A metallurgist staring down a microscope at a polished, etched sample is making a judgment call that decides whether a batch of steel ships or gets scrapped. ASTM grain size rating, phase distribution, inclusion content — these numbers determine strength, toughness, fatigue life, and whether a part survives its service life. The trouble is that the method has barely changed in decades: one person, one field of view, one comparison chart, repeated a handful of times per lot. AI vision changes what a metallography lab can actually see and how fast it can see it. If manual grain counting is the bottleneck between your furnace and your certificate of conformance, book a demo and we will show you automated ASTM rating running on your own micrographs.
Turn Every Micrograph Into an Automated, Repeatable ASTM Grain Size Rating
AI vision reads metallographic images the way a trained metallurgist does — classifying grain size, phase distribution, and inclusion content — but does it on every field of view, every sample, every shift, without the fatigue or the variability that comes with manual counting.
The Comparison Chart Was Built for Speed, Not for Consistency
ASTM E112 gives metallurgists three ways to rate grain size: comparison against a standard chart, the planimetric or Jeffries count within a defined area, and the intercept method, which counts grain boundary crossings along test lines. The comparison method is the one almost every production lab actually uses day to day, because it is fast — a prepared and etched section is held up against a reference chart and the analyst records whichever image looks like the closest match.
That speed comes at a cost the standard itself acknowledges. The comparison method is explicitly described as the faster but more subjective of the three approaches, sufficient for routine quality control but dependent on the analyst's judgment of a visual match. The intercept method is more rigorous and can achieve considerably tighter precision, but it requires an actual count of boundary crossings along test lines, which takes real time and concentration, and its accuracy still depends directly on how many intercepts get counted. On a busy production floor, that thoroughness is often the first casualty of the shift.
Comparison Method
Fastest, most common on the floor, but the standard itself flags it as a subjective visual match rather than a measured value.
Planimetric Count
Counts grains within a defined area of known size, with grains touching the boundary weighted at half value. More rigorous, more time-consuming.
Intercept Method
Counts boundary crossings along test lines to calculate mean lineal intercept length. The most precise manual method, and the slowest to do properly.
Every one of these methods was designed around what a human analyst can reasonably do with a microscope, a chart, and a limited number of minutes per sample. None of them were designed around what becomes possible once a trained model can apply the same rigorous counting logic to every field of view on every sample, instantly and without fatigue.
Three Structural Properties, One Automated Read
A metallographic image carries more information than a single grain size number, and a capable AI system is built to extract all of it from the same micrograph rather than requiring a separate manual procedure for each property.
Each of these three reads happens on the same captured image, which means a single micrograph produces a complete structural profile instead of three separate manual procedures each competing for the same limited lab time.
From Pixel-Wise Segmentation to a Number a Metallurgist Trusts
The technical challenge in automating metallography is not spotting that a boundary exists, it is learning to trace it the way an experienced eye does, distinguishing a true grain boundary from etching artifact, scratch, or phase contrast that only looks like one. This is a segmentation problem, and it is where the underlying deep learning architecture matters most.
Image Capture and Preparation
A polished, etched sample is imaged under consistent magnification and lighting, producing the same kind of micrograph a metallurgist would otherwise study directly through the eyepiece.
Pixel-Wise Segmentation
A trained segmentation network classifies every pixel into grain interior, grain boundary, or a specific phase, producing a labeled map of the entire structure rather than a handful of manually traced lines.
Geometric Measurement
From the segmented map, the system derives mean lineal intercept length, grain area, phase fraction, and inclusion morphology using the same underlying geometry ASTM E112 defines, just applied automatically and exhaustively.
Standard-Referenced Output
Measurements convert to the ASTM grain size number, phase area percentages, and inclusion ratings, output in the same terms a metallurgist already reports against, so results slot directly into existing quality documentation.
The reason segmentation-based approaches have proven effective is that they learn structure the way a trained eye does — from labeled examples of what a real boundary looks like against real background noise — rather than relying on a fixed brightness threshold that breaks down the moment etching contrast or lighting varies even slightly between samples.
Bring Your Own Micrographs to the Demo
The most convincing way to evaluate automated metallography is against your own material, your own etch, and your own historical ASTM ratings. Send a set of sample images ahead of time and we will show you side-by-side results comparing automated output against your lab's manual determination.
Two Metallurgists, Two Comparison Charts, Two Different Answers
The comparison method's greatest practical weakness is not speed, it is reproducibility. Two competent analysts looking at the same borderline field of view can reasonably select adjacent chart images and report grain size numbers that differ by a full unit or more, particularly on structures that fall between two reference images rather than matching one cleanly. That variability is quietly baked into quality records across the industry, and it rarely shows up until a customer audit or a failure investigation asks why two certificates for what should be the same material report different numbers.
This is not an argument that manual metallography is being done poorly. It is an argument that even skilled, careful analysts are working within the inherent limits of a visual comparison method performed on a small sample of fields, and that a consistent, exhaustive automated read closes a gap the standard itself was designed around, not against.
Grain Structure Is a Readout of What Happened in the Furnace
Grain size and phase distribution are not abstract lab numbers, they are a direct record of thermal history. Grain growth accelerates with time and temperature above the recrystallization threshold, phase fractions shift with cooling rate, and inclusion content traces back to melt cleanliness and deoxidation practice. That means a metallography lab that can turn results around in minutes instead of hours becomes a genuine process control input rather than a lagging quality checkpoint.
Heat Treatment Verification
Confirm target grain size and phase balance were achieved immediately after a quench or temper cycle, catching a furnace deviation before the next batch runs the same recipe.
Casting and Solidification Control
Track columnar-to-equiaxed transition and grain refinement across a cast cross-section to validate that inoculation and cooling practice are producing the intended structure.
Forging and Hot Working Records
Document grain flow and recrystallization behavior across a forging to support both quality certification and root-cause investigation if a part underperforms in service.
Cleanliness and Inclusion Rating
Quantify non-metallic inclusion type, size, and distribution to support melt cleanliness certification for fatigue-critical and aerospace-grade material.
What Changes When Rating Stops Being the Bottleneck
That answer took seconds because every field of view had already been measured, not just the one or two a technician had time to manually count. On a manual process, answering that same question means going back to the microscope, re-preparing focus across the section, and spending real time re-counting fields the original inspection never touched.
Getting From Pilot Samples to a Validated Lab Workflow
Automated metallography is not deployed by pointing a generic model at your microscope on day one. It is validated against your specific materials, etchants, and historical ratings before it becomes part of your certified quality process.
Baseline Against Historical Data
Run the model against a set of previously rated samples with known manual ASTM results, establishing agreement before anything changes in the lab workflow.
Calibrate to Your Etch and Material
Fine-tune segmentation against your specific alloys, etchant chemistry, and imaging setup, since grain boundary contrast varies meaningfully between material families.
Run in Parallel
Automated results run alongside manual rating for a defined period, giving your metallurgists direct comparison and the chance to flag any disagreement before automation carries production weight.
Adopt as Primary Record
Once agreement is confirmed, automated rating becomes the documented result for routine samples, with manual review retained for edge cases and any result outside expected range.
What Metallurgists and Quality Managers Ask Before Adopting AI Rating
Stop Rating Grain Size One Field of View at a Time
See how automated ASTM grain size rating, phase segmentation, and inclusion classification perform against your own micrographs, your own alloys, and your own historical results. If it doesn't match what your metallurgists already trust, it's not ready for your floor, and that is exactly what the pilot is built to prove.
