z,\lambda)$78; saliency-based location approximation outperformed direct coordinate mapping, $T_2=T(E_2;Z,\lambda)$79 versus $T_2=T(E_2;Z,\lambda)$80; the best auxiliary-view detector was Cascade R-CNN with mAP $T_2=T(E_2;Z,\lambda)$81 and UM $T_2=T(E_2;Z,\lambda)$82; and an optimal confidence threshold of $T_2=T(E_2;Z,\lambda)$83 gave the highest mAP $T_2=T(E_2;Z,\lambda)$84 with balanced precision/recall (Tao et al., 2024).

In the petawatt study, the discussion section formulates explicit source scaling and technical avenues. The yield scalings are

$T_2=T(E_2;Z,\lambda)$85

Proposed routes to improved performance include low-density foams ahead of the foil to increase hot-electron conversion, yielding a factor of $T_2=T(E_2;Z,\lambda)$86–$T_2=T(E_2;Z,\lambda)$87; contrast improvement by double plasma mirrors to obtain a smaller X-ray spot and a harder spectrum; higher repetition rates of $T_2=T(E_2;Z,\lambda)$88–$T_2=T(E_2;Z,\lambda)$89; and micro-focus X-ray geometry with TOF-gated epithermal neutron imaging for sub-$T_2=T(E_2;Z,\lambda)$90 resolution and dynamic radiography (Cohen et al., 14 Apr 2026).

A plausible implication is that the three literatures point toward complementary futures rather than a single convergent apparatus. One direction pursues richer physical channels, as in simultaneous MeV X-ray and neutron beams; another pursues richer inferential architectures, as in cross-view expert-model refinement; and a third defines the boundary conditions of what two-channel transmission can and cannot recover, as in the non-monotonic dual-energy inverse map. Taken together, they delineate dual-ray analysis as a technically heterogeneous but conceptually coherent research area centered on paired measurements, paired models, and the controlled reduction of ambiguity.