The phase map is a model output, not an observation
Put a dual-phase steel under the SEM, run a standard EBSD scan, and the software will happily paint you a two-color phase map: ferrite here, martensite there. The colors look like data. They are not. In most DP steels, ferrite, bainite, and martensite are all indexed using the same body-centered cubic structure, so a conventional Hough-transform indexing routine has no crystallographic difference to key on. The phase boundary in that legend is a decision made by a threshold, a classifier, or a simulation — and if your report doesn’t name which one, the number it produces is not defensible.
This matters most when the question is not academic. A stamped DP980 bracket cracks at a sheared edge, or a formed rail shows a split at a bend line, and someone asks whether the hard phase is continuous, coarse, or clustered at the feature that failed. A phase fraction quoted without its discrimination method invites the reader to over-read it. The mechanism is on the table: standard indexing cannot separate these phases crystallographically, so every ferrite/martensite map you have ever seen rests on a secondary signal. The job is to know which signal, and where it fails.
Why standard EBSD indexing cannot do it
Hough-based indexing works by detecting Kikuchi bands, measuring the angles between them, and matching that geometry to candidate crystal structures. That is powerful when the phases differ in structure — ferrite (BCC) versus austenite (FCC) in a duplex stainless steel, for example. It is nearly blind when the phases share a structure. Oxford Instruments states the limitation directly: in some steels, ferrite, bainite, and martensite are usually indexed using the same phase structure, and conventional live EBSD acquisition may be unable to differentiate them (Oxford Instruments, Advanced Phase Discrimination).
Martensite is body-centered tetragonal in principle, but its tetragonality in low-carbon commercial steels is small. A study on a 0.1 wt.% C dual-phase steel put the problem plainly: it is difficult to accurately distinguish BCC and BCT phases with small tetragonality using standard EBSD software (PubMed 33226120). The same source notes that standard Hough-transform indexing cannot resolve the slight c/a variation, though pattern-matching research indicates it may be possible with more demanding methods.
So the phase map you are looking at is not a crystallographic measurement of martensite. It is an inference. What follows is what each inference route actually measures, and how each one fails.
The discrimination routes, and what each one really measures
Pattern quality: band slope and image quality
The most common route uses the fact that martensite is heavily distorted and ferrite is not. Pattern quality drops where the lattice is strained, so a threshold on a quality metric separates the two populations. The metric matters. In one DP steel study, band slope exhibited much stronger phase contrast than the conventional band contrast, and because it is highly sensitive to lattice defects with little orientation dependence, it provided a convenient threshold for phase discrimination (PubMed 23920165). Image quality has been used the same way: a study from U.S. Steel Košice notes that IQ reflects lattice distortion and is therefore used to determine martensite fraction (Černík et al., Structure Evaluation of Dual Phase Steels by EBSD).
What this route measures is strain, not phase identity. A heavily deformed ferrite grain near a sheared edge can have pattern quality as poor as martensite. A tempered or auto-tempered martensite can recover enough to look like ferrite. The threshold is a cut through a continuous distribution, and the fraction you report moves when the cut moves. That is not a reason to abandon the method; it is a reason to report the threshold and to check whether the assignment is stable.
KAM and grain-based classification
Kernel average misorientation and grain characteristics add a second dimension. Martensite laths carry high local misorientation; ferrite grains are comparatively clean. A trained classifier can combine EBSP quality and KAM to separate the phases, and Oxford Instruments describes this as one of the most common applications of its phase reclassification tool — separating martensite and ferrite in steels (Oxford Instruments). The same source describes the mechanism: the user selects discriminating parameters, defines classes, trains on selected regions, and the software builds a look-up table applied to the whole map.
This is more robust than a single threshold, but it is not assumption-free. The classifier is only as good as the training areas, and those areas are chosen by an operator. If the training set contains no recovered ferrite or no coarse martensite, the classifier will not have learned them. The output is a model, and the model’s training data belong in the report.
Chemistry-assisted indexing
If the phases differ in composition, EDS can help. In chemistry-assisted indexing, reference spectra are collected for the phases that cannot be separated crystallographically, and each point’s spectrum is matched against them during indexing. Oxford Instruments notes the drawback: the spatial resolution of simultaneous EDS is constrained by the electron–sample interaction volume, on the order of 1 µm diameter at 20 kV on many materials, which makes sub-micron phase separation difficult (Oxford Instruments).
In a DP steel, the chemical contrast between ferrite and martensite is often modest — partitioning during intercritical annealing concentrates carbon in austenite that becomes martensite, but the absolute difference in a low-carbon grade is small. Chemistry-assisted indexing is a strong tool where the chemistry difference is large. In a 0.1 wt.% C DP steel, it is not the first lever you reach for.
Kikuchi band width
Phases with the same structure but different lattice parameters produce Kikuchi bands of different widths, because band width scales inversely with d-spacing. This can discriminate same-structure phases, but the resolution limit is severe: in a low-resolution, high-speed pattern, an individual Kikuchi band is typically only 5–10 pixels across, so a 10% difference in d-spacing becomes a sub-pixel difference in band width (Oxford Instruments). For ferrite versus martensite, the lattice parameter difference is far smaller than 10%. Band width is not the practical route here.
Tetragonality mapping
The most direct attack on the problem is to measure the tetragonality itself, by comparing experimental patterns with simulated patterns and detecting distortion away from a cubic lattice. The 0.1 wt.% C study did exactly this, using a strain-minimization routine across many grains to determine the pattern centre accurately, then producing tetragonality maps and phase maps. With dynamically simulated patterns, the method reached close to 90% accuracy for grain phase identification when compared with an image-quality identification method (PubMed 33226120).
Read that number carefully. Ninety percent agreement with another method is not ground truth; it is two imperfect methods agreeing most of the time. The same study reports that small errors in determining the pattern centre can cause significant errors in tetragonality measurement and lead to erroneous phase results, and that the tetragonality error is on the order of 1%, producing a commensurate ~0.2% error in carbon content — which makes the technique unsuitable for estimating total carbon in most commercial steels below 0.1 wt.% C. It is a research-grade tool with documented sensitivity, not a push-button production method.
What this means for a failure investigation
Suppose a DP steel component fails at a bend and the question is whether a band of martensite at the bend line initiated the crack. A phase map produced by band-slope threshold alone will show you a dark region and call it martensite. Before that assignment goes into a failure report, run a second, independent discrimination on the same area — KAM, a trained classifier, or a tetragonality check if the capability exists — and see whether the assignment is stable or threshold-sensitive. If the dark region moves when the threshold moves, you have learned something important about the material and about your method.
The practical question in most investigations is not the exact volume fraction. It is whether the hard phase is continuous, coarse, or clustered at the feature that cracked. That question survives imperfect phase discrimination better than a fraction does, because it is a question about morphology and spatial arrangement, which the map shows directly, rather than about a number that depends on a cut through a distribution.
There is also a preparation dependency that belongs in the caveat list. The Košice study reports that determination of austenite content by EBSD was not precise because the measurement is highly dependent on perfect sample preparation (Černík et al.). The same sensitivity applies to phase discrimination by pattern quality: surface damage from mechanical polishing depresses pattern quality everywhere and compresses the contrast between ferrite and martensite. Electropolishing or ion milling is not optional if you intend to threshold on pattern quality.
Bench checklist for the next DP steel phase map
This is a recommendation, not a standard. It is what the evidence above supports as defensible practice.
- Prepare for pattern quality, not just for etching. Mechanical polish followed by electropolish or ion milling. Record the final preparation step. If pattern quality is the discriminator, surface damage is a confounder.
- Acquire at a step size that resolves the martensite morphology you care about. If the question is whether a hard band is continuous at a bend line, the step size must be small relative to the band width. State the step size in the report.
- Run at least two independent discrimination signals on a reference area. Band slope plus KAM, or a trained classifier plus a threshold map. Compare the phase assignment, not just the fraction.
- Record the method, the threshold or training areas, and the known failure modes. A phase fraction without its discrimination method is not a measurement; it is a number.
- Report morphology alongside fraction. Continuity, size, and spatial clustering of the hard phase at the feature of interest are more robust to discrimination error than a volume fraction.
- Caveat the number in the report. State that ferrite and martensite share a BCC indexing structure, name the discrimination route used, and note its documented limit — interaction volume for EDS-assisted work, pattern resolution for band-width work, pattern-centre sensitivity for tetragonality work, training-set bias for classifier work.
FAQ
Can EBSD tell ferrite from martensite at all?
Not by crystal structure alone with standard Hough indexing, because both are commonly indexed as BCC. It can separate them by secondary signals — pattern quality, KAM, chemistry, or tetragonality — each with its own limits. The phase map is an inference from those signals, and the method should be reported with the result.
Is a band-slope threshold good enough for a failure report?
It is a reasonable first pass, but it measures strain, not phase identity. Deformed ferrite and recovered martensite can cross the threshold. Cross-check with a second signal on the same area before the assignment goes into a report, and state the threshold.
Why not just use EDS to separate them?
Chemistry-assisted indexing works where the compositional difference is large. In low-carbon DP steels the difference is modest, and the EDS interaction volume — on the order of 1 µm at 20 kV on many materials — limits spatial resolution for fine martensite. It is one lever, not the only one.
Does the phase fraction I report depend on the method?
Yes. Different discrimination routes cut the same microstructure differently, and each has documented failure modes. That is why the fraction should travel with its method, acquisition conditions, and caveats. An unqualified number invites over-reading.