Limited-angle X-ray CT can help address a difficult measurement problem in industrial and experimental systems: visualizing multiphase flow inside an opaque metal pipe. Conventional optical techniques struggle when light refracts at gas-liquid interfaces, while conventional CT can blur rapidly moving structures because its source and detector must rotate mechanically. Electron beam X-ray tomography (EBXT) avoids that mechanical rotation, but its fixed geometry can leave projection data incomplete. Reconstruction methods must therefore contend with both limited-angle artifacts and the complex behavior of energy-integrating detectors.
Why multiphase flow imaging is challenging
In EBXT, electromagnetic fields steer an electron beam to generate X-rays at successive positions along a target. A fixed detector array records the resulting projections. Removing mechanical rotation can support faster scanning of moving objects, making the approach relevant to time-resolved flow measurement. However, a fixed source-detector arrangement may provide only limited angular coverage. Missing projection directions make the reconstruction problem ill-posed: different internal structures can produce data that are difficult to distinguish.
Flow structures also vary greatly in size. Some bubbles may be smaller than the reconstruction voxel, so a binary classification of each voxel as either liquid or gas can misrepresent the material present. A volume-fraction model instead estimates the proportion of each material within a voxel. This allows intermediate values to represent partially filled voxels rather than forcing every location into a single phase.
A statistical model for energy-integrating detectors
Energy-integrating semiconductor detectors measure the accumulated energy transferred by incoming X-ray photons. Their measurements do not follow the simple Poisson model commonly associated with photon-counting detectors. The underlying statistics are more complicated because each photon can transfer a varying amount of energy to the detector.
The reconstruction approach described in the study represents these measurements with a nonlinear Gaussian approximation. Its predicted mean and variance depend on the estimated transmitted photon signal, rather than treating measurement uncertainty as fixed. The model is designed to approximate the compound Poisson behavior of energy-integrating measurements while remaining practical for iterative reconstruction.
The method also accounts for the energy dependence of X-ray attenuation through a dual-material model. In the simulated application, the materials are titanium and water. Their known attenuation properties are combined with estimated material volume fractions to predict how X-rays travel through the pipe and its contents. Modeling two materials can help address beam hardening, in which lower-energy photons are attenuated more strongly as the beam passes through an object, changing the spectrum reaching the detector.
Using physical constraints to narrow the solution
A single energy-integrated measurement does not uniquely determine the fractions of two materials. The reconstruction is therefore underdetermined unless additional information is introduced. A key constraint is that each material fraction must be nonnegative and the sum of the modeled fractions cannot exceed one. The remaining portion can represent air, whose X-ray attenuation is treated as negligible in the model.
This volume-fraction-sum constraint limits the range of physically plausible solutions. It is especially useful when projection data are incomplete: rather than allowing estimates to vary freely, the algorithm keeps each voxel’s material fractions within a meaningful range. The constraint supports material separation and, in the reported simulation, also helped reduce limited-angle and beam-hardening artifacts.
Matching regularization to material structure
Constraints alone do not guarantee a unique or stable reconstruction. The method adds regularization to guide the solution toward images with appropriate spatial characteristics. It uses an edge-preserving hyperbola regularizer to discourage unwanted variation without excessively smoothing boundaries. It also applies an ℓ0 sparsity penalty to the titanium fraction image.
The two regularizers are not applied identically to both materials. The approach treats the metal image as having comparatively flat regions and sharper boundaries, while the liquid fraction can vary more continuously. Applying material-specific regularization reflects those differences and helps the reconstruction manage artifacts without imposing the same assumptions on every image.
Because the objective function is nonlinear and includes constraints, the images are estimated iteratively. A preconditioned gradient method with an approximate majorization-minimization strategy sets update sizes, while an accelerated proximal-gradient approach handles the sparsity penalty. In practical terms, these methods are ways to make successive image updates more manageable while respecting the model’s constraints.
What the simulation results indicate
The method was evaluated using a synthetic phantom representing a water-filled titanium pipe, titanium rods and bubbles of varying sizes. The comparison included a simple Poisson approach and nonlinear Gaussian approximation (NGA) reconstructions with and without the volume-fraction-sum constraint. The reported results are simulations, not a demonstration of performance on operating plant equipment.
Across the tested source settings, NGA with the volume-fraction-sum constraint produced the lowest reported water-region root mean square error. For example, at the simulated 20 kW setting, the reported error was 0.096 with the constraint, compared with 0.106 for NGA without it and 0.302 for the simple Poisson method. The constrained NGA approach also reduced visible limited-angle patterns in the reconstructed images.
The results have an important qualification: at very low photon levels, the Gaussian approximation can become less accurate because the underlying measurement distribution remains skewed. In the simulation, this was associated with increased noise in the metal region under the lower-power condition. Detector noise, charge sharing and X-ray scatter were also excluded from the simulation, so the results should be interpreted within those modeling limits.
Practical implications for CT system design
For engineers assessing high-speed CT for multiphase flow, the reconstruction algorithm is part of the measurement system—not simply a final image-processing step. Detector statistics, spectral attenuation, angular coverage and prior knowledge about the materials all affect the estimated image. A useful evaluation should therefore consider whether the measurement model matches the detector, whether the material constraints reflect the application and how image quality changes when photon levels are low.
The study’s central lesson is that limited-angle data need more than conventional image smoothing. Combining a detector-aware statistical model with physically meaningful volume fractions and material-specific regularization can improve simulated reconstructions of multiphase flow in a metal pipe. For industrial and research teams, these techniques provide a framework to evaluate, while recognizing that validation with the intended hardware and measurement conditions remains essential.