Sounding the Grain Noise: Software-Defined Ultrasound for Cracks Smaller Than the Continuum

Ultrasonic inspection finds cracks in metal by sending a short sound pulse into the part and listening for echoes. The simulators used to design these inspections treat the metal as a smooth, uniform solid. That assumption holds for millimetre-scale defects. It fails for the cracks that matter most for fatigue life: cracks tens to hundreds of micrometres long that start at corrosion pits and grow through individual grains. At that scale every grain boundary also scatters sound. This "grain noise" is a coherent echo pattern fixed to the part's microstructure, and the smooth-solid simulators do not produce it.

We propose to build an inspection loop borrowed from wireless engineering, where a radio first probes the channel and then adapts to it. A software-defined ultrasonic transmitter first spends a few seconds measuring the grain-noise spectrum on a region of the part known to be free of cracks. It then reshapes its transmitted spectrum and recalibrates its detector against that measurement before scanning. We pair this with a matched benchmark of simulated and physically measured raw data on identical fatigue-cracked steel specimens. The benchmark lets us quantify, as a function of wavelength relative to grain size, how far simulation-designed inspections overestimate what real hardware can detect.

Core Claim

Our central claim is that in the regime where sub-continuum structure dominates, the useful control is where the transmitted energy sits in frequency, not how much energy is sent or how long the code is. The reason is that grain clutter is caused by the transmitted pulse itself, so it grows exactly as fast as the crack echo does.

A simple model makes this precise. Let $X(f)$ be the transmitted spectrum, $H_c(f)$ the crack's frequency response, $S_g(f)$ the grain backscatter per unit transmitted power, and $N_0$ the electronic noise level. After matched receive filtering, detectability scales as

$$ \rho[X] = \int \frac{|X(f)|^2\,|H_c(f)|^2}{|X(f)|^2\,S_g(f) + N_0}\,df, $$

maximized subject to a total-energy budget $\int |X(f)|^2 df \le E$ and the transmitter's peak-amplitude limit. As $E$ grows, the integrand saturates at $|H_c(f)|^2/S_g(f)$. Past that point, longer codes or more power add nothing, and only moving energy into frequency bands where the crack-to-grain ratio is high improves detection. The quantity $S_g(f)$ depends on the specific microstructure, and continuum simulators set it to zero. For both reasons, we argue it must be measured on the part.

The claim makes three falsifiable predictions:

  1. A waveform designed from a continuum simulation will place energy at the wrong frequencies once wavelength approaches grain size.
  2. The resulting gap between predicted and measured detectability will grow with $k\bar d$, where $k$ is the wavenumber and $\bar d$ the mean grain size.
  3. A short in-situ measurement of the grain-noise spectrum will close most of that gap.

What Exists and What Is Missing

Work that motivates the proposal. Research on amorphous and granular solids shows that continuum elasticity fails below an emergent length scale.

Together these papers set the target defect population and confirm nobody is measuring it.

Work that supplies methods. - SUPRA: open-source software-defined ultrasound processing for real-time applications moves ultrasound receive processing onto a commodity GPU at about 5 ms per frame. Its transmit path stays fixed, and it assumes homogeneous soft tissue. - Design of the transceiver for a new type HF/VHF ionospheric sounding system streams arbitrary waveforms to digital-to-analog converters in radio. That is the transmit capability we transfer to acoustics. - Spectrum Fitting Approach for Passive Wireless SAW Sensor Interrogation Using Software-Defined Radio shows a software-defined radio driving an acoustic device, but only as a narrowband frequency tracker on an ideal crystal.

We differ from all three by making the transmitted spectrum a runtime decision driven by a measured property of the scattering medium.

Work that constrains the design. - Defect Detection Algorithm for Wing Skin with Stiffener Based on Phased-Array Ultrasonic Imaging combines full-matrix capture, total focusing, and MVDR and coherence-factor beamforming on a GPU, reaching 1 mm sensitivity and 0.5 mm resolution in aluminium. It treats grain noise as nuisance and uses standard pulses. - Buried Defect Detection Method for a Blowout Preventer Seal Ring Groove Based on an Ultrasonic Phased Array reports simulation-to-experiment echo agreement within 1 dB, but only for EDM notches of 5 mm and larger.

These results warn us not to assume continuum simulators are poor in general; their agreement at millimetre scale is exactly what we expect to degrade with $k\bar d$.

Inverse-problem and detection work evaluated only in simulation. - Stochastic identification of imperfections in a submerged shell structure assumes a continuum forward model and Gaussian noise. - Multichannel Adaptive Signal Detection in Space-Time Colored Compound-Gaussian Autoregressive Processes proposes a detector for spiky, correlated clutter, the CG-PGLRT, which needs only two or three clutter training cells. It is tested only on synthetic K-distributed clutter.

We reuse that detector as a component and supply what it lacked: real clutter, measured on the part.

Research Program

Instrument. We will build a software-defined ultrasonic phased-array instrument with three parts:

  • Transmit: a streaming arbitrary-waveform transmitter, following the ionosonde design, feeding a linear power stage into an array.
  • Receive: a GPU pipeline modelled on SUPRA, extended to full-matrix capture and total focusing.
  • Front-end correction: the analog front end distorts the waveforms software assumes are ideal. Before any experiment, we measure the end-to-end transfer function and harmonic distortion of the amplifier and transducer chain with a hydrophone and apply predistortion. Without this step, "designed" spectra would not be the spectra actually emitted.

Sounding protocol. We first scan a reference patch of the part that is assumed crack-free. A broadband chirp is transmitted, and spatial averaging of the focused backscatter power spectrum estimates $S_g(f)$, together with depth-dependent attenuation. The crack response $H_c(f)$ cannot be measured before a crack is found. We therefore use a prior over crack depths from 50 to 500 µm, computed by a scattering model for small planar cracks. We then:

  1. Solve the design problem above for $X(f)$ under the amplitude limit.
  2. Synthesize a phase-coded waveform with that magnitude spectrum.
  3. Fit the CG-PGLRT clutter model to the measured patch rather than to an assumed distribution.

The whole loop, from sounding to redesign, should take seconds on the GPU. We report it as a time cost.

Specimens. We will manufacture S355 steel coupons, matching the grade studied in the pitting paper. Heat treatment gives three grain-size populations: as-received (about 22 µm, verified by EBSD) and two coarsened batches targeting roughly 80 and 200 µm. These batches, combined with centre frequencies from 5 to 25 MHz, span $k\bar d$ from well inside the Rayleigh scattering regime to near unity.

Fatigue cracks will be grown from electrochemically produced pits under cyclic bending. Crack depth is tracked by periodic surface replicas and established afterwards by sectioning or X-ray computed tomography, giving ground truth on 60 or more cracks per grain-size group. EDM notches of 0.2 to 1 mm in the same coupons provide a continuum-regime control.

Simulation half of the benchmark. Each coupon geometry is simulated twice:

  • with a homogeneous isotropic continuum model, the CIVA-style standard;
  • with a grain-resolved finite-element model using synthetic Voronoi microstructures with matching grain-size statistics, plus a specimen-specific variant seeded from EBSD maps where available.

We will release raw full-matrix data from both simulations and the measurements on identical geometries. This directly addresses the absence of paired simulated and physical sensor data noted in the simulation-based-evaluation direction.

Baselines

Naive reference. The do-nothing point is a conventional inspection: a fixed broadband bipolar pulse at the transducer's nominal centre frequency, delay-and-sum total focusing, and a fixed amplitude threshold set for a target false-alarm rate on crack-free patches. Every system will run on the same array, coupons, scan positions, and total transmitted energy per firing. This tells us whether short cracks are detectable at all with standard practice.

Strongest existing imaging method. This is the pipeline from the wing-skin paper: total focusing plus MVDR and coherence-factor adaptive beamforming, with its near-surface artifact suppression, on a GPU. The paper reports 1 mm detection sensitivity and 0.5 mm spatial resolution in aluminium, and an array performance indicator of 4.4898 for its plain total-focusing baseline. We found no reported result on Lacuna for detection of microstructurally small fatigue cracks below 1 mm by any method, so our beating condition is defined on our own data.

We will implement that receive chain faithfully and give it the same tuning budget as ours: grid search over its regularization and coherence parameters on a held-out calibration coupon. It beats us, or we beat it, on crack depth detected at 90% probability with 95% confidence, at matched false-alarm rate. Beating it means a statistically separated POD curve, by bootstrap over cracks, in at least the two coarser grain populations.

Strongest coded-excitation competitor. This is a fixed linear chirp spanning the transducer band with matched-filter pulse compression, again at equal energy. It is the obvious use of a software-defined transmitter that does not use the sounding step.

Ablations isolating the key ingredient. These separate the claim that the sounding step matters from the claim that the hardware is programmable:

  • Continuum-designed waveform: the same optimization run with $S_g(f)$ taken from the continuum simulation (effectively zero, so the optimizer just maximizes crack echo).
  • Grain-model-designed waveform: the same optimization with $S_g(f)$ from the grain-resolved simulation.
  • Detector ablation: the CG-PGLRT with clutter parameters fit to simulated rather than measured clutter.
  • Receive-side combination: our designed waveform combined with MVDR and coherence-factor beamforming, since transmit and receive gains may stack.

Why not just…?

Why not just use a long chirp or complementary-code excitation with pulse compression? Pulse compression trades waveform length for gain against noise that is independent of the transmission. Grain backscatter is not independent: it is generated by the same waveform and compresses along with the crack echo. The design equation predicts this baseline improves detectability only until electronic noise falls below grain clutter, then flattens.

We test this directly by reporting two separate quantities as transmit energy is increased: signal-to-electronic-noise ratio, and signal-to-grain-noise ratio. If chirps keep improving the second quantity, our central claim is wrong.

Why not just sweep the centre frequency and pick the best fixed tone burst? This is the most serious threat, because much of the benefit of spectral shaping may come from choosing one good band. We include an oracle version: an exhaustive sweep of narrowband bursts, choosing the best centre frequency per coupon using the crack ground truth.

If the sounding-designed waveform does no better than this oracle, the honest conclusion is that in-situ sounding is a practical way to find the right band without ground truth, and that multiband shaping adds nothing. We would report the contribution in exactly those reduced terms. The sounding step still has to be shown to recover the oracle's choice; a simulation-chosen frequency is the alternative it competes with.

Why not just run a better, grain-resolved simulation instead of measuring on the part? A polycrystal model with the right grain-size statistics may predict $S_g(f)$ well enough. If it does, the hardware loop is unnecessary for design. That is a positive outcome for simulation and a useful finding.

The grain-model-designed ablation settles this question. We expect it to close part but not all of the gap, for two reasons. Real S355 has texture, banding, and pearlite colonies that a Voronoi model with matched mean grain size omits. And in service, the EBSD maps needed for a specimen-specific model are unavailable.

We will report how much of the continuum-to-measured gap each simulation tier closes, so the answer is quantitative rather than argued.

Why not just improve the receive side with adaptive beamforming? MVDR and coherence-factor weighting exploit spatial incoherence. Below the scattering length, grain clutter is partly spatially coherent across a small aperture. The wing-skin results were obtained at millimetre scale in aluminium, where this matters less. The combined baseline shows whether transmit and receive improvements are additive or redundant.

Proposed Experiments and Decisive Checks

First decisive check: the reality gap. For each grain population and centre frequency, we compute the gap in decibels between simulated and measured signal-to-grain-noise ratio for each crack, under both continuum and grain-resolved simulation. We then plot it against $k\bar d$. The continuum-regime EDM notches anchor the plot: they should reproduce agreement of roughly 1 dB, as in the blowout-preventer study, and if they do not, the instrument rather than the physics is at fault. The prediction is a continuum gap that rises steadily with $k\bar d$ and a grain-resolved gap that rises more slowly.

Second decisive check: probability of detection. We produce probability-of-detection curves versus true crack depth for all systems at a fixed false-alarm rate, with the 90/95 crack depth as the headline metric. The claim survives only if all three conditions hold:

  • the sounding-designed system beats the chirp and the continuum-designed waveform in the coarse-grained populations;
  • it at least matches the oracle fixed burst;
  • the measured-clutter detector beats the simulated-clutter detector in false alarms at equal detection.

Third check: transfer of simulation-tuned detectors. We report how false-alarm rates of detectors calibrated on simulated clutter inflate on real data. This gives a measured counterpart to the purely synthetic CG-PGLRT evaluation.

Fourth check: practicality. We report sounding-plus-redesign latency on the GPU and the area of reference patch needed for a stable $S_g(f)$ estimate. If a usable estimate needs a patch larger than the component, the method is a laboratory tool rather than an inspection tool, and we will state so.