Physics-Informed Machine Learning

Updated 17 March 2026

Physics-informed machine learning is a computational paradigm that embeds known physical laws into ML models via tailored loss functions and network architectures.
It improves data efficiency and model generalization by integrating differential equations, conservation laws, and variational principles into the learning process.
The approach has practical applications in fluid dynamics, combustion, structural mechanics, and biomedical imaging, delivering enhanced speed and accuracy even in low-data regimes.

Physics-informed machine learning (PIML) integrates prior knowledge of physical laws—typically represented as differential equations, conservation laws, symmetries, or variational principles—directly into the construction, training, or deployment of machine learning models. Unlike purely data-driven approaches, PIML systematically embeds these constraints into models via loss penalization, architecture design, or operator-theoretic principles, enabling physically plausible predictions, improved data efficiency, and enhanced generalization, particularly in scientific, engineering, and industrial domains characterized by data scarcity or high model uncertainty.

1. Foundations and Motivations

The central objective of physics-informed machine learning is to unify the strengths of data-driven inference and mechanistic modeling by minimizing a composite cost that encodes both empirical fit and physical consistency. This paradigm is motivated by shortcomings of traditional deep learning in many scientific domains: when data is scarce or out-of-distribution, models can overfit, produce physically impossible solutions, or generalize poorly. PIML addresses these challenges by enforcing known physical laws—such as governing PDEs/ODEs, conservation constraints, symmetries, and variational principles—thus shifting model inductive bias from pure data interpolation toward physically valid solution spaces (Meng et al., 2022, Hao et al., 2022).

Key rationales for PIML include:

Data efficiency: Physics constraints reduce the effective hypothesis space, lowering sample complexity required for accurate learning.
Physical plausibility: Direct enforcement of conservation, invariance, or boundary conditions prevents egregiously unphysical outputs.
Robust generalization: Models are regularized to extrapolate beyond the training distribution, especially important in operational settings or for lifetime assessment scenarios (Hao et al., 2022, Nghiem et al., 2023).

2. Mathematical Frameworks and Model Classes

Physics-informed machine learning encompasses a variety of mathematical formulations, unified by the principle of embedding physical knowledge in the model or training objective. The canonical form for continuous systems is:

$\min_{f \in \mathcal{H}} L_{\text{data}}(f; \mathcal{D}) + \lambda_{\text{phys}} L_{\text{phys}}(f; \mathcal{F})$

where $L_{\text{phys}}$ enforces residuals of physical laws (e.g., PDEs), and $\lambda_{\text{phys}}$ weights their relative importance (Hao et al., 2022, Meng et al., 2022, Nghiem et al., 2023).

2.1 Physics-Informed Neural Networks (PINNs)

PINNs use standard neural network architectures to parameterize the solution to a system and enforce the governing equations through a differentiable loss (Toscano et al., 2024, Hao et al., 2022, Shukla et al., 2022). For a PDE $\mathcal{N}[u]=0$ , the loss typically takes the form

$L(\theta) = \frac{1}{N_\text{data}}\sum_{i=1}^{N_\text{data}} |u_\theta(x_i) - y_i|^2 + \frac{1}{N_\text{phys}}\sum_{j=1}^{N_\text{phys}} |\mathcal{N}[u_\theta](x_j)|^2 + ...$

with additional penalty terms for boundary or initial conditions (Nghiem et al., 2023, Toscano et al., 2024).

2.2 Operator Learning and Neural Operators

Neural operators, such as DeepONet and Fourier Neural Operator (FNO), learn mappings between function spaces, permitting rapid surrogate modeling of families of PDEs or parameterized boundary conditions (Toscano et al., 2024, Hao et al., 2022). Architecture design is often inspired by discretization schemes or operator-theoretic frameworks, with physics loss terms included in training (Chen et al., 2023).

2.3 Hybrid and Kernel-based Formulations

Physics-informed kernel methods cast the problem as regularized kernel regression in a reproducing kernel Hilbert space (RKHS) tailored by physical priors (Doumèche et al., 2024, Doumèche, 11 Jul 2025, Doumèche et al., 2024). For linear differential constraints $\mathscr{D}(f) \simeq 0$ , the RKHS norm is modified to penalize PDE violations: $\| f \|_\mathrm{RKHS}^2 = \lambda \| f \|_{H^s}^2 + \mu \| \mathscr{D}(f) \|_{L^2(\Omega)}^2$ yielding closed-form, globally optimal estimators and transparent convergence guarantees.

2.4 Extreme Learning Machines (ELM/PIELM)

PIELM replaces iterative backpropagation with single-layer, fixed-weight architectures trained using direct least-squares solution of the linearized physics-informed loss vector, providing huge gains in speed and precision for problems with analytically accessible forms (Zhuang et al., 24 Oct 2025, Guo et al., 1 Oct 2025).

3. Techniques for Encoding Physics

Different forms of physical knowledge are systematically embedded within PIML frameworks by several mechanisms:

Soft constraints: PDE, ODE, or algebraic residuals are penalized by adding their squared norms to the loss (Toscano et al., 2024, Shukla et al., 2022).
Hard constraints: Model architectures are built to exactly satisfy boundary conditions, invariants, or conservation laws (e.g., using trial functions, stream-function representations, or output transformations) (Wu et al., 3 Sep 2025, Hao et al., 2022).
Architectural priors: Physics-aware layers impose invariance/equivariance (e.g., E(n)-equivariant GNNs, Hamiltonian layers), variationally inspired structures, or domain decomposition for large-scale problems (Meng et al., 2022, Shukla et al., 2022).
Operator/Kernel construction: Choosing or learning kernels/operator architectures that admit analytic enforcement of physics (e.g., Green’s function kernels or spectral representations) (Doumèche et al., 2024, Doumèche et al., 2024).

The weighting of data vs. physical penalties ( $\lambda_\text{phys}$ ) is crucial and may be fixed, adaptively learned (e.g., via self-adaptive methods), or optimally selected by Bayesian model-evidence metrics such as the Physics-Informed Log Evidence (PILE) score (Daniels et al., 30 Oct 2025).

4. Application Domains and Benchmark Results

Physics-informed machine learning has achieved significant gains across numerous domains:

Fluid dynamics and turbulence: PINNs and neural operators provide mesh-free surrogates for Navier–Stokes and related PDEs, enabling simulation, system identification, and control with vastly reduced computational cost and robust generalization under data scarcity (Shukla et al., 2022, Toscano et al., 2024, Hao et al., 2022).
Combustion and reacting flows: PIML surrogates accelerate stiff chemical kinetics, reconstruct velocity and scalar fields from sparse data, and perform closure modeling in RANS/LES with improved physical consistency and uncertainty quantification (Wu et al., 3 Sep 2025).
Structural mechanics and materials: PIML frameworks model time-dependent, nonlinear, and ill-posed inverse problems (e.g., grade prediction in flotation (Nasiri et al., 2024), load identification (Kapoor et al., 2023), soil-pile interaction (Guo et al., 1 Oct 2025)) by embedding known ODE/PDE structure and leveraging measurement data.
Biomedicine and imaging: PINNs, neural ODEs, and neural operators produce interpretable models for biofluid dynamics, elastography, pharmacokinetics, and medical imaging, often achieving sub-voxel or sub-millisecond accuracy where traditional data-driven or physics-only solvers fail (Ahmadi et al., 6 Oct 2025, Toscano et al., 2024).

Benchmark studies report orders-of-magnitude speedups over classical solvers, accuracy gains of 2–3× or more in low-data regimes, and improved robustness to noisy or adversarial data (Doumèche, 11 Jul 2025, Doumèche et al., 2024).

5. Computational Strategies and Scalability

Scaling PIML to large, multiscale, or high-dimensional problems introduces unique computational demands (Shukla et al., 2022):

Domain and time decomposition: Methods such as XPINN/cPINN, parareal-PINN, or operator learning partition complex domains or long time intervals, allowing parallel/distributed training and adaptation to local stiffness.
Graph-based approaches: Physics-informed graph networks (PIGNs) employ graph exterior calculus and message-passing to impose discrete conservation laws on arbitrarily structured meshes or systems-of-systems (Shukla et al., 2022).
Kernel and random feature approximations: Truncated Fourier or Nyström methods accelerate kernel-based estimators, especially for moderate-dimensional problems (Doumèche et al., 2024).
Self-adaptive and multi-fidelity training: Loss component weights and collocation points may be dynamically reweighted based on error gradients, balancing convergence and stability (Toscano et al., 2024, Wu et al., 3 Sep 2025).

The curse of dimensionality remains a significant limitation, with most approaches scaling well up to d ≈ 5–10 for PINNs, kernel methods, or mesh-free neural operators (Chen et al., 19 Jan 2026).

6. Uncertainty Quantification and Model Selection

Rigorous uncertainty quantification is essential for trust in PIML predictions, particularly in safety-critical or high-regulation domains (Daniels et al., 30 Oct 2025). Key techniques include:

Bayesian PINNs and neural ODEs: Posterior over weights or solutions via MCMC, variational inference, or ensemble models.
Physics-informed Gaussian processes: Joint modeling of solution and PDE residuals provides fast, closed-form inference and calibrated predictive intervals.
Marginal likelihood diagnostics (PILE score): The normalized negative log-evidence of the joint GP posterior integrates fit and complexity, providing a principled hyperparameter selection criterion. Data-free PILE enables a priori kernel assessment for a given PDE (Daniels et al., 30 Oct 2025).

Uncertainty-aware metrics are increasingly used to prevent overfitting to either data or physics constraints, and to guard against ill-posed solutions or spurious generalization.

7. Open Challenges and Future Research Directions

Active research areas and unresolved issues in PIML include:

Automated physics-prior selection: Moving beyond expert-crafted priors toward data-driven or neural-architecture search for optimal physical constraints (Meng et al., 2022, Toscano et al., 2024).
Unified benchmarks: Absence of widely accepted, cross-domain benchmark suites limits reproducibility and fair model comparison (Hao et al., 2022).
Optimization and training pathology: High-order derivatives, nonconvex loss landscapes, and spectral bias hinder efficient convergence; advances in activations, adaptive loss-weighting, and domain decomposition are actively pursued (Toscano et al., 2024).
High-dimensional scalability: Leveraging low-dimensional solution manifolds or compressed/physics-adapted network representations remains a key priority (Chen et al., 19 Jan 2026, Doumèche et al., 2024).
Generalization under mismatch and noise: Methods for robust inference and model misspecification quantification, as in Bayesian model checking or via discrepancy NNs, are under development (Ahmadi et al., 6 Oct 2025).
Integration with LLMs and foundation models: Combining PIML with language-based code synthesis and symbolic reasoning offers workflows for automated discovery and model assembly (Ahmadi et al., 6 Oct 2025).

Continued progress in these areas is expected to establish physics-informed machine learning as a central methodology in computational science and engineering.

References:

(Meng et al., 2022, Hao et al., 2022, Nghiem et al., 2023, Shukla et al., 2022, Doumèche et al., 2024, Doumèche, 11 Jul 2025, Doumèche et al., 2024, Daniels et al., 30 Oct 2025, Toscano et al., 2024, Wu et al., 3 Sep 2025, Zhuang et al., 24 Oct 2025, Guo et al., 1 Oct 2025, Chen et al., 19 Jan 2026, Nasiri et al., 2024, Ahmadi et al., 6 Oct 2025, Ghosh et al., 2021, Chen et al., 2023, Kapoor et al., 2023, Raymond et al., 2021).