Physics-Informed Neural Networks Show Major Validation Gaps in Structural Dynamics

A comprehensive study-level audit of forward structural-dynamic physics-informed neural networks reveals that satisfying governing differential equations during machine learning does not guarantee numerical reliability or physical credibility. Published in Frontiers in Mechanical Engineering by researchers Borun Das of the University of Louisiana at Lafayette and Abir Nag Bulbul of United International University, the investigation examines twenty-seven peer-reviewed studies published through September 2026. While physics-informed neural networks, commonly termed PINNs, integrate mathematical models directly into loss functions, structural dynamics imposes stringent verification and validation demands. Dynamic systems involve continuous oscillations, higher-order spatial derivatives, long-term temporal error accumulation, complex frequency spectra, and changing operating conditions. Existing scientific literature has frequently treated equation residual minimization or isolated surface agreement as sufficient evidence of model accuracy. However, this full-text audit demonstrates that reported validation across the published literature remains highly uneven, with significant structural deficiencies appearing in high-frequency capture, derivative quantities, and physical experimentation.

To evaluate the credibility of forward predictive models systematically, the authors operationalized four core evidence dimensions grounded in established computational mechanics standards from the American Society of Mechanical Engineers. These dimensions comprise equation and numerical consistency, structural response verification, dynamic fidelity, and predictive scope evaluated alongside benchmark provenance. Rather than presenting subjective narrative summaries of individual study successes, the review resolved each dimension into concrete categories governed by explicit decision rules applied consistently across full texts. Each documented dimension connects directly to an operational reporting chain linking the specific engineering quantity of interest, the formal error metric, the selected reference and its underlying uncertainty, the intended-use tolerance, the observed numerical error, and the ultimate validation conclusion. The audit applied this structured framework across diverse structural applications, encompassing simple beams, cables, continuous bridges, multi-storey frames, and nonlinear dynamic systems subjected to seismic ground motions or moving operational loads.

Across the twenty-seven core studies evaluated in the quantitative synthesis, numerical robustness was widely discussed but rarely demonstrated with conclusive rigor. The authors found that only five investigations performed formal refinement sequences or repeated training runs with documented statistical variability to evaluate solution sensitivity. Twenty-one studies presented much weaker numerical evidence, such as basic two-level comparisons or exploratory parameter sensitivity tests linked to solution error, while one paper reported only a single run tracking total loss. Structural response verification beyond primary displacement fields proved equally scarce across the corpus. Only nine studies verified a secondary kinematic quantity such as velocity, acceleration, or rotational deflection. More critically, only six studies validated derived engineering quantities, such as bending moments, internal shear forces, strains, stresses, or interstorey drift, which are vital for real-world structural integrity assessments.

The audit revealed even sharper reporting deficits within dynamic fidelity, spectral behavior, modal response, and independent experimental validation. Long-horizon dynamic fidelity was documented with temporal resolution over extended intervals or complete physical events in only eight studies, while ten reported solely aggregate time-averaged errors and nine presented neither. Spectral evidence appeared in only one study, with partial spectral evidence in three others, and modal-output validation was verified in just three papers. Furthermore, independent physical validation was achieved in only two of the twenty-seven studies, alongside partial physical validation in two others, whereas the vast majority relied strictly on analytical formulations or numerical simulations. When evaluating cross-dimensional performance, the audit observed that evidence does not accumulate within individual research projects: no study satisfied more than five of the eight defined evidence categories, six satisfied none, and the median study satisfied only one.

These widespread empirical gaps highlight fundamental theoretical and numerical mechanisms governing physics-informed machine learning that cannot be resolved by residual minimization alone. Theoretical analyses cited in the audit demonstrate that minimizing differential equation residuals does not directly yield general solution error bounds without strict supplementary mathematical assumptions. In second-order dynamic partial differential equations, formal mathematical bounds derived for solutions, temporal derivatives, and spatial gradients require residual terms that canonical neural network loss functions routinely fail to optimize. Training dynamics also suffer from severe gradient imbalances among competing loss terms, distortion from optimization failure modes in complex regimes, and frequency-dependent learning rates explained by neural tangent kernel analyses. Additionally, enforcing boundary conditions through soft penalty terms can yield deceptive optimizations where diminishing loss values accompany surging solution errors, unlike rigid mathematical enforcement where constraints are satisfied strictly by construction.

In structural engineering mechanics, verification requirements must extend far beyond primary kinematic fields because governing safety decisions rely on derivative-dependent physical quantities. The audit underscores that high accuracy in displacement predictions provides no mathematical guarantee of accuracy in internal stresses or moments. Several evaluated studies demonstrated that approximation errors grow substantially with higher differentiation orders when relying on standard automatic differentiation in separable network architectures. Introducing auxiliary stress or strain variables into the formulation directly mitigates derivative degradation, proving that structural safety assessments necessitate direct verification of derived engineering quantities rather than unverified extrapolations from matching displacement plots.

The review also examined temporal error growth across extended dynamic response horizons, demonstrating that short-duration matching frequently masks catastrophic long-term model divergence. Parallel investigations of vibrating suspension cables revealed that predicted trajectories deteriorated rapidly after approximately four seconds of vibration. For strongly nonlinear structural dynamics, dividing simulations into shorter temporal segments improved solution precision, though at significantly increased computational expense. The review emphasizes that dynamic reliability must be demonstrated across physically meaningful durations spanning dozens of fundamental natural periods, rather than being inferred from agreement within brief, artificially truncated initial time windows.

Dynamic fidelity presents acute challenges regarding spectral bias and modal content preservation under multifrequency or moving excitations. Standard coordinate-based neural networks exhibit well-documented spectral bias, learning low-frequency response components rapidly while struggling to resolve higher-frequency harmonics. In moving-load beam benchmarks, time-domain formulations failed to capture higher vibration frequencies; yet proposed frequency-domain remedies were validated merely against time-domain displacement rather than rigorous spectral metrics. Conversely, embedding natural frequencies within Fourier feature mappings stabilized seismic response predictions across multiple earthquake records, whereas poorly chosen Fourier scales caused relative prediction errors to spike past fifty percent. Experimental cantilever studies demonstrated that frequency response functions, fast Fourier transform processing, and laser-vibrometer measurements can validate natural frequencies and damping ratios, proving that spectral and modal criteria represent foundational engineering requirements rather than optional machine-learning diagnostics.

Assessing model generalizability and benchmark provenance introduced further comparability dilemmas across the twenty-seven reviewed studies. The authors argue that predictive scope must specify precise physical variations—such as changes in load amplitudes, material distributions, structural geometries, or uninstrumented sensor locations—rather than relying on vague claims of model generalization. In plate surrogate benchmarks, networks exhibited substantially higher error when confronted with unseen boundary conditions than with altered geometries or load magnitudes. Furthermore, benchmark provenance dictates evidentiary strength: analytical and numerical simulations verify mathematical solvers, whereas laboratory experiments and field-monitoring data capture unavoidable modeling discrepancies and measurement noise. In one bridge monitoring study, a frozen model demonstrated load extrapolation up to ninety-five kilonewtons after training on sixty-five kilonewtons, but its field validation was confined entirely to strain sensors, leaving deflection and moment predictions reliant on computational references.

To overcome these systemic validation deficits, the authors propose a transparent minimum reporting framework for future forward structural-dynamic investigations. Rather than adopting binary labels of validated or unvalidated models, future studies should report complete evidence chains defining specific quantities of interest, explicit error metrics, reference provenance, tolerance thresholds, observed errors, and domain limitations. Researchers must distinguish between physics-only solvers, data-assisted methods, and reduced-order hybrid models, isolating the error contributions of embedded numerical integrators or modal superpositions. Furthermore, computational cost accounting must separate expensive offline training and data generation from rapid online neural inference to prevent misleading efficiency comparisons against traditional finite-element methods. By shifting from isolated demonstrations toward standardized, multi-dimensional verification benchmarks, computational mechanics can establish when physics-informed neural networks achieve genuine engineering credibility for safety-critical structural applications.

Subject of Research: Verification and validation evidence distribution in forward structural-dynamic physics-informed neural networks

Article Title: Beyond governing-equation residuals: verification and validation of physics-informed neural networks for forward structural-dynamic response prediction

Article References: Das, B., & Bulbul, A. N. (2026). Beyond governing-equation residuals: verification and validation of physics-informed neural networks for forward structural-dynamic response prediction. Frontiers in Mechanical Engineering, 12, 1979675. https://doi.org/10.3389/fmech.2026.1979675

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1979675

Keywords: physics-informed neural networks, structural dynamics, verification and validation, computational mechanics, dynamic fidelity, spectral bias, modal analysis, finite element method, equation consistency, predictive scope, benchmark provenance, derivative-dependent quantities, beam vibration, seismic response, moving loads, error accumulation, surrogate modeling