Physics-Informed Learning

• Physics-informed learning is a framework that integrates known physical laws and constraints into machine learning models to improve data efficiency and ensure realistic outcomes.
• It utilizes methods such as penalty terms, hard constraints, and operator-based regularization to enforce differential equations and maintain physical consistency during training.
• This approach is applied across diverse domains like fluid dynamics, biomedical imaging, and control systems, offering robust, computationally efficient predictions in data-scarce environments.

Physics-informed learning is a rigorous methodological framework in which prior knowledge of physical laws, symmetries, or constraints is embedded directly into statistical or machine learning models. By systematically incorporating these domain-specific priors—typically expressed in the form of differential, integral, or algebraic operators—physics-informed approaches constrain the learned model class, enhance data efficiency, improve generalization, and guarantee physical plausibility even in the presence of limited or noisy data (Meng et al., 2022, Doumèche, 11 Jul 2025, Ahmadi et al., 6 Oct 2025). This paradigm underlies a vast array of modern model architectures, theoretical analyses, and applications at the intersection of computational science, engineering, and machine learning.

1. Conceptual Foundations and Motivation

At its core, physics-informed learning unifies data-driven modeling and first-principles physical knowledge by augmenting standard empirical risk minimization objectives with penalty terms, constraints, or architectural modifications derived from known physical relationships. These include partial differential equations (PDEs), ordinary differential equations (ODEs), conservation laws, stability conditions, symmetry requirements, and more general operator-theoretic structures. Typical motivations for adopting this framework are:

• Data scarcity and cost: High-fidelity simulation, laboratory, or field data are often scarce or prohibitively expensive to collect. Physics-informed models leverage knowledge of governing laws to interpolate accurately between limited supervision (Meng et al., 2022, Doumèche, 11 Jul 2025).
• Physical consistency: Embedding physical constraints excludes unphysical model outputs, ensuring physically meaningful predictions even under distribution shift, adversarial corruptions, or noise (Ahmadi et al., 6 Oct 2025, Ghosh et al., 2021).
• Interpretability and parameter discovery: Model decompositions often permit the identification of physically interpretable parameters or dynamics, aiding inverse problems and system identification.
• Computational efficiency: Once trained, physics-informed surrogates can deliver real-time or amortized inference at orders-of-magnitude lower computational cost than direct numerical solvers (Ghosh et al., 2021, Doumèche, 11 Jul 2025).

2. Mathematical Formulations and Model Classes

Let $u_\theta(x)$ denote a machine learning model parameterized by $\theta$, aiming to approximate the solution of a physical system described by an operator equation $\mathcal{L}[u](x) = 0$ on domain $x \in \Omega$. The paradigm encompasses several canonical forms:

• Residual-penalty loss (soft constraints):

$$J(\theta) = \frac{1}{N}\sum_{i=1}^N |u_\theta(x_i) - y_i|^2 + \lambda \frac{1}{M}\sum_{j=1}^M |\mathcal{L}[u_\theta](x_j^{(r)})|^2,$$

where $(x_i, y_i)$ are labeled data and $x_j^{(r)}$ are collocation points for physics enforcement (Doumèche, 11 Jul 2025, Meng et al., 2022).

• Constrained/hard-satisfaction architectures (hard constraints):

Model ansätze are constructed to automatically satisfy boundary or initial conditions, e.g.,

$\hat u(x, t) = g(x) + t N(x, t; \theta)$

to enforce Dirichlet data (Meng et al., 2022).

• Operator or functional approaches:

Learning the solution operator $\mathcal{G}_\theta: a(\cdot) \mapsto u(\cdot)$ (e.g., with DeepONet, FNO) directly in function space with embedded physical loss (Ahmadi et al., 6 Oct 2025, Marcandelli et al., 2 Feb 2026).

• Kernel and variational formulations:

Physics-informed regularization is interpreted as an RKHS norm or as a physics-based kernel (PIKL), yielding closed-form solutions in terms of operator Green’s functions or Fourier series (Doumèche et al., 2024, Doumèche, 11 Jul 2025).

3. Model Architectures and Integration Methods

Representative Architectures

• Physics-Informed Neural Networks (PINNs): Feed-forward or shallow conv/deconv neural networks parameterize $u_\theta(x)$, with physics and data losses jointly enforced (Meng et al., 2022, Shi et al., 2022).
• Physics-Informed Convolutional Networks (PICN): Shallow, convolutional architectures with fixed, pre-trained finite-difference kernels compute discrete derivatives efficiently and are effective across irregular domains (Shi et al., 2022).
• Neural Operators and Operator Learning: Architectures (e.g., DeepONet, FNO, PINO, PhIS-FNO) parameterize mappings between function spaces and enable mesh-independent inference with embedded physical losses or curriculum stages (Ahmadi et al., 6 Oct 2025, Marcandelli et al., 2 Feb 2026).
• Echo State Networks and RNNs: Reservoir computing models are trained with additional physics-based penalties and LMI constraints for control-oriented applications and guaranteed stability (Doan et al., 2020, Ravasio et al., 26 Mar 2026).
• Hybrid and Structured Models: Architectures incorporating both black-box and physically-structured sub-networks; e.g., models with enforced symplecticity, sparsity, or multi-scale modularity (Nghiem et al., 2023, Liu et al., 2021).

Methods of Physics Integration

• Penalty Terms: Softly penalize violations of PDEs/operator residuals during training (as in PINNs/PINOs).
• Architectural Ansatz: Construct solution forms or network layers that guarantee satisfaction of physical hard constraints.
• Operator-based Regularization: Employ physics-based kernels, Green's function expansions, or operator spectral decompositions in the model parameterization or loss (Doumèche et al., 2024, Doumèche, 11 Jul 2025).
• Diffusion Models with Physics-informed Guidance: Diffusion-based generative models (PILD, Pi-fusion) inject physics via virtual residual observations or direct score guidance during sampling (Zeng et al., 29 Jan 2026, Qiu et al., 2024).

4. Theoretical Foundations and Statistical Properties

Analytical studies of physics-informed learning have produced precise generalization bounds, convergence theorems, and quantitative insight into the impact of physical priors:

• Statistical Learning Rate Acceleration: For empirical risk minimization with physics-informed regularization, alignment of the true solution with the kernel of the regularizing operator (e.g., when $\theta$0 for differential operator $\theta$1) yields a transition from slow Sobolev minimax rates $\theta$2 to fast parametric rates $\theta$3, even under temporally dependent (mixing) data (Scampicchio et al., 29 Sep 2025, Doumèche et al., 2024).
• Singular Learning Theory: PINNs are "singular" from a statistical learning viewpoint—the parameter identification map is highly non-injective with flat regions of the loss landscape. Tools such as the Local Learning Coefficient (LLC) quantify the effective complexity and uncertainty, and explain generalization and extrapolation limits in physics-informed models (Barajas-Solano, 11 Feb 2026).
• Kernel Methods and RKHS Structure: For linear operators, physics-informed learning corresponds to learning in RKHSs determined by the operator's Green's function, with data and regularization terms dictating the covariance structure and sample complexity (Doumèche, 11 Jul 2025, Doumèche et al., 2024).
• Robustness and Data Efficiency: Imposing explicit physics constraints (even through weak or indirect supervision) robustly regularizes models and reduces required labeled data or supervision (Ghosh et al., 2021, Dang et al., 17 Apr 2025).

5. Integration with Generative, Probabilistic, and Operator Learning

Recent advances merge physics-informed loss mechanisms with probabilistic generative modeling and operator learning:

• Diffusion-based Physics-Informed Learning: PILD and Pi-fusion generalize diffusion models to enforce physical laws either by virtual residuals sampled from heavy-tailed distributions (e.g., Laplace) or by direct score guidance using PDE residuals, achieving high-fidelity generative surrogates for ODE/PDE constrained systems (Zeng et al., 29 Jan 2026, Qiu et al., 2024).
• Curriculum and Stagewise Optimization: Multi-stage optimization strategies, such as the curriculum-based PhIS-FNO, sequentially enforce boundary conditions and PDE residuals while re-initializing optimizers, leading to improved convergence and stability in unsupervised operator learning (Marcandelli et al., 2 Feb 2026).
• Constrained Learning with Static Data: Constrained physics-informed learning enables recovery of ODE-type dynamics from static or incomplete observational data via graph-structured message passing over known physical graph topologies (Dang et al., 17 Apr 2025).
• Physics-Augmented and Generative Modeling: The duality between physics-informed (discriminative, penalty-based) and physics-augmented (generative, hard-constraint, model-structural) approaches highlights an emerging taxonomy for physics-constrained ML frameworks (Liu et al., 2021).

6. Practical Applications and Impact

Physics-informed learning architectures and algorithms are deployed across a broad spectrum of scientific and engineering domains:

• Spatio-temporal dynamical prediction: Fluid dynamics, weather modeling, quantum and classical many-body systems, plasma turbulence, chemical process control (Qiu et al., 2024, Ghosh et al., 2021, Zeng et al., 29 Jan 2026).
• Biomedical science and engineering: Biomedical imaging (PDE-constrained inversion), mechanobiology, pharmacokinetics/dynamics, and simulation-based digital twins for personalized medicine (Ahmadi et al., 6 Oct 2025).
• Dynamical systems and control: System identification, model-based RL, robust and safe optimal control, Lyapunov/barrier function learning, data-driven predictive safety filters (Nghiem et al., 2023, Liu et al., 2021).
• Materials science and optics: Surrogate models for eigenvalue problems in photonic composites and optical waveguides, parameterized closure models for thermodynamic state theory (Ghosh et al., 2021, Chen et al., 2023).
• Engineering design and time-series forecasting: Real-time vehicle tracking, tire-model estimation, multi-fidelity and multi-scale surrogate modeling for power systems and mobility data (Mo et al., 2020, Doumèche, 11 Jul 2025).

Representative empirical findings demonstrate superior accuracy, data efficiency, noise robustness, and computational speed for physics-informed models relative to black-box deep learning and standard numerical solvers, particularly when labeled data are scarce or the physical domain is complex, high-dimensional, or noisy (Ahmadi et al., 6 Oct 2025, Zeng et al., 29 Jan 2026, Doumèche et al., 2024, Shi et al., 2022).

7. Open Challenges, Limitations, and Future Directions

Despite rapid progress, key open issues remain:

• Optimization and Conditioning: PINNs and similar models may suffer from ill-conditioned optimization landscapes (due to high-order derivatives), spectral bias, or pathological convergence properties. Curriculum-based scheduling and learned neural solvers for parametric PDEs offer promising remedies (Marcandelli et al., 2 Feb 2026, Boudec et al., 2024).
• UQ and Generalization: Systematic uncertainty quantification, both epistemic and aleatoric, and principled assessment of out-of-distribution generalization capabilities, especially in singular learning problems, require further theoretical and empirical study (Barajas-Solano, 11 Feb 2026, Ahmadi et al., 6 Oct 2025).
• Architectural Search and Hyperparameter Selection: Optimal balancing between data and physics penalties, design of loss and architecture for multi-physics and high-dimensional systems, and scalable solvers for operator learning remain active research frontiers (Meng et al., 2022, Scampicchio et al., 29 Sep 2025).
• Probabilistic Inference and Bayesian Extensions: Efficient, physically-constrained posterior inference, probabilistic inversion, and generative modeling via diffusion (PILD, Pi-fusion) require deeper mathematical and computational developments (Zeng et al., 29 Jan 2026, Qiu et al., 2024).
• Integration with Foundation Models and LLMs: Modular integration of physics-informed solvers with LLM-based scientific agents for automated design and discovery is an emerging area of research (Ahmadi et al., 6 Oct 2025).
• Benchmarking and Standardization: Broader, standardized benchmark suites and comparative studies are necessary to drive progress and ensure reproducibility (Meng et al., 2022, Scampicchio et al., 29 Sep 2025, Shi et al., 2022).

Physics-informed learning methodologies are increasingly recognized as foundational in advancing the scientific rigor, transparency, and practical impact of modern machine learning in the physical sciences and engineering (Zeng et al., 29 Jan 2026, Meng et al., 2022, Ahmadi et al., 6 Oct 2025, Doumèche, 11 Jul 2025).