A polished cross-section can look nearly featureless in secondary-electron contrast while containing a crystallographic architecture that controls current flow, slip, cracking, diffusion, phase transformation, and reliability. Orientation imaging microscopy makes that architecture visible by attaching a measured phase and lattice orientation to known positions, then transforming the measurements into maps, grains, boundaries, texture, and local statistics. The color image is the beginning of interpretation, not the measurement itself. Every defensible conclusion still depends on the diffraction patterns, acquisition geometry, specimen and crystal reference frames, symmetry definitions, spatial sampling, segmentation rules, unindexed points, and processing history that produced it.

Orientation imaging microscopy is a measurement-and-inference workflow, not one colored map. The term was introduced for automated local-orientation measurements on a grid, originally using electron backscatter diffraction patterns in a scanning electron microscope. In current practice, OIM often describes EBSD acquisition and analysis, while OIM Analysis is also an EDAX product name. The underlying concept is broader than a proprietary file or display: spatial coordinates and crystallographic orientations are paired so morphology can be analyzed together with phase, texture, misorientation, and boundary character. TKD, precession electron diffraction, and four-dimensional STEM can generate related orientation maps, but their signal formation, resolution, uncertainty, and metadata are not interchangeable with surface EBSD.

The central data object is a spatially sampled orientation field. If $\mathbf{v}_c$ is a direction expressed in the crystal frame and $\mathbf{v}_s$ is the same physical direction expressed in the specimen frame, an orientation $g$ may be written under one common active convention as

Other communities and software may store the inverse transformation, passive rotations, different Euler-angle sequences, or different handedness. Numerically valid Euler angles can therefore describe a mirrored or rotated physical sample when imported under the wrong convention. The map axes, surface normal, rolling direction, transverse direction, wafer notch, device-line direction, and detector frame must be tied to physical fiducials. Merely rotating a plot until it looks familiar does not repair a reference-frame error.

Reference frames and crystal symmetry determine what every orientation number means. An orientation is not a direction alone; it maps a crystallographic basis into a declared specimen basis. Crystal symmetry makes several rotation representations physically equivalent. Sample symmetry may further reduce a texture description, but it should be imposed only when specimen processing and sampling justify it. A phase assigned the wrong point group can alter orientation coloring, misorientation distributions, special-boundary classification, and texture strength even when the visible map looks plausible.

An inverse-pole-figure map colors the crystal direction aligned with a selected specimen direction. IPF-X, IPF-Y, and IPF-Z are different maps. “Z” may mean surface normal, beam direction, or a software display axis depending on the workflow. A complete figure states the specimen direction, phase symmetry, color key, map axes, scale, indexed fraction, and whether coordinates were transformed. Pole figures describe selected crystal directions or plane normals in specimen space; inverse pole figures describe specimen directions in crystal space. Confusing them can reverse the physical interpretation of fiber texture or epitaxial alignment.

Orientation representations should be converted with tested library operations rather than hand-edited Euler columns. Euler angles have singularities and depend on convention; rotation matrices are redundant but direct; unit quaternions are compact but double-cover rotations. Averaging Euler-angle components is generally not a crystallographic mean. Mean orientations and interpolation need symmetry-aware operations on the rotation manifold, along with a declared weighting and spread metric.

The acquisition grid samples a finite diffraction response rather than infinitesimal pixels. Step size sets the distance between reported coordinates, while spatial resolution is governed by probe size, scattering volume, specimen geometry, signal formation, detector response, drift, and indexing behavior. Oversampling produces correlated neighboring measurements and more dose; undersampling can miss small grains, thin twins, boundary curvature, or a minority phase. A grain represented by only a few points has size and shape dominated by grid placement.

Square and hexagonal grids encode different neighbors and cell areas. Missing scan lines, backlash, charging, and drift can warp coordinates. A metrology map should retain original coordinates and a documented spatial transform rather than silently forcing points onto an ideal grid. Fiducials, SEM images, scan reversal, and overlap regions can quantify distortion.

Sampling design depends on the inference. A high-resolution field may resolve subgrains but contain too few grains for texture; a wide field may estimate texture while undersampling thin device features. Representative statistics require positions across relevant wafer radii, dies, pattern densities, layers, process splits, and failure classes. Several separated fields usually estimate spatial heterogeneity better than one contiguous, visually attractive region. The number of pixels is not the number of independent grains, and the number of grains is not automatically the effective sample size for a strongly textured or spatially correlated microstructure.

Define the phase, texture, grain, boundary, deformation, or reliability question
  -> Select EBSD, TKD, TEM orientation mapping, or a correlative combination
  -> Establish specimen axes from fiducials, process history, and mounting geometry
  -> Prepare the surface or foil and qualify damage, relief, charging, and representativeness
  -> Calibrate detector projection, stage coordinates, spatial scale, and pattern response
  -> Choose map area, grid, step, exposure, and sampling sites from resolution and statistics
  -> Acquire raw patterns, backgrounds, standards, contextual images, and metadata
  -> Index all plausible phases and retain alternatives, residuals, and unindexed sites
  -> Verify map, specimen, crystal, and display reference-frame relationships
  -> Reconstruct grains with declared symmetry, neighborhood, threshold, and size rules
  -> Compare raw and cleaned maps and run parameter-sensitivity analysis
  -> Compute texture, boundaries, morphology, and local statistics with correct weights
  -> Register chemistry or imaging and validate conclusions independently
  -> Archive raw data, transforms, software, scripts, parameters, and uncertainty

Grains and boundaries are reconstructed objects whose definitions must be exposed. A common grain algorithm connects neighboring indexed points of the same phase when their symmetry-reduced disorientation is below a selected threshold. For orientations $g_1$ and $g_2$, one representation is

where $S_a$ and $S_b$ are symmetry operations in the phase group $\mathcal{G}$. Software conventions may reduce the same physical relationship differently, but the minimum physical disorientation must respect symmetry. The segmentation threshold is an analysis parameter, not a universal law separating grains from subgrains. Recrystallized material with a clear bimodal neighbor distribution may be insensitive over a reasonable range; a deformed, gradient-rich, or noisy map may change grain count dramatically.

Phase boundaries, twins, low-angle boundaries, and coincidence-site-lattice labels add separate rules. A two-dimensional map measures a boundary trace and lattice disorientation. Full five-parameter grain-boundary character also needs the boundary-plane normal in crystal coordinates, which a single planar section generally does not supply. A special misorientation within a tolerance is not proof of the corresponding coherent boundary plane or property. Boundary-length fractions on a section are not automatically three-dimensional boundary-area fractions.

Grain size has multiple legitimate definitions. For a reconstructed planar grain of area $A$, the equivalent-circle diameter is

but Feret diameters, intercept lengths, area-weighted means, number-weighted means, and three-dimensional estimates answer different questions. Edge grains may be excluded, truncated, or weighted; twins may be counted as boundaries or merged; small grains may fall below the resolution cutoff. A report should name the metric, weighting, edge rule, minimum grain size, twin treatment, and uncertainty. Histograms alone can conceal these choices.

Cleanup changes the microstructure model and must remain reversible. Wild-spike removal, confidence standardization, neighbor orientation correlation, dilation, fill, smoothing, and minimum-grain filtering can suppress isolated errors. The same operations can erase a real nanoscale phase, close a crack, bridge a thin twin, move a boundary, inflate texture, or create an apparently continuous grain. Unindexed points are not empty background: they may mark pores, topography, amorphous material, pattern overlap, charging, damage, or a phase missing from the library.

Processing should begin from an immutable raw or as-indexed dataset and produce versioned derivatives. Each operation needs parameters, order, software version, and the count of points changed. Comparing raw, lightly processed, and sensitivity-case results reveals whether a conclusion is robust. Grain size, phase fraction, boundary fraction, KAM, GOS, and texture should be recalculated across reasonable cleanup and segmentation choices when they drive a decision.

Pattern quality and indexing confidence should not be conflated. Image quality or band contrast describes aspects of the diffraction signal; confidence or fit describes preference under a particular indexing model. A sharp pattern can be assigned to the wrong phase because of pseudosymmetry or an incomplete library. A weak pattern can have the correct orientation with larger uncertainty. Vendor metrics have different scales and meanings and are not universal probabilities. Raw-pattern review at unexpected phases, critical boundaries, low-confidence regions, and representative good regions is indispensable.

Texture and local-misorientation statistics need sampling and scale attached. A crystallographic texture may be represented by an orientation distribution function $f(g)$ over orientation space. Under a normalized measure,

with the domain reduced by appropriate crystal symmetry and any justified specimen symmetry. Kernel bandwidth, harmonic order, weighting, grain versus point sampling, and incomplete spatial coverage alter the estimated distribution. Point-weighted texture emphasizes area on the measured section; grain-weighted texture gives each reconstructed grain equal influence. Neither is universally correct. The physical question determines the weighting.

Kernel average misorientation, grain orientation spread, grain reference orientation deviation, and local orientation gradients summarize different relationships. Their magnitude depends on angular noise, step size, neighbor shell, exclusion threshold, grain segmentation, cleanup, and reference choice. KAM is not a direct universal plastic-strain scale. Inferring geometrically necessary dislocation density adds derivatives, Burgers vectors, slip-system assumptions, an incomplete view of the dislocation tensor, and a length scale. A color bar labeled “strain” without that model and calibration overstates the data.

Texture and boundary distributions also carry statistical uncertainty. Neighbor pairs share pixels and are not independent observations. Large grains contribute many points and boundary segments; spatial clustering reduces effective sample size. Bootstrap or hierarchical resampling by grain, field, die, or specimen can better reflect the sampling design than resampling individual pixels. Lot-level claims require lot-level replication, not millions of points from one cross-section.

Correlative registration connects orientation to mechanism only when its error is measured. OIM becomes especially powerful when registered to secondary- or backscattered-electron images, EDS or wavelength-dispersive chemistry, cathodoluminescence, AFM, Raman maps, mechanical strain fields, electrical failure sites, or TEM. Registration may require translation, rotation, scale, affine distortion, or nonlinear correction. Fiducials distributed across the field and held-out check points provide a residual error; visual alignment at one feature does not establish nanoscale correspondence elsewhere.

For semiconductor and packaging applications, orientation imaging can relate interconnect texture and boundaries to electromigration, liner or silicide phases to contact resistance, solder and intermetallic grains to crack paths, GaN or SiC domains to epitaxial defects, bonded-metal grains to interface evolution, and ferroelectric or phase-change orientations to switching variability. Those are structure-property hypotheses. Establishing causality requires controlled process splits, representative sampling, independent chemistry or imaging, and electrical or mechanical outcomes—not merely spatial coincidence between two attractive maps.

A reproducible OIM deliverable preserves raw patterns, acquisition conditions, calibration, preparation, coordinates, specimen axes, crystal symmetry, orientation convention, phase library, alternatives, confidence definitions, spatial transform, processing scripts, texture settings, statistical unit, software, and validation. It reports indexed fraction and uncertainty beside polished maps. Read orientation imaging microscopy through the coordinate-frame-symmetry-segmentation-sampling-and-provenance lens.