pith. machine review for the scientific record. sign in

arxiv: 2604.15181 · v1 · submitted 2026-04-16 · 💻 cs.LG · math.DS

Recognition: unknown

One-shot learning for the complex dynamical behaviors of weakly nonlinear forced oscillators

Teng Ma , Luca Rosafalco , Wei Cui , Lin Zhao , Attilio Frangi
Authors on Pith no claims yet

Pith reviewed 2026-05-10 11:11 UTC · model grok-4.3

classification 💻 cs.LG math.DS
keywords one-shot learningsparse identificationnonlinear dynamicsMEMS resonatorsfrequency response curvesjump phenomenaweakly nonlinear oscillatorsharmonic balance
0
0 comments X

The pith

A model trained on one excitation time history predicts softening, hardening and jump phenomena across wide excitation ranges in weakly nonlinear oscillators.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a one-shot learning method to recover the global frequency-response behavior of weakly nonlinear forced oscillators from a single excitation trajectory. It does so by first applying the Generalized Harmonic Balance method to convert the multi-frequency forced response into a set of slow-varying evolution equations for amplitudes and phases, then using an evolutionary sparse regression procedure to identify the coefficients of those equations. Once the governing slow equations are learned from data at one excitation level, the model extrapolates to forecast the entire frequency-response curve, including bistable regions and discontinuous jumps, at many other levels. The approach is shown to work on two MEMS devices: a nonlinear beam resonator and a micromirror. This removes the need for dense experimental sampling when mapping nonlinear microsystem dynamics.

Core claim

By decomposing the forced response of a weakly nonlinear oscillator into slow-varying amplitude and phase equations via the Generalized Harmonic Balance method and then applying evolutionary sparse identification to those equations, the complete nonlinear model can be recovered from a single forced time history. The recovered model then reproduces the full frequency-response curve, including amplitude-dependent frequency shifts and jump discontinuities, for excitation levels different from the training point.

What carries the argument

MEv-SINDy, which performs sparse regression on the slow-varying differential equations obtained by applying the Generalized Harmonic Balance decomposition to the multi-frequency forced response.

If this is right

  • Softening and hardening nonlinearities are predicted accurately from training at one excitation strength.
  • Jump phenomena appear correctly in the extrapolated frequency-response curves without additional training data.
  • Only a single excitation test is needed to characterize the global dynamics of the tested MEMS resonators and micromirrors.
  • Data acquisition effort for nonlinear microsystem design drops from repeated sweeps to one trajectory.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same decomposition-plus-sparse-regression pipeline could be applied to other weakly nonlinear systems such as Duffing or van der Pol oscillators under external forcing.
  • Embedding the method in an online experiment loop might allow continuous refinement of the model while the device operates.
  • Including additional harmonics in the balance step could extend the range to moderately nonlinear regimes while preserving the one-shot property.

Load-bearing premise

The response must remain weakly nonlinear so that the Generalized Harmonic Balance decomposition into a small set of slow-varying equations stays accurate when the training data come from only one excitation level.

What would settle it

Experimental frequency-response measurements at an excitation amplitude far from the single training level that show a jump location or peak amplitude differing markedly from the model's prediction would refute the extrapolation claim.

Figures

Figures reproduced from arXiv: 2604.15181 by Attilio Frangi, Lin Zhao, Luca Rosafalco, Teng Ma, Wei Cui.

Figure 1
Figure 1. Figure 1: One-Shot Learning framework of weakly nonlinear forced oscillator: The starting point of our framework is the dynamical response at a single excitation configuration, which is also the training data of MEv-SINDy. During the training stage, the multi-frequency dynamics are separated into distinct harmonic components. We use the generalized harmonic balance method to reformulate the sparse regression problem… view at source ↗
Figure 2
Figure 2. Figure 2: Identification process of the slowly varying evolutionary variables [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two examples of mean chamfer distance of frequency response functions (MCDRC): FRC1 (blue [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Benchmark 1D forced oscillator. Training data of numerical quadratic and cubic nonlinear [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Benchmark 1D forced oscillator. Influence of external forcing amplitude [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Benchmark 1D forced oscillator. Numerical validation of the Generalized Harmonic Balance (GHB) [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Benchmark 1D forced oscillator. Comparison between predicted and reference frequency response [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Schematic representation of the beam MEMS resonator. (a) Geometry and mesh of doubly clamped [PITH_FULL_IMAGE:figures/full_fig_p032_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Beam MEMS resonator training data in time, frequency and evolutionary domain. The model is [PITH_FULL_IMAGE:figures/full_fig_p032_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FRCs of beam MEMS resonator. Comparison between predicted and reference frequency response [PITH_FULL_IMAGE:figures/full_fig_p034_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Schematic representation of the MEMS micromirror: (a) Optical picture of the Micromirror; (b) [PITH_FULL_IMAGE:figures/full_fig_p035_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: MEMS micromirror training data in time, frequency and evolutionary domains: The model is [PITH_FULL_IMAGE:figures/full_fig_p036_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: FRCs of MEMS micromirror. Comparison between predicted and reference frequency response [PITH_FULL_IMAGE:figures/full_fig_p037_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Robustness evaluation of proposed approach across various training conditions. The predictive [PITH_FULL_IMAGE:figures/full_fig_p038_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Quantitative and qualitative error analysis of FRF predictions. (a) MCDRF distribution across all [PITH_FULL_IMAGE:figures/full_fig_p039_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of Frequency Response Curves (FRCs) identified from different training sets. (left) [PITH_FULL_IMAGE:figures/full_fig_p043_16.png] view at source ↗
read the original abstract

Extrapolative prediction of complex nonlinear dynamics remains a central challenge in engineering. This study proposes a one-shot learning method to identify global frequency-response curves from a single excitation time history by learning governing equations. We introduce MEv-SINDy (Multi-frequency Evolutionary Sparse Identification of Nonlinear Dynamics) to infer the governing equations of non-autonomous and multi-frequency systems. The methodology leverages the Generalized Harmonic Balance (GHB) method to decompose complex forced responses into a set of slow-varying evolution equations. We validated the capabilities of MEv-SINDy on two critical Micro-Electro-Mechanical Systems (MEMS). These applications include a nonlinear beam resonator and a MEMS micromirror. Our results show that the model trained on a single point accurately predicts softening/hardening effects and jump phenomena across a wide range of excitation levels. This approach significantly reduces the data acquisition burden for the characterization and design of nonlinear microsystems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces MEv-SINDy, an evolutionary sparse regression method that augments SINDy with multi-frequency handling and integrates it with the Generalized Harmonic Balance (GHB) decomposition to obtain slow-varying amplitude/phase evolution equations from a single forced time history. It claims that the resulting model, trained at one excitation level, extrapolates to predict the full nonlinear frequency-response curve—including softening/hardening behavior and jump phenomena—across wide ranges of excitation for a nonlinear beam resonator and a MEMS micromirror.

Significance. If the extrapolation holds with verifiable accuracy, the approach would materially reduce the experimental burden of mapping nonlinear frequency responses in microsystems, replacing dense frequency sweeps with a single trajectory measurement. The combination of GHB with evolutionary sparsity is a plausible route to data-efficient modeling of weakly nonlinear forced oscillators.

major comments (3)
  1. [Abstract] Abstract: the central claim that 'the model trained on a single point accurately predicts softening/hardening effects and jump phenomena across a wide range of excitation levels' is presented without any quantitative error metrics (RMSE, relative error on jump locations, or backbone-curve deviation), cross-validation protocol, or explicit statement of the training amplitude relative to the onset of nonlinearity. This leaves the extrapolation performance unsupported by visible evidence.
  2. [Validation sections] Validation sections (beam resonator and micromirror): the manuscript does not demonstrate that the GHB first-harmonic truncation remains accurate once response amplitude increases into the jump regime, nor does it report residual higher-harmonic content or amplitude-dependent coefficient drift. Without such checks, the assumption that the identified slow equations remain valid far outside the training trajectory is untested.
  3. [MEv-SINDy algorithm description] MEv-SINDy algorithm description: it is not shown whether the evolutionary term selection or any scaling parameters in the library are constrained so that the recovered coefficients are independent of the single training trajectory; if the sparsity threshold or library terms are tuned post-hoc to the training data, the reported predictions risk being circular rather than genuinely extrapolative.
minor comments (2)
  1. [Abstract] The abstract introduces the acronym MEv-SINDy with its expansion, but the subsequent text should consistently use the expanded form on first use in each major section for clarity.
  2. [Results figures] Figure captions for the MEMS results should explicitly state the excitation amplitude used for training versus the range of predictions shown, to allow immediate visual assessment of the extrapolation distance.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of clarity and validation. We address each major comment below and will incorporate revisions to strengthen the presentation of results and methods.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that 'the model trained on a single point accurately predicts softening/hardening effects and jump phenomena across a wide range of excitation levels' is presented without any quantitative error metrics (RMSE, relative error on jump locations, or backbone-curve deviation), cross-validation protocol, or explicit statement of the training amplitude relative to the onset of nonlinearity. This leaves the extrapolation performance unsupported by visible evidence.

    Authors: We agree that quantitative support strengthens the abstract. In the revised manuscript we will add explicit RMSE values for the predicted frequency-response curves, relative errors on jump locations, a statement of the training amplitude relative to nonlinearity onset, and a brief description of the cross-validation protocol employed. revision: yes

  2. Referee: [Validation sections] Validation sections (beam resonator and micromirror): the manuscript does not demonstrate that the GHB first-harmonic truncation remains accurate once response amplitude increases into the jump regime, nor does it report residual higher-harmonic content or amplitude-dependent coefficient drift. Without such checks, the assumption that the identified slow equations remain valid far outside the training trajectory is untested.

    Authors: We acknowledge the need for explicit validation of the GHB truncation. The revised validation sections will include quantitative checks on first-harmonic accuracy in the jump regime, reported residual higher-harmonic content extracted from the original time histories, and an assessment of amplitude-dependent drift in the identified coefficients to confirm validity of the slow equations outside the training trajectory. revision: yes

  3. Referee: [MEv-SINDy algorithm description] MEv-SINDy algorithm description: it is not shown whether the evolutionary term selection or any scaling parameters in the library are constrained so that the recovered coefficients are independent of the single training trajectory; if the sparsity threshold or library terms are tuned post-hoc to the training data, the reported predictions risk being circular rather than genuinely extrapolative.

    Authors: The MEv-SINDy implementation uses a fixed candidate library and evolutionary selection with sparsity thresholds set a priori according to the method's design criteria, independent of any specific training trajectory. In the revised algorithm description we will explicitly document these fixed parameters and include supplementary analysis showing that the recovered coefficients remain stable across different single-trajectory choices, thereby confirming the extrapolative nature of the predictions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained data-driven identification plus standard extrapolation

full rationale

The paper introduces MEv-SINDy as a sparse regression procedure that identifies coefficients in GHB-derived slow equations from a single measured time history. Subsequent predictions at new excitation amplitudes are obtained by numerically integrating those fixed-coefficient equations under different forcing terms. This is ordinary forward simulation of an identified model and does not reduce to the training data by construction, nor does it rely on self-citations, uniqueness theorems imported from prior author work, or renaming of known results. The weak-nonlinearity assumption and library completeness are explicit methodological premises, not hidden definitional loops. No load-bearing step equates a fitted quantity to its own prediction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that GHB can reliably reduce the non-autonomous dynamics to slow evolution equations and that evolutionary sparse regression can recover the correct sparse model from a single trajectory without additional constraints.

free parameters (1)
  • Sparsity threshold and library terms
    The evolutionary SINDy step selects which nonlinear terms to retain; these choices are fitted to the single time history and directly affect the learned equations.
axioms (1)
  • domain assumption The forced response admits a decomposition via Generalized Harmonic Balance into a set of slow-varying evolution equations
    Invoked to handle non-autonomous multi-frequency behavior from limited data.
invented entities (1)
  • MEv-SINDy algorithm no independent evidence
    purpose: To perform one-shot learning of governing equations for forced oscillators
    Newly introduced method whose correctness is not independently verified outside the two MEMS examples.

pith-pipeline@v0.9.0 · 5463 in / 1338 out tokens · 51859 ms · 2026-05-10T11:11:19.326097+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 4 canonical work pages

  1. [1]

    Model neurons: from hodgkin-huxley to hopfield, in: StatisticalMechanicsofNeuralNetworks: ProceedingsoftheXlthSitgesConferenceSitges, Barcelona, Spain, 3–7 June 1990, Springer. pp. 5–18. 42 Amsallem, D., Zahr, M.J., Farhat, C.,

    1990

  2. [2]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 28, 594–611

    One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence 28, 594–611. doi:10.1109/ TPAMI.2006.79. Franco, N., Manzoni, A., Zunino, P.,

    2006

  3. [3]

    Greco, C

    Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems. arXiv preprint arXiv:1808.01346 . Guillot, L., Cochelin, B., Vergez, C.,

    work page arXiv

  4. [4]

    ArXiv abs/2005.11699

    Physics-based polynomial neural networks for one-shot learning of dynamical systems from one or a few samples. ArXiv abs/2005.11699. URL:https://api.semanticscholar.org/CorpusID:218870280. Jiao, A., He, H., Ranade, R., et al.,

    work page arXiv 2005

  5. [5]

    1038/s41467-025-63076-z, doi:10.1038/s41467-025-63076-z

    URL:https://doi.org/10. 1038/s41467-025-63076-z, doi:10.1038/s41467-025-63076-z. Krack, M., Gross, J.,

    work page doi:10.1038/s41467-025-63076-z

  6. [6]

    Advanced Science n/a, e19707

    Encoding cumula- tion to learn perturbative nonlinear oscillatory dynamics. Advanced Science n/a, e19707. doi:https://doi.org/10.1002/advs.202519707. Maday, Y., Rønquist, E.M.,

    work page doi:10.1002/advs.202519707

  7. [7]

    Autoencoder and its various variants, in: 2018 IEEE international conference on systems, man, and cybernetics (SMC), IEEE. pp. 415–

    2018