Seeing Through Carbon

Background

Fuselages, wing spars, Formula 1 monocoques, high-end bicycles, wind turbine blades, and other parts that require strength without weight are made of carbon fiber reinforced plastic (CFRP). Plies with different fiber orientations are stacked into curved, variable-thickness laminates, often joined to the substructure through adhesive bondlines of uncertain thickness, which makes ultrasonic inspection challenging in parts with complex geometry. The material is fully anisotropic—the direction of propagation and the ply being traversed determine the velocity, attenuation, and mode conversion. That anisotropy is the central obstacle to reliable nondestructive testing (NDT) of composite structures.

A missed delamination or disbond at a wing spar is a safety-of-flight issue, and any technique that cannot resolve indications in complex regions, such as under stringers, is only doing part of the job. In practice, A-scan and image interpretation remain largely manual and operator-dependent, carrying a real risk of missed indications under time pressure. Fiber waviness, one of the more strength-critical defect types in a laminate, is close to acoustically invisible under conventional inspection.

Up Until Now

One of today’s most used inspection techniques for composite inspection is pulse-echo (and pitch-catch) ultrasonics. It is well-established and widely deployed, but it encounters limits with CFRP.

Attenuation is strongly frequency-dependent, so ultrasonic methods are pushed toward low-MHz or sub-MHz frequencies to achieve a usable signal-to-noise ratio through the laminate. Lower frequency means coarser resolution. Pulse-echo is well-suited to delamination detection and thickness measurement, where the indication is a distinct echo between the interface and backwall on a B- or C-scan. It is far less suited to wrinkles, kissing bonds, or subtle stiffness and can also be blind to defects oriented parallel to the beam axis rather than normal to it, such as some types of cracking.

Modeling these wavefields is just as hard as measuring them. Full anisotropy, complex part geometry, and ply-by-ply stacking sequence make wave propagation a demanding computational problem, and most legacy solvers were not built to handle fully anisotropic, spatially varying viscoelastic stiffness across a complete part.

Outlook

To better visualize composite defects, we undertook a study using guided waves, looking—counterintuitively—toward lower rather than higher frequencies. These waves are in the 100 kHz range and propagate along the laminate rather than straight through it. They travel farther with lower attenuation, they detect defect types that pulse-echo misses, and they can be excited and received with simple, low-cost piezoelectric transducers.

The first question an inspection design needs to answer is whether a transducer layout is sensitive to a given type of defect. Physics-informed digital twins can answer that: spectral-element simulations on GPUs efficiently solve the full elastodynamic wave equation through fully anisotropic, viscoelastic laminates, curvature included, fast enough to screen many scenarios computationally rather than through large numbers of physical reference test pieces and trial and error.

As shown in the figure, we calibrated a digital twin for a composite wind turbine blade and validated it against a three-component laser Doppler vibrometry (LDV) system. Simulated and measured wavefields agreed closely enough in arrival time and waveform shape that an LDV is unnecessary for routine use. Once wave propagation through a part’s geometry is understood, a handful of sparse sensor points can capture the same information.

Full-waveform inversion (FWI) is what makes that sparse layout possible. FWI fits the entire recorded waveform—amplitude and phase together—against the digital twin’s forward model to recover the material properties behind it. A few sensor locations then stand in for a dense scanning grid, using piezoelectric sensors already available, and can cover meters of material per acquisition.

In addition to the standard B- and C-scans, the inversion turns the output images into a quantitative 3D map of material properties, stiffness, local wave speed, and attenuation. Because FWI works from the whole waveform, it is sensitive to indications that produce only a small perturbation in scattering, phase, or local wave speed. That is a broader detection envelope than pulse-echo or pitch-catch. It still reliably detects porosity, delamination, and impact damage, but can also resolve defect types that leave only a faint signature in the waveform, including fiber waviness.

These results can be compared against a library of synthetic defect cases generated within the digital twin, building a dataset for AI/ML training that covers corner cases no archive of physical reference standards could capture. In effect, a model-based probability-of-detection exercise is run before the part is scanned. The output is a yes/no call with an attached characterization probability. A comparable approach has already worked outside CFRP, in concrete inspection, on a similarly subtle defect.

Combining this quantitative imaging capability with generative AI will let us build solutions that create accurate initial velocity and material models, generate synthetic data, perform rapid analysis and interpretation of data, generate automated reports, and more. Pairing full-waveform inversion of ultrasonic NDT with generative AI will lead to autonomous, highly accurate, and predictive nondestructive evaluation (NDE) and structural health monitoring (SHM) systems for advanced materials such as CFRP and glass fiber reinforced plastic (GFRP).

The bigger shift over the next several years is what gets built from that data. Instead of a digital twin keyed to the manufacturing CAD model, picture one built from the part’s as-measured ultrasonic data, reflecting as-built properties rather than design intent. That makes it possible to compare factories and batches, and to link manufacturing signatures to in-service performance.

Our vision for the next five to seven years is that composite inspection converges on something simple: excite the component with a broadband guided-wave source, and let a GPU-based digital twin—trained on simulation and field data—reconstruct what is inside and return a defensible accept/reject call with a stated confidence level.

Acknowledgments

This work was supported by the US Air Force under SBIR Phase I Contract FA864926P0085 (July 2026), “Automated Inspection Reports with Physics-Informed Digital Twins for Aerospace Composites.”