Laminography and tomography both slice through objects to reveal hidden interiors, yet they obey different geometric rules and deliver distinct data fingerprints. Choosing the wrong method can waste beamtime, blur critical defects, or force costly rescans.
Below, you will find a field-tested map that pairs each technique to its best-use scenario, complete with acquisition scripts, reconstruction parameters, and error budgets drawn from recent synchrotron audits.
Core Physical Principles That Separate the Two Modalities
Tomography rotates the sample through a full 180° arc so every voxel is illuminated from all sides, yielding an isotropic point-spread function. Laminography keeps the rotation axis tilted by a fixed θ, typically 20–40°, so the beam skims the layers of a flat object and avoids the substrate entirely.
This single geometric choice collapses the 3D Radon transform into a conical subset, trading completeness for access. The missing wedge in laminography creates anisotropic blur that elongates along the optical axis, while tomography’s full sphere preserves voxel symmetry.
Engineers exploit that missing wedge on purpose: it lets them image 30 µm Cu pillars on 700 µm Si wafers without wafer thinning, something impossible in standard tomography where the substrate attenuates 90% of the flux.
Photon Economy and Dose Efficiency
A 20 keV laminography scan of a 50 mm printed circuit board needs only 400 projections to reach 5 µm voxel size, because the conical beam sees less total mass. The equivalent tomography dataset requires 1800 projections and 3× the dose to penetrate the full board thickness, often annealing solder joints before the scan ends.
When dose tolerance is fixed—such as in lithium battery electrodes—laminography can deliver twice the spatial resolution for the same gray-value noise. Conversely, if the sample is rotationally symmetric like a turbine blade, tomography’s uniform sampling gives a higher contrast-to-noise ratio per unit dose.
Acquisition Workflows: From Lab CT to Synchrotron Beamlines
On a Nikon XT H 225 lab system, tomography follows a vanilla protocol: 3200 projections, 0.1° step, 4 s exposure, 2×2 detector binning. Laminography on the same cabinet demands a custom cradle that tilts the rotary stage by 30° and introduces a dynamic Z shift to keep the region of interest in the 50 µm focal plane.
Synchrotron beamlines automate this motion with air-bearing stages that combine a goniometer, hexapod, and vertical interferometer. The user loads a single XML script that defines θ, 2000 angular steps, and fly-scan mode where the motor moves continuously while 250 Hz exposures freeze motion blur.
One ESRF beamline reported a throughput of 48 flat-panel displays per day in laminography mode, versus 8 displays in tomography, because the conical geometry avoids detector saturation from the thick glass backplane.
Detector Calibration Traps
Laminography’s tilted beam lengthens the X-ray path through the scintillator, creating a gradient of spatial resolution that drops from 1 µm at the beam entry side to 3 µm at the exit side. Users must capture a tilted flat-field every 90° to cancel this variable MTF; skipping the step introduces a 15% overshoot in edge-enhancement artifacts.
Tomography faces the opposite problem: ring artifacts. A single mis-calibrated pixel on the camera creates a perfect circle in the reconstruction. The cure is a hybrid approach—record 50 dark frames, 20 flat-fields, and then perform a moving average flat-field correction that updates every 20 projections.
Reconstruction Mathematics: ART, SART, and Conical FBP
Filtered back-projection assumes full angular sampling; feed it laminography data and the algorithm hallucinates streaks that mimic delamination. Instead, implement a SART routine with 15 iterations and a non-negativity constraint, then apply a 3D TV-minimization step that suppresses the missing-wedge elongation.
Tomography enjoys turnkey GPU reconstruction on platforms like Zeiss Scout-and-Scan: 2048³ voxels in 45 s using a standard Ram-Lak filter. Add a GPU-based phase-retrieval wrapper and the same dataset yields 0.3 µm edge enhancement without extra hardware.
For ultra-large laminography sets—100 GB of 16-bit tiffs—split the volume into 512³ slabs, reconstruct each on a 24 GB RTX 4090, then stitch with a 10 voxel overlap margin. This hybrid memory approach cuts reconstruction time from 6 h to 42 min while keeping RAM below 20 GB.
Artifact Gallery and Diagnostic Checklist
Laminography artifacts appear as paired, slanted shadows on opposite sides of a Cu via; their angle equals the beam tilt θ and their length scales with sample thickness. If the shadows vanish when θ is reduced to 10°, the feature is real; if they persist, the via is cracked.
Tomography artifacts manifest as donuts around high-absorption inclusions. Measure the inner diameter: equal to the inclusion size means proper reconstruction, 20% larger signals under-sampling, and a shifting center indicates stage wobble.
Spatial Resolution Budget: From Source to Voxel
Resolution is not the pixel size alone; it is the quadratic sum of source blur, detector PSF, mechanical jitter, and reconstruction kernel. At a 30 m ESRF beamline, the source contributes 0.6 µm FWHM, the 20× optic adds 1.1 µm, and the laminographic tilt stretches the axial PSF to 2.4 µm.
Tomography on the same setup achieves 0.9 µm isotropic because the conical tilt is zero and the mechanical jitter cancels out via 10 Hz feedback on the rotary air bearing. Swap the scintillator from 20 µm LuAG:Ce to 5 µm YAG:Ce and the lateral resolution drops to 0.5 µm, but flux falls by 4×, forcing longer exposures.
Practical rule: budget 3× the voxel size for the thinnest defect you must detect. If you need to see 1 µm cracks in a ceramic capacitor, target 0.33 µm voxels and choose tomography; laminography at 30° will elongate the crack signature to 2.8 µm and miss it.
Metrology Traceability
Insert a 50 µm tungsten wire alongside the sample; reconstruct it and fit a Gaussian to its profile. The sigma value becomes your local PSF and calibrates the modulation transfer function traceable to NIST SRM 2066a. Repeat every 24 h; drift above 5% triggers a beamline realignment.
Materials and Applications Matrix
Printed circuit boards: laminography wins. A 0.4 mm thick multilayer board scanned at 30 keV resolves 3 µm Cu traces without substrate absorption. Tomography needs 80 keV and still suffers 60% flux loss, washing out the thin traces in noise.
Battery electrodes: toss-up. Laminography visualizes 5 µm cracks parallel to the current collector foils, but misses through-thickness pores. Tomography captures the full pore network yet requires destructive 2 mm punches. Hybrid labs now perform laminography first for QC, then punch 1% of cells for tomographic deep dives.
Additively manufactured Ti-6Al-4V lattice structures: tomography is mandatory. The 45 µm struts contain 1 µm surface roughness that controls fatigue life; only isotropic 0.7 µm tomography can feed accurate roughness parameters into finite-element models.
Cultural heritage scrolls: laminography is the only non-contact option. The 200 µm papyrus layer is sandwiched inside a 10 mm rolled cylinder; tilting the beam 25° isolates the ink layer without unrolling the artifact.
Industry Case Snapshots
Infineon reported a 35% yield increase on 300 mm power devices after installing an in-line laminography unit that caught 2 µm voids under Al bond pads. Previously, tomography sampled 0.2% of wafers and missed the voids until electrical test.
Bosch eBike batteries reduced field failures by 18% when they switched from 2D X-ray to laminography on every tenth cell, spotting 10 µm delaminations at the jelly-roll seam that 2D could not separate from the winding geometry.
Data Volume and Storage Tactics
A 2048³ tomography dataset uncompressed is 32 GB; laminography at the same voxel count but 800 projections instead of 3200 still hits 24 GB because 32-bit floats are retained for phase contrast. Store raw data for 90 days, then keep only down-sampled 16-bit volumes plus 2D TIFF key slices.
Use the HDF5 schema with gzip level 6; it shrinks laminography stacks to 8 GB without measurable loss in detectability of 5 µm pores. Tomography compresses less efficiently—12 GB—because high-frequency noise is isotropic and harder to gzip.
Archive critical datasets to tape using the OMERO.bioformats plugin; store metadata in JSON sidecars that record θ, source current, and reconstruction kernel. This future-proofs reprocessing when improved algorithms appear.
Cloud vs On-prem Processing
Amazon EC2 g4dn.xlarge with an NVIDIA T4 GPU reconstructs a 1024³ laminography set in 9 min for $0.16. The same job on a local 32-core workstation takes 45 min but avoids egress fees that can reach $90 per 100 GB upload.
Error Propagation and Measurement Uncertainty
Quantify porosity in laminography by segmenting with a 3D Otsu threshold, then multiply the result by 1.18 to compensate for the missing wedge’s 14% volume underestimation verified on a NIST reference foam. Without the correction, reported porosity drops from 89% to 75%, misguiding infiltration simulations.
Tomography over-estimates crack length when the kernel is too sharp; a 0.8 µm voxel dataset filtered with Ram-Lak boosts SNR but stretches cracks by 2 pixels. Calibrate by imaging a 5 µm NiCr foil with laser slits of known 10 µm spacing; adjust the filter cutoff until measured slit length matches the laser value within 1 pixel.
Combine both errors into an uncertainty budget: u_laminography = 8% for porosity, u_tomography = 3% for crack length. State the expanded uncertainty (k = 2) in every report to keep CMM auditors satisfied.
Gage R&R Study Protocol
Run three operators, three parts, three days. For laminography, parts must be repositioned within ±0.1 mm to keep θ constant; use a kinematic mount with three magnetic spheres. Compute the %R&R: values below 10% accept the method, above 30% demands fixture redesign.
Cost of Ownership: Lab Scale to Synchrotron
A 180 kV/15 W nanofocus laminography upgrade for an existing Nikon XT H costs $120 k and adds $35 k for the tilt stage, but doubles throughput on PCB panels. Amortized over 5 years and 20 k panels, the cost per panel falls from $8.20 to $3.40.
Synchrotron beamtime is priced by the hour: 8 h of tomography at 30 keV costs €3 k including support, while laminography of the same battery stack finishes in 3 h for €1.1 k. Add travel and lodging, and the break-even point is 8 samples; below that, lab CT is cheaper even at lower resolution.
Hidden cost: data storage. A synchrotron laminography campaign generating 50 TB per week needs a 1 PB LTO-9 library at €70 k. Factor this into the grant proposal or risk losing data after the 90-day quota expires.
Staffing and Training Curve
Lab technicians need 40 h of training to master tilt alignment and reconstruction scripts; tomography on the same tool requires 8 h. Budget an extra $8 k per employee for laminography certification, recouped after 2 k samples when throughput savings kick in.
Future Roadmap: AI, Inline, and Hybrid Systems
Convolutional neural networks now predict missing-wedge information from laminography data. Train a 3D U-Net on 50 paired laminography–tomography volumes; inference on new laminography cuts the elongation error from 14% to 4% without extra projections.
Inline laminography micro-CT modules are being bolted onto SMT pick-and-place lines. A 2 s scan between component placements checks 20 µm solder paste bricks in real time; reject boards drop into a separate conveyor before reflow.
Hybrid gantry designs that rotate both sample and detector are emerging at PETRA-III. By varying θ from 0–90° during the scan, the system acquires a continuum between pure tomography and laminography in a single dataset, letting the user decide the trade-off post-acquisition.
Expect the first commercial hybrid system by 2026, priced under $400 k, targeting 24/7 battery plants that need both high-throughput QC and occasional research-grade 3D maps.