Physics-Informed Nonlinear System Identification

• Physics-informed nonlinear system identification integrates physical governing equations with data-driven methods to produce interpretable, physically consistent nonlinear models.
• It employs residual-based, soft-constraint, and latent-space techniques to enforce structural properties such as energy balance, passivity, and stability during the identification process.
• Applications in power systems, industrial processes, and fluid dynamics demonstrate its ability to achieve high accuracy and computational efficiency under noisy, incomplete data conditions.

Physics-informed nonlinear system identification denotes a family of identification methods that learn nonlinear dynamical models from data while enforcing mechanistic structure such as governing differential equations, energy balance, passivity, stability, modal invariance, or other control-relevant properties. In the recent literature, it is consistently presented as an intermediate regime between purely white-box modeling and purely black-box regression: the model is trained from observed trajectories, but its hypothesis class, loss, or constraints are shaped by prior physics so that the identified dynamics remain interpretable, physically admissible, and more reliable for extrapolation, control, and inverse estimation (Sivaranjani et al., 6 Dec 2025, Linka et al., 2022, Donati et al., 14 Feb 2025).

1. Conceptual scope and problem classes

At the most general level, the identification target is a nonlinear input-state-output relation, often written as

$\dot{x}(t)=f(x(t),u(t)),\qquad y(t)=h(x(t),u(t)),$

or, in discrete hybrid form,

$$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$

where the known part $f,h$ carries physical structure and the unknown part $\Delta$ or the parameters $\theta$ are inferred from data (Sivaranjani et al., 6 Dec 2025, Donati et al., 14 Feb 2025). This formulation already shows that physics-informed identification is broader than inverse PINNs with fixed governing equations. It includes parameter estimation, residual-model learning, latent-coordinate discovery, reduced-order modeling, and unknown-input reconstruction.

The literature distinguishes several epistemic regimes. In one regime, the governing law is known up to parameters, and identification amounts to estimating those parameters jointly with the state trajectory, as in swing-equation identification from PMU data or damped-oscillator-constrained PINNs (Stiasny et al., 2020, Linka et al., 2022). In a second regime, the physics is only partially known, so the model is written as a mechanistic core plus a learned discrepancy term, as in residual-force correction, kernel correction, or black-box compensation of unmodeled dynamics (Garg et al., 2021, Donati et al., 9 Sep 2025, Donati et al., 14 Feb 2025). In a third regime, the original equations may be unavailable, but the learned representation is nevertheless constrained to carry physical meaning through latent modal coordinates, port-Hamiltonian structure, or interpretable surrogate coefficients (Rostamijavanani et al., 23 Jan 2025, Rettberg et al., 2024, Mansur et al., 16 Apr 2026).

This breadth explains why the topic connects simultaneously to nonlinear system identification, scientific machine learning, reduced-order modeling, system-theoretic structure preservation, and inverse problems. The surveyed control-oriented view makes this explicit by treating identification as an optimization problem in which control-relevant properties can be imposed through direct parameterization, soft constraints, and hard constraints (Sivaranjani et al., 6 Dec 2025).

2. Residual-based and soft-constrained formulations

The most visible strand is the residual-based formulation associated with PINNs and related architectures. Here a neural surrogate for the state trajectory is trained not only against measurements but also against a physics residual derived from the governing ODE or PDE. A representative loss is

$$L(\Theta;t)=(1-\varepsilon)L_{\rm data}+\varepsilon L_{\rm phys},$$

with derivatives computed by automatic differentiation and residuals enforced at measurement or collocation points (Linka et al., 2022). The same template appears in power-system frequency dynamics, where the network approximates generator angles, their derivatives are obtained by automatic differentiation, and the swing equation residual contributes a physics loss that is minimized jointly with the measurement mismatch (Stiasny et al., 2020).

Industrial process identification illustrates the same principle in a gray-box recurrent setting. For an induced draft evaporative cooling system, the basin temperature $T_b$ is identified from fan power, time of day, and weather variables, while the residual

$f=\frac{dT_b}{dt}-H(T_b,W;\beta)$

enforces first-order thermal dynamics. The resulting PhyNN and PhyLSTM architectures do not merely fit $T_b$; they also penalize violations of the cooling-tower thermodynamic model, with the recurrent variant using teacher forcing and, in its best version, output feedback (Lahariya et al., 2022).

Several recent variants modify the residual mechanism rather than abandoning it. Weak-form latent-space dynamics identification replaces pointwise derivative fitting by test-function-weighted integral relations, thereby avoiding direct differentiation of noisy trajectories. In WgLaSDI, an autoencoder and WENDy are trained simultaneously, and the weak-form latent and physical derivative penalties replace the strong-form SINDy terms used in gLaSDI (He et al., 2024). A related equation-discovery line uses B-splines for analytical derivatives, sequentially regularized derivatives for denoising, uncorrelated component analysis for dictionary pruning, and physics-informed spline fitting so that the spline representation is gradually updated while satisfying the candidate governing equation (Pal et al., 2024).

Soft-constraint methods are powerful because they are architecture-agnostic and naturally accommodate continuous-time models, noisy data, and collocation-based regularization. At the same time, the literature emphasizes that they are sensitive to scaling and to the weighting between data and physics; a fixed physics weight can bias the solution too far toward either interpolation or pure equation fitting (Linka et al., 2022).

3. Structured gray-box models and interpretable latent representations

A second major strand embeds physics by construction rather than primarily through residual penalties. In gray-box correction frameworks, the mechanistic model is retained and only its model-form error is learned. One formulation rewrites the true nonlinear dynamics as an approximate governing equation plus a residual force $\mathbf R$, estimates $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$0 jointly with the state via a dual Bayesian filter, learns a mapping $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$1 using Gaussian process regression, and reinserts that learned term into the governing equation for forward prediction under seen and unseen forcing (Garg et al., 2021). A kernel-based analogue writes

$$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$2

assumes $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$3 lies in an RKHS, and estimates the physical parameters and kernel correction simultaneously rather than in two stages (Donati et al., 9 Sep 2025).

Other methods move the identification problem into a latent space chosen to carry physical semantics. Lift & Learn uses known governing equations to define a lifting map $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$4 such that the lifted dynamics have quadratic structure,

$$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$5

then performs POD in lifted coordinates and identifies reduced linear and quadratic operators by least squares (Qian et al., 2019). Physics-constrained normalizing flows use exact invertibility and latent-space independence to identify nonlinear normal modes directly from response data only, with the learned inverse map $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$6 interpreted as a nonlinear modal transformation (Rostamijavanani et al., 23 Jan 2025).

Energy-based structure preservation appears in latent port-Hamiltonian identification. There the encoder maps physical states to a low-dimensional latent variable $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$7, the latent dynamics are constrained to

$$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$8

and the matrices are parameterized through triangular factorizations so that $$x_{k+1}=f(x_k,u_k;\theta)+\Delta(x_k,u_k),\qquad z_k=h(x_k;\theta),$$9, $f,h$0, and $f,h$1 hold by construction (Rettberg et al., 2024). SOLIS pursues a related interpretability objective through a state-conditioned second-order surrogate

$f,h$2

which recasts identification as learning a Quasi-LPV representation and recovers local natural frequency, damping ratio, and DC gain as state-dependent quantities (Mansur et al., 16 Apr 2026).

Collectively, these methods show that “physics-informed” need not mean only “residual-constrained neural network.” It can also mean that the coordinate system, residual correction, or latent state-space model is chosen so that the identified system preserves modal, energetic, or control-theoretic structure (Sivaranjani et al., 6 Dec 2025).

4. Partial observations, uncertainty, and inverse estimation

A recurring difficulty is that real identification problems rarely provide dense, noise-free, full-state observations. One response is explicit uncertainty quantification. Bayesian PINNs combine physics-informed losses with posterior inference over network and physics parameters, yielding credible intervals and allowing the same model to solve forward and inverse problems. In the damped-oscillator study, classical Bayesian inference, BNNs, and BPINNs are compared directly, with the conclusion that BPINNs provide the most expressive joint treatment of physics and uncertainty but are also the most computationally demanding and sensitive to scaling (Linka et al., 2022).

A second response is to redesign the identification objective around non-uniform data. One recent framework treats missing measurements, multiple runs, and aggregated observations within the same nonlinear state-space formulation. Missing samples are handled by restricting the loss to observed indices, multiple trajectories are stacked with distinct initial states, and aggregated measurements are converted into an equivalent extended-state problem. The paper also proves explicit bounds showing that the missing-data error depends on $f,h$3, whereas the aggregation error grows roughly like $f,h$4 (Donati et al., 14 Feb 2025).

A third response integrates state estimation into the physics-informed loop. In the kernel-based hybrid framework, UKF and Unscented Rauch–Tung–Striebel smoothing reconstruct latent states before the parametric-plus-kernel model is estimated (Donati et al., 9 Sep 2025). APSMC couples Kalman-filter-based state estimation with proximal-gradient updates of state-space matrices, interprets the time-varying matrix $f,h$5 as a real-time Jacobian tracker for nonlinear structural dynamics, and claims convergence toward the optimal physically consistent state-space estimate as data accumulate (Chen et al., 10 May 2025).

The most stringent inverse-estimation formulation appears when unknown inputs must be reconstructed from partial measurements. In autonomic cardiac regulation, a neural inverse estimator is trained under data fidelity and dynamical consistency constraints, but the design is further restricted by left-invertibility conditions obtained by differential-algebraic elimination. The paper derives admissibility conditions, introduces penalties for sigmoid non-saturation and non-degeneracy, and analyzes a conservative Lipschitz bound for the inverse map. This makes well-posedness in the sense of Hadamard a design criterion rather than an afterthought (Sadoun et al., 2 Jun 2026).

5. Representative applications and reported performance

The empirical record spans industrial processes, power grids, high-dimensional PDEs, structural systems, and chaotic ODEs. The reported results are heterogeneous because the benchmarks, data regimes, and objectives differ, but they show that physics-informed identification is used both for accurate short-horizon prediction and for structure-preserving extrapolation.

Domain	Method	Reported result
Evaporative cooling system	PhyLSTMWF	less than 2% system response estimation error; with 7 months of training data, training loss converges in about 100 iterations (Lahariya et al., 2022)
Power-system frequency dynamics	PINN	for System A and some parameters in System B, PINN achieves <1% relative error; tested with Gaussian noise up to 5% and uniform noise up to ±5% (Stiasny et al., 2020)
High-dimensional nonlinear PDE ROMs	WgLaSDI	with data that contains 5–10% Gaussian white noise, WgLaSDI achieves 1–7% relative errors and 121 to 1,779x speed-up (He et al., 2024)
Parametric latent ROMs	gLaSDI	compared with the high-fidelity models, gLaSDI achieves 17 to 2,658x speed-up with 1 to 5% relative errors (He et al., 2022)
Flow over a cylinder	Physics-constrained NF	streamwise velocity MSE $f,h$6 versus POD $f,h$7; transverse velocity MSE $f,h$8 versus POD $f,h$9 (Rostamijavanani et al., 23 Jan 2025)
Online structural estimation	APSMC	predicts 19 consecutive 10-second time series using only a single initial 10-second segment for model updating; minimum NMSE 0.398% (Chen et al., 10 May 2025)
Sparse noisy chaotic reconstruction	PIDM-DP	reconstruction RMSE improvements of up to $\Delta$0 over an unconstrained diffusion baseline; on Rabinovich-Fabrikant OOD, RMSE $\Delta$1 versus $\Delta$2 (Dabral, 26 May 2026)

Beyond these results, several studies emphasize secondary benefits that are central to system identification rather than mere forecasting. In the evaporative-cooling study, the identified dynamics support direct estimation of flexibility metrics such as state of charge and rate of charge (Lahariya et al., 2022). In the power-systems study, the continuous-time neural estimator produces differentiable state trajectories and jointly estimates inertia and damping from PMU data (Stiasny et al., 2020). In SOLIS, the point is not only reconstruction accuracy but coherent physical rollouts and recovery of a parameter manifold described by local frequency, damping, and gain (Mansur et al., 16 Apr 2026). In latent port-Hamiltonian identification, the gain is a rapidly computable reduced model with passivity and stability in the latent space (Rettberg et al., 2024).

6. Limitations, misconceptions, and open directions

A common misconception is that adding physics always improves identification monotonically. The evidence is more qualified. In the evaporative-cooling benchmark, increasing training data improves PhyLSTMWF, but for simpler feedforward models one month of data can sometimes outperform seven months because the model must generalize across stronger seasonal variation; the paper explicitly notes that more data does not always help (Lahariya et al., 2022). Likewise, Bayesian PINNs show that a fixed data-physics weight can be problematic: very small $\Delta$3 reduces to a plain NN, whereas $\Delta$4 can converge to a trivial zero-amplitude solution if the model is underconstrained (Linka et al., 2022).

Another misconception is that physics-informed identification removes optimization pathologies. Several papers state the opposite. NF training is described as somewhat unstable, requiring a very small learning rate and becoming computationally expensive for high-dimensional systems (Rostamijavanani et al., 23 Jan 2025). The control-oriented survey emphasizes that hard constraints may make optimization nonconvex or difficult, strong structure can introduce model bias, and constrained architectures may be less expressive even when they preserve stability or dissipativity (Sivaranjani et al., 6 Dec 2025). Sparse equation discovery remains highly sensitive to correlated dictionaries, which motivates additional mechanisms such as UCA and physics-informed spline refitting (Pal et al., 2024).

A third misconception is that forward consistency suffices for inverse estimation. The autonomic-regulation work argues that it does not: uniqueness, admissibility, and stability of the inverse map must be enforced explicitly through left-invertibility and conditioning constraints (Sadoun et al., 2 Jun 2026). The same theme appears indirectly in non-uniform-observation identification, where aggregation creates an inherent information bottleneck, and in chaotic reconstruction, where a trajectory can pass through sparse observations while still drifting off the true attractor unless the ODE is enforced during inference (Donati et al., 14 Feb 2025, Dabral, 26 May 2026).

Open problems are stated clearly in the survey literature. These include identification in networked, switched, and time-varying systems; experiment design for informative yet safe excitation; bridging flexible nonlinear function approximation with behavioral and direct data-driven guarantees; sample-complexity theory for structured model classes; and standardized control-relevant benchmarks (Sivaranjani et al., 6 Dec 2025). This suggests that the field is moving from the initial demonstration that physics priors can improve nonlinear identification toward a more discriminating agenda: identifying which priors are appropriate, how they should be enforced, and what guarantees they deliver for prediction, extrapolation, control, and inverse inference.