🔬 applied physics

Identifiability Limits in Ultrasonic Microstructure Characterisation: A Canonical and Stochastic Framework

This paper establishes that the identifiability of ultrasonic microstructure parameters is fundamentally constrained not just by the structural rank of the forward operator, but primarily by the anisotropic sensitivity geometry and the balance between sensitivity magnitude and intrinsic stochastic variability, thereby dictating the necessity of combining complementary observables for robust characterization.

Original authors: Wei Yi Yeoh

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wei Yi Yeoh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out what's inside a sealed, opaque box by tapping on it and listening to the sound it makes. This is essentially what scientists do when they use ultrasound to look inside metal materials without cutting them open. They send sound waves in, and the way those waves bounce back (echoes) or slow down tells them about the tiny crystal grains inside the metal.

However, there's a catch: just because you can hear a sound doesn't mean you can uniquely identify exactly what caused it. This paper asks a fundamental question: "Is the information in the sound wave actually enough to solve the puzzle, or are we just guessing?"

The authors use two different "test boxes" to figure this out.

1. The Simple Test: The "Echo Chamber" (Canonical Model)

First, they looked at a very simple, perfect block of metal (like a solid titanium slab) sitting in water. They treated it like a basic echo chamber.

The Problem: They tried to measure four things (density, thickness, how much the sound fades, and how hard the sound bounces off the surface) using only three "clues" (how long the echo took, how loud the echoes were, and how much the sound bounced off the front).
The Analogy: Imagine trying to guess the exact weight and height of a person, but you only have a scale that gives you a combined "weight-height" number. You can't separate the two.
The Finding: Because they had more unknowns than clues, the math broke down. Even with perfect, noise-free data, some combinations of properties were impossible to tell apart. It's like trying to solve a math equation with too many variables and not enough numbers. They also found that if the metal is too similar to the water (or too different), the "clues" become too weak or too strong to be useful, making the puzzle unsolvable.

2. The Complex Test: The "Noisy Crowd" (Surrogate Microstructure)

Next, they moved to a more realistic scenario. Real metal isn't a perfect block; it's a messy crowd of tiny crystals pointing in different directions. They created a computer model of this "messy crowd" using two main settings:

Correlation Length ( $D$ ): How big the "clumps" of similar crystals are.
Texture Coherence ( $T$ ): How orderly or chaotic the crystal directions are.

They measured two things: how fast the sound traveled (velocity) and how much it faded away (attenuation).

The Good News: Unlike the simple block, they had two clues for two unknowns. Mathematically, this should be solvable. The "map" from the metal's properties to the sound was full of information.
The Bad News (The "Imbalance"): The clues weren't equal.
- Velocity was very sensitive to how orderly the crystals were ( $T$ ) but barely cared about the size of the clumps ( $D$ ).
- Attenuation was very sensitive to the size of the clumps ( $D$ ) but also cared a bit about orderliness.
- The Metaphor: Imagine trying to guess the ingredients in a soup by tasting it. If the saltiness tells you exactly how much salt is in there, but the spiciness tells you almost nothing about the pepper, you can't figure out the pepper amount just by tasting. One clue is loud and clear; the other is a whisper.
The "Noise" Factor: Real materials are random. Even if you make two identical-looking metal blocks, the tiny crystals inside will be slightly different. This creates "background noise" in the sound.
- The authors found that when this natural randomness is taken into account, the "whisper" (the sensitivity to clump size) gets drowned out by the noise. Even though the math says the answer exists, the noise makes it impossible to find in the real world.

3. The Solution: Combining Clues

The paper tested what happens if you use just one clue versus using both.

Using only Velocity: The answer was a long, stretched-out guess. You knew the "orderliness" but had no idea about the "clump size."
Using only Attenuation: You knew the "clump size" but had no idea about the "orderliness."
Using Both: When they combined the two clues, the "stretched-out" guess snapped into a tight, precise point. Because the two clues looked at the problem from different angles, they covered each other's weaknesses.

The Big Takeaway

The main lesson of this paper is that having a signal isn't enough; you need the right kind of signal.

Don't just count your clues: It's not enough to have the same number of measurements as unknowns. The measurements must be "balanced" (sensitive to all parts of the problem).
Watch out for the noise: Even if the math works perfectly, the natural randomness of the material can hide the answer if the signal is too weak.
Mix your methods: To get a clear picture, you need to combine different types of measurements (like speed and fading) that complement each other.

In short, the authors built a framework to tell engineers: "Before you try to reverse-engineer a material's structure, check if your ultrasound setup is actually capable of hearing the difference. If the clues are too weak or too similar, no amount of fancy computer math will fix it."

Technical Summary: Identifiability Limits in Ultrasonic Microstructure Characterisation

Problem Statement
Ultrasonic microstructure characterisation relies on inverting measured wave responses (attenuation, velocity, backscatter) to infer material descriptors such as grain size, texture, and heterogeneity. While inversion methods and machine learning have advanced, a fundamental question remains: do the measured wavefields contain sufficient independent information to uniquely recover the target descriptors? Often, ambiguities in inferred properties are attributed to inversion algorithms, yet they may stem from intrinsic non-uniqueness in the forward mapping. This work addresses the "identifiability" of ultrasonic measurements, distinguishing between limitations caused by the inversion method and those inherent to the information content of the measurement itself.

Methodology
The study employs a unified framework analyzing the geometry of the forward operator across two distinct modeling levels:

Canonical Pulse-Echo Model: A simplified, one-dimensional homogeneous model (titanium block in water) is used to isolate fundamental mechanisms. The forward map relates four parameters (density $\rho$ , thickness $d$ , attenuation $\alpha$ , impedance $Z$ ) to three observables (time-of-flight, echo amplitude ratios, front-face reflection). Identifiability is assessed via analytical Jacobian analysis, singular value decomposition (SVD), and condition number evaluation to quantify rank deficiency, parameter coupling, and interface-driven saturation.
Stochastic Surrogate Microstructures: To address heterogeneity, the framework extends to two-dimensional media modeled as Gaussian Random Fields (GRFs). The microstructure is parameterized by a correlation length ( $D$ $D$ ) and a texture-coherence parameter ( $T$ $T$ ). The forward map predicts mean attenuation and velocity.
- Structural Identifiability: Analyzed using the Jacobian of the ensemble-mean feature map to determine if the system is full rank and to assess sensitivity anisotropy.
- Practical Identifiability: Extended via a variance-weighted Fisher Information framework. This incorporates the intrinsic covariance of the observables arising from stochastic microstructure realisations. Sensitivity is normalized by variability to determine if parameter changes are distinguishable from natural microstructural fluctuations.
- Validation: Inversion experiments using grid-based search methods on the objective function landscape verify the theoretical predictions regarding constraint and recovery.

Key Contributions

Geometric Baseline for Identifiability: The paper establishes a geometric baseline for identifiability limits using a canonical model, revealing that non-uniqueness arises not only from dimensional underdetermination but also from severe parameter coupling and impedance-dependent saturation.
Anisotropic Sensitivity in Heterogeneous Media: It demonstrates that even when a surrogate forward map is structurally full rank (number of observables equals number of parameters), practical identifiability is governed by strong anisotropy. Sensitivity is unevenly distributed, with one parameter direction (texture-coherence) often dominating the other (correlation length).
Variance-Weighted Fisher Framework: The work introduces a variance-weighted approach to practical identifiability. It shows that recoverability is not determined by structural rank alone but by the balance between sensitivity magnitude and intrinsic stochastic variability. When variability is high, the weakest parameter direction becomes effectively unrecoverable, even if the mean map is theoretically invertible.

Results

Canonical Model: The analysis reveals that the canonical system is rank-deficient (4 parameters, 3 observables), leading to a near-null direction caused by the coupling of density and thickness. Even in reduced, full-rank configurations, the condition number is highly sensitive to impedance contrast. At high contrasts, reflection saturation reduces sensitivity, elongating the information geometry and destabilizing inversion.
Surrogate Model (Structural): The map from $(D, T)$ to $(\alpha, v)$ is structurally full rank. However, the singular value spectrum shows significant anisotropy; velocity is highly sensitive to $T$ but weakly to $D$ , while attenuation responds to both but is dominated by $D$ . This creates an imbalanced information geometry where one direction is inherently weaker.
Surrogate Model (Practical): When stochastic variability is incorporated, the "whitened" Jacobian reveals that the weaker parameter direction ( $D$ ) is further suppressed in regimes of high scattering variability (low $T$ , high $D$ ). The practical identifiability index confirms that $T$ primarily governs whether the system remains resolvable.
Inversion Behavior: Inversion landscapes confirm the theoretical analysis. Using a single observable (either attenuation or velocity) results in elongated, weakly constrained objective functions with poor localisation. Combining both observables yields complementary constraints, resulting in a compact, well-localised minimum at the ground truth.

Significance and Claims
The paper argues that the limits of ultrasonic inversion are governed primarily by the geometric structure of the forward map and the information content available in the measurement, rather than by the sophistication of the inversion algorithm.

Measurement Design: The results imply that measurement design should be guided by the local information geometry (sensitivity balance and variance) rather than signal amplitude alone. Multi-observable strategies are necessary to constrain complementary parameter directions.
Machine Learning Context: The authors claim that data-driven methods cannot recover information that is weakly expressed in the measurement or suppressed by intrinsic variability. Increasing model complexity or training data does not overcome fundamental identifiability limits imposed by the forward map.
Diagnostic Utility: The proposed framework serves as a diagnostic tool to distinguish between structural non-identifiability (rank deficiency), poor conditioning (anisotropy), and variability-limited resolution, providing a systematic way to evaluate observable selection before attempting inversion.

The study concludes that within the feature-level framework considered, identifiability is a property of the measurement physics and stochastic nature of the material, necessitating a shift from purely algorithmic improvements to a deeper understanding of the information geometry of ultrasonic responses.

1. The Simple Test: The "Echo Chamber" (Canonical Model)

2. The Complex Test: The "Noisy Crowd" (Surrogate Microstructure)

3. The Solution: Combining Clues

The Big Takeaway

More like this