Verifying Physics-Informed Neural Network Fidelity using Classical Fisher Information from Differentiable Dynamical System PubID: pubid: 0000–0000/00$00.00 © 2021 IEEE

Josafat Ribeiro Leal Filho, Antônio Augusto Fröhlich · 2026

Read Verifying Physics-Informed Neural Network Fidelity using Classical Fisher Information from Differentiable Dynamical System PubID: pubid: 0000–0000/00$00.00 © 2021 IEEE by Josafat Ribeiro Leal Filho, Antônio Augusto Fröhlich (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-11638.txt.

Introduction

Josafat Ribeiro Leal Filho and Antonio Augusto Frohlich verify PINN fidelity. Their 2026 preprint uses classical Fisher information for the check. Physics-informed neural networks solve differential equations by learning. They embed physical laws into the training process itself.

Judging how fully a PINN captures dynamics is still hard. Trajectory prediction alone says little about full behavior. The team draws Fisher information from a differentiable system. That quantity can measure sensitivity in the model's fit. A differentiable dynamical system supplies the needed gradients. The method aims at rigorous fidelity scores for PINNs. Such scores matter where PINNs guide real engineering.

Catalog record

CREATORJosafat Ribeiro Leal Filho, Antônio Augusto Fröhlich
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-11638

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Verifying Physics-Informed Neural Network Fidelity using Classical Fisher Information from Differentiable Dynamical System PubID: pubid: 0000–0000/00$00.00 © 2021 IEEE Josafat Ribeiro Leal Filho, Antônio Augusto Fröhlich Abstract Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool for solving differential equations and modeling physical systems by embedding physical laws into the learning process. However, rigorously quantifying how well a PINN captures the complete dynamical behavior of the system, beyond simple trajectory prediction, remains a challenge. This paper proposes a novel experimental framework to address this by employing Fisher information for differentiable dynamical systems, denoted g F C g_{F}^{C} . This Fisher information, distinct from its statistical counterpart, measures inherent uncertainties in deterministic systems, such as sensitivity to initial conditions, and is related to the phase space curvature and the net stretching action of the state space evolution. We hypothesize that if a PINN accurately learns the underlying dynamics of a physical system, then the Fisher information landscape derived from the PINN’s learned equations of motion will closely match that of the original analytical model. This match would signify that the PINN has achieved comprehensive fidelity capturing not only the state evolution but also crucial geometric and stability properties. We outline an experimental methodology using the dynamical model of a car to compute and compare g F C g_{F}^{C} for both the analytical model and a trained PINN. The comparison, based on the Jacobians of the respective system dynamics, provides a quantitative measure of the PINN’s fidelity in representing the system’s intricate dynamical characteristics. I Introduction Physics-Informed Neural Networks (PINNs) constitute a class of machine learning models that explicitly incorporate governing physical laws—most commonly expressed in the form of partial or ordinary differential equations—directly into the training process through the neural network loss function [ 20 ] . By embedding prior physical knowledge into the learning architecture, PINNs reduce reliance on large labeled datasets and provide a principled framework for learning dynamical systems constrained by known physics. Consequently, PINNs have been successfully applied across a broad range of scientific and engineering domains, including fluid dynamics, solid mechanics, and inverse parameter identification, often achieving accurate solutions in regimes characterized by sparse, noisy, or incomplete data [ 10 ] . Despite these advances, it is now well recognized that PINNs exhibit notable limitations. Previous studies have documented difficulties in training PINNs for stiff systems, chaotic dynamics, multiscale problems, and long-time integration, where optimization may stagnate or converge to physically inconsistent solutions [ 26 , 12 ] . PINNs are also known to suffer from spectral bias, resulting in poor representation of high-frequency solution components [ 25 ] , as well as from imbalance among competing loss terms, which can lead to solutions that satisfy governing equations numerically while failing to capture correct dynamical behavior [ 17 ] . These observations expose a fundamental limitation of prevailing evaluation practices for PINNs, which largely rely on pointwise error norms or trajectory-level matching with reference solutions. While such metrics are necessary, they are insufficient to fully characterize the fidelity of a learned dynamical model. In particular, they often fail to detect discrepancies in local sensitivity, stability, and geometric structure that govern predictability and robustness. In this work, we refer to this broader notion of model fidelity as fidelity : the extent to which a learned model captures not only the evolution of states, but also the local response of trajectories to perturbations and the underlying geometry of the dynamical flow. Sensitivity analysis has long been formalized through the concept of Fisher information. In classical statistics, Fisher information quantifies how sensitively a probability distribution depends on its parameters, thereby measuring the amount of information that observable data carry about those parameters. This notion admits a geometric interpretation as a metric on the space of distributions. The concept was later generalized to quantum systems through the quantum Fisher information, which quantifies the distinguishability of nearby quantum states and plays a central role in quantum estimation theory [ 22 ] . In both settings, Fisher information provides a principled measure of sensitivity encoded in an underlying geometric structure. More recently, this conceptual information has been extended to deterministic dynamical systems. Sahbani, Das, and Green introduced a classical Fisher information metric, denoted g_{F}^{C} , defined directly on the phase space of sufficiently smooth deterministic states without recourse to probability distributions [ 22 ] . Instead, g_{F}^{C} is constructed from Lyapunov vectors in the tangent space, capturing the local stretching and contraction induced by the system’s flow. This formulation yields a geometric measure of sensitivity that is intrinsically linked to local trajectory curvature, stability properties, and the differential structure of the dynamics [ 22 ] . As such, g_{F}^{C} provides a natural tool for probing whether a learned model has accurately captured the local structure of a dynamical system, rather than merely reproducing its trajectories. In this work, we propose an experimental framework that leverages this classical Fisher information metric as a validation and diagnostic tool for Physics-Informed Neural Networks. Our central hypothesis is that a PINN that faithfully learns the governing dynamics of a physical system should reproduce not only the observed state evolution, but also the Fisher information geometry associated with the true dynamics. Specifically, we assess fidelity by comparing the Fisher information landscape computed from the PINN’s learned dynamics with that of an analytical or high-fidelity reference model. Agreement at this level would indicate that the PINN has captured essential geometric, sensitivity, and stability properties of the system. The contributions of this work are twofold. First, we introduce classical Fisher information as a principled metric for evaluating the dynamical fidelity of PINNs. Second, we design an experimental framework for computing and comparing Fisher information landscapes between learned and reference systems. I-A Related Work The validation of neural networks for dynamical systems has traditionally followed evaluation paradigms inherited from regression and forecasting. While these approaches have gradually incorporated physical priors and geometric constraints, a fundamental gap remains between predictive accuracy and dynamical fidelity. This section synthesizes prior work by focusing not on a catalog of techniques, but on the structural limitations that motivate the need for deeper validation criteria. I-A 1 Limits of Trajectory-Based Validation Most learned dynamical models are validated using trajectory-level error metrics such as Mean Squared Error (MSE), Average Displacement Error (ADE), or Final Displacement Error (FDE) [ 3 ] . These metrics assess how closely predicted trajectories match observed ones, and are indispensable for benchmarking short-term predictive performance. However, they provide only a weak guarantee that the learned model captures the underlying dynamics. Trajectory-based metrics are agnostic to the causal structure of the vector field. As emphasized in reliability and validation studies [ 21 ] , low prediction error does not imply that a model has learned essential physical properties such as stability, invariants, or sensitivity to perturbations. Neural networks can minimize MSE by behaving as interpolators over the training data, effectively memorizing trajectories while misrepresenting the local flow of the system. This shortcoming is particularly acute in dynamical systems, where generalization is not merely a matter of extrapolating in state space, but of preserving the qualitative structure of the phase portrait. Moreover, for physical systems, it is often unclear how to meaningfully partition data

Read the complete text