Source-linked AI summary
Physics-informed learning for the inverse problem in resonant ultrasound spectroscopy
Alejandro Cubillos Muñoz, Manuela Rivas, Julian Rincon
TL;DR
Inferring elastic constants from finite RUS spectra is a nonlinear constrained inverse-isospectral problem because measured frequencies must correspond to symmetry- and stability-admissible tensors. The paper uses physics-adapted geometric, spectral, and elastic-ratio variables so regression predicts reduced ratios while scale recovery and final reconstruction remain analytic. Benchmarks report consistent reconstruction with few-percent fixed-geometry errors, while the study remains limited to isotropic and cubic synthetic spectra.
Problem
RUS inversion uses finite resonance data to recover elastic tensors despite nonlinear frequency dependence, spectral nonuniqueness, and symmetry and stability constraints.
Method
The pipeline separates size and elastic scales, aspect ratios, scale-free spectral features, and stability-respecting elastic ratios before feed-forward regression and analytic reconstruction.
Results
Fixed-geometry benchmarks report cubic MAPE values of 4.14(3.87)%, 8.31(8.50)%, and 2.44(2.86)%, and isotropic values of 4.0(3.6)% and 0.4(0.3)%.
Takeaways & Limitations
The inverse problem becomes constrained regression in variables adapted to geometry, scaling, crystal symmetry, and thermodynamic stability.
Takeaways & Limitations
The study is limited to isotropic and cubic elasticity and synthetic spectra generated with the same Rayleigh-Ritz model used for training.
Abstract
from arXiv · showhide
Inferring elastic constants from resonant ultrasound spectra is a nonlinear and typically overdetermined inverse problem based on finite spectral data. We formulate the Rayleigh-Ritz inverse problem as a constrained inverse-isospectral problem on the set of physically admissible elasticity tensors. This induces effective low-dimensional variables for the inverse map on the admissible elasticity manifold: length and elastic scales, aspect-ratio coordinates, scale-free spectral features, and stability-respecting elastic ratios. We use these variables to construct a physics-informed learning pipeline in which a regression model acts only on reduced spectral and geometric features, while scale recovery and final elastic-constant reconstruction are imposed analytically. For the full cubic benchmark, the reconstructed constants have MAE values of $20.37(35.15)$, $24.30(41.33)$, and $2.13(3.66)~\mathrm{GPa}$ for $C_{11}$, $C_{12}$, and $C_{44}$. In the fixed-geometry benchmark, the corresponding cubic MAPE values are $4.14(3.87)\%$, $8.31(8.50)\%$, and $2.44(2.86)\%$, while the isotropic values are $4.0(3.6)\%$ and $0.4(0.3)\%$ for the bulk and shear moduli. The inverse problem then becomes a constrained regression problem in variables adapted to the geometry, scaling, crystal symmetry, and thermodynamic stability of Hookean elasticity.
I. INTRODUCTION
RUS infers elastic constants from finite resonance data through a nonlinear forward problem whose admissible solutions are constrained by geometry, symmetry, and stability. The paper motivates physics-adapted variables so learning handles a reduced inverse map rather than rediscovering known structure.
- Elastic constants characterize mechanical response and connect to microscopic structure, phase transitions, anharmonicity, and couplings to other excitations.
- RUS recovers elastic constants from vibration frequencies, with Rayleigh-Ritz reducing the forward calculation to a finite-dimensional generalized eigenvalue problem.
- Finite spectra, uncertain peaks, absent mode shapes, crystal symmetry, and stability constraints make inversion a constrained inverse-isospectral problem rather than direct inversion.
- Raw regression obscures distinct effects of size, aspect ratio, elastic scale, and stability while forcing models to rediscover correlations induced by shared elastic constants.
- The proposed coordinates separate size, aspect ratio, spectral scale, and elastic magnitude from stability-respecting ratios, leaving regression to learn a reduced spectral-shape map.
- A. Spring-mass analogs for RUS problems: Spring-mass analogs distinguish overdetermined, determined, and underdetermined inverse problems according to parameter constraints and available eigenvalue information.
B. Forward problem: elastic constants to resonances
The Rayleigh-Ritz forward problem maps density, geometry, elasticity, and symmetry to resonance frequencies through finite-dimensional mass and stiffness matrices. Removing rigid-body modes and normalizing scale exposes a reduced dependence on shape and elastic parameters.
- The forward problem computes resonances from density, geometry, elasticity tensor, and crystal symmetry for a Hookean body.
- Rayleigh-Ritz restricts displacement fields to a finite variational subspace, converting continuum elasticity into a finite-dimensional generalized eigenvalue problem.
- The kinetic and elastic energies generate a positive-definite mass matrix E(ρ,L) and a stiffness matrix Γ(C,L), with stability making Γ positive semidefinite.
- After rigid-body zero modes are removed, the positive eigenvalues provide squared resonance frequencies; homogeneous density factors from the mass matrix as E(ρ,L)=ρE(L).
- The selected forward spectrum is the positive-eigenvalue map C 7→ν = spec+,nω[AN(C,ρ,L)], with nω determined by the experimental window or protocol.
- Separating an overall length scale R from aspect-ratio coordinates (η,β) and using size-normalized spectra removes trivial scaling while retaining dimensionless geometric dependence.
C. Inverse problem: resonances to elastic constants
RUS inversion searches for physically admissible elasticity tensors whose predicted finite spectrum matches measured resonances. The paper frames this as constrained inverse-isospectral reconstruction and implements it with reduced features, ratio prediction, and analytic scale recovery.
- Measured resonances constrain the positive spectrum of a Rayleigh-Ritz operator, while symmetry and thermodynamic stability restrict tensors to a physically admissible set.
- A complete spectrum permits an isospectral orbit, but finite measurements provide spectral constraints within a low-dimensional physical elasticity manifold.
- The physical domain is not a vector space: stable tensors obey positivity inequalities, and arbitrary linear combinations or sign changes need not remain stable.
- The inverse reconstruction need not be unique or exact: noisy data may have an empty exact preimage, while multiple admissible tensors may match the measurements.
- The physics-informed pipeline separates elastic-ratio prediction from analytic scale recovery using the measured size-normalized spectrum.
- The inverse procedure selects an admissible tensor by comparing computed and measured spectra, using a weighted least-squares or finite-multiset distance when correspondence is uncertain.
A. Spectral redundancy and reduced coordinates
Raw spectra contain strong redundancy from common elastic scaling, sample size, geometry, and shared dependence on a small number of elastic constants. The reduction replaces these correlations with shape coordinates and scale-free spectral features that retain complementary information.
- Raw isotropic inputs show strong correlations among squared-frequency features, indicating substantial redundant spectral information.
- A common elastic scale rescales the spectrum coherently, while sample size changes the boundary-value problem and shared elastic dependence further correlates resonances.
- Size normalization removes trivial uniform dimension scaling, while R, η, and β separate overall size from two dimensionless aspect-ratio coordinates.
- Ordering side lengths makes each aspect ratio unique and removes permutation redundancy from the geometric representation.
- For isotropic benchmarks, six resonances yield five ratios ξ0,...,ξ4 that depend on G/K at fixed shape but not on overall elastic magnitude.
- The reduced isotropic ratios occupy different ranges and respond differently to G/K, preserving complementary spectral-shape information after scale removal.
- For cubic benchmarks, twenty spacings χ0,...,χ19 retain local spectral shape while removing common scale, although neither feature set is statistically independent.
B. Stability-respecting parameterization and inverse pipeline
The pipeline encodes stability-respecting elastic ratios as bounded angular variables, learns only the reduced inverse map, and recovers scale and elastic constants analytically. This parameterization spans the admissible cubic target space while separating geometry, spectral shape, and elastic magnitude.
- Stability-respecting target variables: Bounded angular variables encode stability-respecting elastic ratios, so the model need not learn the admissible stability domain directly.The absolute elastic scale is recovered analytically from the normalized spectrum.
- Inverse pipeline: The cubic network predicts only (ϕK, ϕa), while a reference forward calculation determines the magnitude M and analytic formulas reconstruct C11, C12, and C44.This separates elastic ratios from overall magnitude in the inverse pipeline.
- Stability-respecting target variables: Positive variables (K, a, µ) map the allowed cubic target space to the positive octant and make the Born stability conditions explicit.The original constants are reconstructed as C11 = K + 2a, C12 = K −a, and C44 = µ.
- Benchmark scope: The full cubic benchmark permits varying sample dimensions and stable cubic elastic constants, with C44 showing the strongest parity-plot agreement and C12 the broadest spread.The broader C12 distribution reflects sensitivity of C12 = K −a to propagated errors in K and a.
- Inverse pipeline: Physics-informed preprocessing transforms raw data into size-normalized spectral variables, scale-free features, and shape coordinates before regression.The transformations reduce the inverse problem before the regression model is introduced.
IV. RESULTS AND DISCUSSION
The reduced angular maps are learnable, and final reconstruction is evaluated across complementary full and fixed-geometry benchmarks. Full cubic errors are reported in absolute units, while fixed geometry yields smaller percentage errors because aspect-ratio dependence is removed.
- Reduced inverse maps: The reduced angular maps achieve test MAE values of approximately 0.024 for ϕK and 0.028 for (ϕK, ϕa), with RMSE values of 0.040 and 0.062.These intermediate metrics assess learnability of the reduced inverse maps; final elasticity reconstruction is the experimentally relevant benchmark.
- Full cubic benchmark: 20.37(35.15) GPa, 24.30(41.33) GPa, and 2.13(3.66) GPa are the full cubic MAE values for C11, C12, and C44, respectively.This benchmark varies sample dimensions and spans the thermodynamically stable cubic domain.
- Benchmark comparison: Fixed geometry produces smaller percentage errors because it removes aspect-ratio dependence from the inverse map.The full benchmark tests parameterization scope, whereas the fixed-geometry benchmark provides a controlled comparison.
- Benchmark comparison: The approach reaches few-percent errors using scalar spectral features and direct regression without image encoding or separate classification and regression stages.The reported pipeline is described as parsimonious in the benchmark discussion.
B. Physical interpretation and limitations
The physics-informed pipeline separates geometric, spectral, and elastic scales before learning, while benchmark behavior reflects geometry, ratio structure, and cancellation-sensitive reconstruction. Fixed geometry improves percentage errors, but lower-symmetry and experimental settings remain outside the demonstrated scope.
- Benchmark behavior: 4.0(3.6)% and 0.4(0.3)% are the fixed-geometry isotropic MAPE values for K and G, respectively.Both fixed-geometry panels use the controlled 3:4:5 aspect ratio.
- Physical interpretation: C44 = µ is reconstructed most accurately, whereas C12 = K − a inherits larger errors through cancellation-sensitive reconstruction.C11 = K + 2a also combines reconstructed variables, but C12 is especially sensitive because it can be small or negative.
- Limitations: Changing side-length ratios shifts modes together, can produce near-degeneracies, and may reduce sensitivity to particular elastic combinations.Separating overall size from aspect ratio does not remove shape dependence from the full benchmark.
- Physics-informed reduction: The pipeline separates sample size, aspect ratio, spectral scale, and stability-respecting elastic ratios before regression.The network predicts reduced spectral-to-ratio maps, while scale recovery and final constants are imposed analytically.
- Benchmark behavior: 4.14(3.87)%, 8.31(8.50)%, and 2.44(2.86)% are the fixed-geometry cubic MAPE values for C11, C12, and C44, respectively.The fixed-geometry benchmark removes aspect-ratio dependence, yielding smaller percentage errors than the full benchmark.
- Limitations: The study is limited to isotropic and cubic elasticity and synthetic spectra generated with the same Rayleigh-Ritz model used for training.Lower-symmetry crystals require additional stability-respecting coordinates, while experiments must address mode assignments, overlap, damping, dimensional errors, and model mismatch.
Appendix A: Rayleigh-Ritz formulation, stability, and isospectral geometry
The appendix constructs a finite-dimensional Rayleigh-Ritz representation of free elastic vibrations and recasts RUS inversion as a spectrum-constrained search over stable, symmetry-restricted elasticity tensors.
- Rayleigh-Ritz formulation: The Rayleigh-Ritz method restricts displacement fields to a finite-dimensional trial space, producing finitely many eigenvalues including rigid-body zero modes.The resulting mass and stiffness matrices define the finite-dimensional vibration problem.
- Stability: The mass matrix is symmetric positive definite, while the elastic matrix is symmetric positive semidefinite for thermodynamically stable elasticity tensors.The exact stability inequalities depend on the crystal symmetry class.
- Rayleigh-Ritz formulation: The Euler-Lagrange equations yield the generalized eigenvalue problem Γ(C)a = ω^2E(ρ)a, whose positive eigenvalues provide the physical RUS frequencies.Rigid-body translations and rotations generate zero eigenvalues under free boundary conditions.
- Isospectral geometry: A mass-weighted transformation converts the generalized problem into a Euclidean symmetric representative without changing its generalized eigenvalues.This permits ordinary orthogonal diagonalization in the space of real symmetric matrices.
- Isospectral geometry: A spectrum alone identifies an isospectral orbit rather than a unique symmetric matrix, whereas physical RUS inversion searches only the elasticity manifold generated by stable tensors.The admissible search is therefore constrained by symmetry and thermodynamic stability rather than spanning arbitrary elastic constants.
Appendix B: Supplementary data and benchmark information
Appendix B documents the synthetic-data domain, polynomial Rayleigh-Ritz implementation, reduced spectral diagnostics, and supplementary benchmark evaluations.
- Supplementary data: Synthetic datasets sample geometries, densities, and stable elastic constants over prescribed ranges before solving the Rayleigh-Ritz problem and retaining positive resonances.These ranges define the supervised computational domain rather than a complete classification of possible materials.
- Rayleigh-Ritz implementation: The numerical Rayleigh-Ritz space uses polynomial basis functions in normalized coordinates, with truncation degree Ng determining the finite-dimensional basis.The resulting vector-valued space has dimension dN = 3Nb, and positive eigenvalues are converted into reduced spectral variables.
- Spectral diagnostics: Figure 6 reports the logarithmic frequency distribution, whose departure from normality supports transforming spectral variables before regression.The diagnostic concerns the generated resonance frequencies used in the synthetic data.
- Spectral diagnostics: The reduced cubic-feature correlation matrix shows substantially weaker pairwise correlations than the raw spectral representation.The reduced variables include normalized spacings and aspect-ratio coordinates used by the cubic angular model.
- Benchmark information: The isotropic dataset contains 252474 entries, while the cubic dataset contains 858387 entries, with separate train, validation, and test splits for angular regression.The isotropic model targets ϕK, whereas the cubic model targets (ϕK, ϕa).
- Benchmark information: Full-benchmark percentage-error distributions remain aspect-ratio dependent, while fixed-geometry diagnostics remove that dependence and are smaller in percentage terms.Relative-error interpretation requires care when the target modulus, especially C12, is small.