Carbon Fiber & Composites

Carbon fiber composite manufacturing for aerospace and EV structural applications has a defect tolerance measured in fractions of a percent — and a traditional inspection workflow that catches most defects after the cure cycle has already spent the autoclave time, energy, and tooling cost. Niobia AI connects AFP inline inspection, cure monitoring, and NDT image analysis into a single part genealogy that shifts the detection point upstream of where the cost is committed.

From Fiber to Structural Part

Carbon fiber begins as polyacrylonitrile (PAN) precursor: stabilization at 200–300 °C oxidises the polymer in air, carbonization at 1,000–1,500 °C in nitrogen removes non-carbon elements, and surface treatment improves adhesion to epoxy resin. The resulting fiber tows (typically 3K to 50K filaments) are converted to prepreg by impregnating with partially-cured epoxy resin and rolling onto a backing paper.

Composite layup uses automated fiber placement (AFP) for complex curvature parts or hand layup for lower-volume or highly contoured geometry. AFP deposits prepreg tows at rates of 20–50 kg/hour for large aerospace structures. After layup, the part is debulked, bagged, and placed in the autoclave for cure at 120–180 °C under 5–8 bar pressure for 2–4 hours, depending on the resin system.

Defect Taxonomy: Layup, Cure, and In-Service

Layup defects include fiber misalignment (out-of-spec above 2° for most aerospace applications), gap and overlap between AFP tows (gap creates a resin-rich zone with lower shear strength; overlap creates a stress concentration), and dry spots where resin coverage is insufficient. AFP systems detect these inline using laser profilometry and vision, but false-positive rates on complex curvature can be high without AI-assisted classification.

Cure defects include voids (target <0.5% area fraction on ultrasonic C-scan), porosity from entrapped air or volatile release, and delamination from poor ply consolidation. Voids above 0.5% reduce interlaminar shear strength by 7–10% per percent void content — the relationship that drives the acceptance criterion. Cure cycle deviations (slow ramp rate, temperature non-uniformity in the autoclave, pressure application timing) are the primary drivers of void content.

Where Most Composite Programs Get This Wrong

AFP layup and autoclave cure are treated as independent checkpoints in most composite shops. The defects that escape NDT at final inspection rarely originate where they are found. AFP systems generate gigabytes of inspection data per part; NDT generates a C-scan image and a pass/fail record. The question — "did the fiber misalignment flags in Zone 3 of the layup correlate to the delamination indication in the post-cure scan?" — requires co-registering two datasets with different coordinate systems that were never designed to interoperate.

The second gap is the absence of a cure deviation database. Most facilities maintain autoclave run logs but do not systematically correlate temperature non-uniformity patterns (zone-specific thermocouple deviations) to the specific part locations where void content exceeds acceptance limits. Without that correlation, the same autoclave zone will continue producing defects across thousands of parts.

What AI Process Intelligence Changes

AFP inline AI classification reduces false-positive rates on complex curvature from 30–40% (for rule-based systems) to 5–10%, enabling real-time correction during layup rather than post-layup inspection that requires destructive rework. When those classifications are linked to the cure record and post-cure NDT by part serial number, the layup-to-void-content correlation is available after 50–100 parts — enough data to identify which AFP flags actually predict structural defects.

Automated NDT image analysis replaces analyst-dependent interpretation with AI-consistent classification. On ultrasonic C-scans, AI analysis achieves 95%+ agreement with expert analysts on void detection and significantly higher consistency on borderline indications — the cases where analyst variation is highest and the accept/reject decision is most consequential.

Carbon fiber composite manufacturing intelligence is a long-cycle-time problem: parts take days to make and are designed for decades of service. The data infrastructure that links layup to cure to NDT to in-service inspection is the foundation that allows structural life prediction models to be built — and it has to be started before the parts are in service, not after.