Who it serves
Metallurgists, materials researchers, quality engineers, laboratory technicians, process engineers, and supplier-quality teams.
Computer vision · Materials AI
MetalloVision
An AI-powered microstructure characterization platform that turns metallographic images into consistent measurements, explainable findings, and inspection-ready reports.
Explore the overview ↓Concept case study · 2026
AI-assisted
microstructure intelligence
microstructure intelligence
ProblemManual image analysis is slow, subjective, and difficult to reproduce at scale.
SolutionHuman-in-the-loop computer vision for measurement, defect detection, and reporting.
OutcomeFaster decisions, consistent evidence, and traceable metallurgical knowledge.
01 Business overview
From microscope image
to confident decision.
Metallographic inspection is central to material qualification, failure analysis, process development, and production quality. Yet much of the work still depends on manual thresholding, visual estimation, and specialist interpretation.
MetalloVision standardizes the repeatable part of that workflow while keeping a metallurgist in control of final judgment. It creates a shared evidence layer between laboratories, production teams, suppliers, and customers.
Why it matters
Reduces analysis time, improves repeatability, captures expert feedback, and makes inspection results easier to compare and audit.
Where it fits
Incoming inspection, heat-treatment validation, weld assessment, failure analysis, R&D experiments, and production troubleshooting.
Business success measures
Turnaround timeImage upload to approved report
RepeatabilityAgreement across analysts and sites
Review effortCorrections required per analysis
AdoptionApproved analyses completed monthly
02 Product capabilities
One workflow.
Multiple measurements.
Every prediction remains connected to the source image, scale calibration, model version, confidence, reviewer, and correction history.
Phase segmentation
Pixel-level identification of ferrite, pearlite, martensite, bainite, and non-metallic inclusions.
Quantitative analysis
Phase fraction, ASTM-aligned grain-size indicators, object morphology, and spatial distribution.
Defect detection
Detection and localization of cracks, pores, voids, scratches, and preparation artefacts.
Inspection reporting
Traceable, review-ready reports combining images, measurements, confidence, and model metadata.
03 Technical overview
Designed for accuracy,
traceability, and learning.
The platform separates image processing, model inference, measurement, review, and reporting so each stage can be tested and improved independently.
A human-in-the-loop design handles ambiguous boundaries, unfamiliar alloys, and preparation artefacts without presenting uncertain predictions as facts.
Ingest
Upload optical or SEM micrographs with alloy, process, magnification, scale, and sample metadata.
Prepare
Validate image quality, normalize contrast, calibrate scale, tile large images, and preserve the original.
Infer
Run segmentation and detection models; calculate grain, phase, inclusion, and defect measurements.
Review
Overlay predictions, expose confidence, and let a metallurgist correct masks or reject uncertain results.
Report
Generate a versioned inspection report and export structured results for quality and R&D systems.
Computer vision
U-Net or DeepLabV3+ for semantic segmentation; Mask R-CNN or YOLO segmentation for inclusions and defects; transfer learning with class-aware augmentation.
Image analytics
OpenCV and scikit-image for calibration, morphology, connected components, boundary analysis, intercept-based grain indicators, and overlay generation.
Application & APIs
React or Streamlit review interface, FastAPI services, PostgreSQL metadata, object storage for images and artefacts, and asynchronous batch processing.
MLOps & governance
MLflow experiment tracking and registry, Docker deployment, CI/CD validation, dataset versioning, audit logs, model monitoring, and role-based approval.
04 Data & evaluation
Measure more than
model accuracy.
Technical metrics are paired with metallurgical acceptance criteria and workflow outcomes.
SegmentationmIoU, Dice/F1 per phase, boundary F1, confusion matrix
DetectionmAP, precision, recall, false positives per image, size-stratified recall
MeasurementPhase-fraction error, grain-size agreement, repeatability and reproducibility
ReliabilityCalibration error, uncertainty coverage, out-of-distribution detection, drift
ProductReview time, correction rate, approval rate, report turnaround, analyst adoption
Key risks and controls
- Domain shift: stratified datasets across microscopes, etchants, magnifications, alloys, and sites.
- Annotation variability: written labeling protocol, dual review, consensus masks, and inter-rater measurement.
- False confidence: calibrated probabilities, uncertainty overlays, quality gates, and mandatory expert approval.
- Traceability: immutable originals, versioned models, complete audit history, and reproducible report inputs.
05 Delivery roadmap
Build narrow.
Validate deeply. Scale carefully.
The first release focuses on a small, well-defined set of microstructures and measurements before expanding to new materials and sites.
MVP
Ferrite/pearlite segmentation, phase fraction, calibrated grain measurements, review UI, and PDF/CSV reports.
Pilot
Martensite and inclusion classes, defect detection, active-learning loop, role-based review, and batch analysis.
Scale
Plant integration, drift monitoring, multi-site model governance, standards templates, and edge inference.
Project vision
Make metallurgical expertise
more scalable—not less human.
MetalloVision is designed as a decision-support system: automation handles repetition, while experts remain responsible for interpretation and approval.
Discuss MetalloVision ↗