How Far Ahead Can a Creep Test See? Truncated-SVD Identifiability Horizons for Long-Term Creep Extrapolation

Engineers run creep tests for hours or weeks and then predict how a material, a pipe in soft soil or a boiler steel, will slowly deform over decades. Creep here means the slow, continuing deformation of a material held under a constant load. The standard approach fits the measured curve with a sum of decaying exponentials, each standing for a spring-and-dashpot element with its own characteristic time, and then evaluates that sum far beyond the test. The fit is numerically ill-posed. Exponentials whose characteristic times are close together are nearly indistinguishable, and exponentials much slower than the test duration barely change within the measured window. Two fits can agree on the data to within the noise and still disagree badly at year twenty.

We propose to treat this as a precise linear-algebra question. Fix a dense grid of candidate characteristic times, which turns the problem into an overcomplete linear system. Regularize the system by truncated singular value decomposition (TSVD): break the system into independent directions ordered by how strongly the data constrain them, and keep only the directions the data actually determine. Then, for every future time $t$, measure how much of the prediction depends on the directions that were thrown away. Because the material's spectrum is nonnegative and its total compliance is physically bounded, this quantity becomes a computable worst-case error bar. The time at which that bar exceeds a tolerance is the test's identifiability horizon: the point beyond which the test data, however carefully fitted, cannot determine the answer. The main deliverable is this horizon, which tells a practitioner whether a 1,000-hour test can support a 100,000-hour prediction. A better fit is secondary.

Core Claim

The research file finds no published work that applies truncated-SVD regularization to the inversion of creep or viscoelastic kernels. The nearest uses of SVD in the creep literature are offline principal-component compression of tabular alloy features before regression (Prediction of High-Temperature Creep Life of Austenitic Heat-Resistant Steels Based on Data Fusion). The nearest operator work learns a low-rank factorization of PDE integral kernels by gradient descent (SVD-NO). Neither touches the conditioning of the hereditary operator, nor asks how far a fit can be trusted.

Our claim is that truncation does more than stabilize the fit. The truncation threshold $\tau$ sets an explicit trade-off between how much noise is amplified and how far ahead in time the prediction is still determined by the data. Combined with a nonnegative, mass-bounded spectrum, this yields per-time error certificates that actually contain the true long-term error. Fixed-branch Prony fits, ridge regression and black-box surrogates do not produce such certificates. The trade-off is the temporal analogue of the precision–resolution trade-off that Parameter selection and numerical approximation properties of Fourier extensions from fixed data establishes for Fourier extensions.

Concretely, write creep compliance as $J(t) = J_0 + \sum_{j=1}^{n} h_j (1 - e^{-t/\tau_j})$. Here $J_0$ is the instantaneous compliance, $\tau_j$ are log-spaced candidate characteristic times spanning far past the test, and $h_j \ge 0$ are spectral weights.

Test data at times $t_1,\dots,t_m \le T$ give a system $Ah = b + \delta b$, where:

  • $A_{ij} = 1 - e^{-t_i/\tau_j}$;
  • $b$ is the measured compliance minus $J_0$;
  • $\delta b$ is noise of size at most $\delta$.

Let $a(t)$ be the row vector that evaluates $J(t) - J_0$ at a future time $t$. Let $A_k^+$ be the rank-$k$ truncated pseudoinverse, and $P_k^\perp$ the projector onto the discarded right singular vectors. Then

$$|J(t) - \hat J_k(t)| \;\le\; \delta\,\|(A_k^+)^T a(t)\|_2 \;+\; \|P_k^\perp a(t)\|_\infty \,\|h\|_1 .$$

The first term is noise amplified into the prediction. The second is the part of the prediction the data cannot see. Nonnegativity of $h$ gives $\|h\|_1 = J_\infty - J_0$, the bounded equilibrium-minus-instantaneous compliance. That quantity can be bounded physically, or from the longest available test.

Minimizing the bound over $k$ at each $t$ gives the certificate. The horizon $t^\ast(\varepsilon)$ is the smallest $t$ at which the minimized bound exceeds a tolerance $\varepsilon$. The bound is elementary, but as far as the file shows it has never been computed or validated for creep.

Research Program

The method side builds on three lines of approximation theory, each in a different role. Frames and Numerical Approximation and Stable and Accurate Least Squares Radial Basis Function Approximations on Bounded Domains prove that truncating singular values of redundant, exponentially ill-conditioned systems keeps approximation power while restoring stability. Those results supply the method and justify a deliberately dense $\tau$-grid in place of a hand-picked handful of elements. Prony systems via decimation and homotopy continuation shows that near-colliding exponentials break standard spectral estimators. It motivates the problem, and it also constrains it: we do not try to recover individual close characteristic times. We certify only functionals of the spectrum, namely future compliance values, which can remain well determined even when individual times are not.

The work has four parts. First, we derive the bound above with the nonnegativity constraint, and study the shape of $t^\ast$ as a function of test duration $T$, noise $\delta$ and grid density. Our expectation is that $t^\ast/T$ grows roughly with $\log(1/\delta)$, because the slowest resolvable exponential contributes a signal of order $T/\tau$ inside the window.

Second, we compute a nonnegative estimate restricted to the retained subspace, as a projected least-squares problem. This keeps the fitted spectrum physical; the file notes that unregularized multi-branch fits typically return alternating-sign, thermodynamically inadmissible moduli.

Third, we extend the bound to tests at several temperatures stitched together by time–temperature shifting. Each temperature adds rows to $A$, and the certificate measures how much each extra temperature extends the horizon. This turns the method into a design tool for choosing which test to run next.

Fourth, we treat the anisotropic case using the relaxation-tensor structure derived in Anisotropic Extended Burgers Model, Its Relaxation Tensor, and Properties of the Associated Boltzmann Viscoelastic System, with one certificate per tensor component.

We deliberately exclude near-instability regimes. The file identifies a real tension in creep buckling, where the mode that drives collapse lives in the near-zero singular subspace that truncation would discard (A Physics-Informed Neural Network Framework for Simulating Creep Buckling in Growing Viscoelastic Biological Tissues). Our truncation acts on the data-to-spectrum map, not on the structural stiffness operator, so it does not filter buckling modes. We state this scope limit explicitly rather than claim the method covers buckling.

Baselines

The naive reference is an unregularized fit of a small fixed number of Maxwell/Kelvin elements by nonlinear least squares, extrapolated directly. It represents common practice. The file's discussion of exponential ill-conditioning (Frames and Numerical Approximation) predicts that it will look fine inside the window and scatter widely outside it. It will be run with the number of elements chosen by in-window cross-validation. Beating it means lower median error at $10T$ and $100T$ with no loss of in-window fit.

The strongest competitor is nonnegative least squares on the same dense $\tau$-grid with Tikhonov (ridge) regularization, using a penalty chosen by generalized cross-validation or the discrepancy principle. It is strong because nonnegativity alone already regularizes heavily, and ridge is the default stabilizer. LoRanPAC reports that explicit truncation beat ridge-regularized RanPAC by over 25% on ImageNet-A in the one-class-at-a-time setting. That result is in continual learning, not creep, so it motivates the comparison but does not predict its outcome. There is no reported result found on Lacuna for ridge or nonnegative least-squares inversion of creep spectra.

This baseline gets the identical grid, data and noise model, and the same hyperparameter search budget as our truncation level. We also give it the analogous filter-factor certificate. Beating it means either lower long-horizon error, or, more importantly, error bars with nominal coverage where its bars under-cover. If the two are tied on both counts, the contribution reduces to the certificate, which holds for either filter, and we will report it that way.

For discrete, well-separated spectra we add an ESPRIT/Prony estimator, with decimation as described in Prony systems via decimation and homotopy continuation. Two ablations isolate the key ingredients. The first is truncation without the nonnegativity bound, which tests whether the $\|h\|_1$ term is what makes the certificate finite. The second fixes $k$ by the discrepancy principle instead of choosing it per time $t$, which tests whether per-horizon rank selection matters.

Why not just…?

Why not just use ridge regression with nonnegativity, as rheologists already do? This is the real threat, and the proposal is designed so that losing to it on point accuracy does not sink the project. The certificate is the novel part. It applies to any spectral filter, including ridge filter factors, and no paper in the file computes or validates horizon certificates for creep. The head-to-head comparison in the experiments settles whether hard truncation also gives tighter bars. LoRanPAC's sensitivity of ridge to $\lambda$ over long sequences is suggestive, but it comes from a different domain and does not settle the question here.

Why not just use the Larson–Miller parameter or a Gaussian-process surrogate? These predict a scalar rupture life, not a compliance curve. Even at that task, the parametric method reaches only $R^2 = 0.752$ on S30432 (Prediction of High-Temperature Creep Life of Austenitic Heat-Resistant Steels Based on Data Fusion). The Gaussian-process result of $R^2 = 0.989$ on modified 9Cr-1Mo (Machine Learning-Based Framework for Predicting Creep Rupture Life of Modified 9Cr-1Mo Steel) comes from cross-validated splits. On the 21-point external set its RMSE rises to 0.279. Neither result says whether accuracy survives when the model is trained only on short tests and evaluated on long ones. Our rupture-life experiment below runs exactly that duration-split test.

Why not just train a neural operator or fine-tune a PDE foundation model? FUsion-based ConstitutivE model (FuCe) documents that purely data-driven constitutive models fail outside the training envelope unless physical structure is imposed, and it leaves time-dependent inelasticity unaddressed. Foundation-model transfer to solids gave at most about a twofold gain over training from scratch (Out-of-distribution transfer of PDE foundation models to material dynamics under extreme loading). SVD-NO learns a low-rank kernel but provides no out-of-window guarantee. We include an SVD-NO-style learned kernel on the synthetic suite to measure this directly.

Proposed Experiments

The file contains no public viscoelastic creep-curve benchmark, so the primary testbed is synthetic with known ground truth. Spectra are drawn from three families:

Each spectrum is sampled to a test duration $T$ between $10^2$ and $10^3$ hours, with heteroscedastic and colored noise, and evaluated out to $10^5$ hours. The metrics are:

  • relative compliance error at $10T$, $100T$ and $1000T$;
  • empirical coverage of the certified error bars, where we expect at least nominal coverage;
  • the accuracy of the predicted horizon $t^\ast$ against the true time at which error crosses $\varepsilon$.

The decisive check is coverage. If the certificate under-covers on colored noise or clustered spectra, the core claim fails regardless of point accuracy.

A second experiment uses multi-temperature synthetic data and verifies the predicted horizon gain from each added temperature. We then validate those predictions by actually adding the temperatures.

A third, real-data experiment reuses the modified 9Cr-1Mo NIMS-derived set (431 points plus 21 external) and the six-grade austenitic set (1060 points). We split them by rupture duration rather than at random: train on tests shorter than $10^3$ hours, test on longer ones. We fit a nonnegative, truncation-regularized Larson–Miller-type master curve and report RMSE in log rupture life alongside its certified band. The reported random-split Gaussian-process numbers serve as context, not as matched baselines, and we rerun the Gaussian process under our duration split.

The outcome that would most change practice is simple to state. The certificate would show, before any long test is run, that a given short-test campaign cannot determine compliance beyond a specific horizon, and longer tests would confirm that the uncertified extrapolations are where the errors occur.