Physics-Constrained Deep Learning

• Physics-Constrained Deep Learning is an approach that integrates governing physical laws, like PDEs and conservation laws, directly into neural network architectures.
• It employs both hard constraints (through architectural modifications) and soft constraints (via loss function penalties) to enforce boundary conditions and physical symmetries.
• This method achieves state-of-the-art results in surrogate modeling, uncertainty quantification, and inverse problem solving across diverse scientific and engineering domains.

Physics-constrained deep learning is an approach in which the structure, constraints, and loss function of a neural network are systematically augmented or designed to ensure that predictions are consistent with established laws of physics, typically in the form of partial differential equations (PDEs), algebraic equations, or conservation laws. This class of models is distinguished from conventional, purely data-driven deep learning by the explicit embedding of governing equations, boundary conditions, or physical symmetries, either as “hard” constraints—enforced by architectural modifications—or “soft” constraints imposed via regularization terms in the loss functional. Physics-constrained methods have achieved state-of-the-art results in a diverse range of scientific tasks where fidelity to governing laws is essential, covering high-dimensional surrogate modeling, uncertainty quantification, inverse problem solving, and efficient solution of PDE-governed forward problems.

1. Fundamental Concepts and Methodologies

The defining feature of physics-constrained deep learning is the explicit imposition of physical laws as constraints during training and inference, fundamentally changing both the learning objective and, often, the architecture of the neural models.

Loss function structure: The canonical objective is an additive (or augmented Lagrangian) combination of data-driven terms and physics-inspired residuals. For a network $\hat{u}_\theta$ approximating a state variable $u(x)$, with PDE operator $\mathcal{F}$ and available data $u^{\text{data}}$, the loss is typically written as

$$\mathcal{L}_\text{total} = \mathcal{L}_\text{data}\bigl(\hat{u}_\theta,u^{\text{data}}\bigr) + \lambda_\text{phys} \mathcal{L}_\text{phys}\bigl(\mathcal{F}[\hat{u}_\theta]\bigr),$$

where $\mathcal{L}_\text{phys}$ penalizes residual violation of the physics (e.g., PDE or ODE).

Hard and soft constraints: Constraints may be “hard”—incorporated by modifying the output layer to enforce boundary conditions identically (e.g., via auxiliary functions or stream function ansatz)—or “soft,” i.e., penalized in the loss function with carefully chosen weights (McDevitt et al., 2022, Sun et al., 2019, Golder et al., 5 Dec 2025, Basir et al., 2021).

Automatic differentiation: All necessary spatial and temporal derivatives for PDE residuals are computed via automatic differentiation through the network, facilitating seamless backpropagation even for high-order, coupled systems.

Projection and constrained optimization: Advanced frameworks formulate the training as a constrained optimization problem, e.g., via Karush-Kuhn-Tucker (KKT) conditions and augmented Lagrangian methods to rigorously enforce equality and inequality constraints (Golder et al., 5 Dec 2025, Basir et al., 2021).

Model architectures: A broad range of network architectures are employed, including fully connected networks for mesh-free surrogates (Sun et al., 2019, McDevitt et al., 2022), convolutional encoders for high-dimensional fields (Zhu et al., 2019), U-Nets with physics-inspired smoothing layers (Yan et al., 2021), and sequential or recurrent architectures for parametric and temporal modeling (Drgona et al., 2020, Qi et al., 8 Mar 2026).

2. Physics Constrained Losses and Constraint Enforcement

Physics-constrained deep learning frameworks integrate diverse categories of physics constraints spanning algebraic, differential, and statistical relationships:

PDE residuals: The physics loss typically consists of the squared norm of the residual of the governing equations sampled at collocation points. For incompressible Navier–Stokes, as in cavity flow surrogates (McDevitt et al., 2022), momentum and continuity residuals are enforced by automatic differentiation:

$$R_u = \bigl[u\,\partial_x u + v\,\partial_y u\bigr] + \partial_x p - \frac{1}{Re}\nabla^2 u$$

with $L_\text{phys}$ aggregating the squared residuals over many collocation points.

Boundary and initial conditions: Hard enforcement is achieved via output transformations incorporating analytic masks or stream functions so that Dirichlet/Neumann conditions are identically satisfied (McDevitt et al., 2022, Sun et al., 2019). Alternatively, boundary conditions are imposed via additional penalty terms or as Lagrangian constraints in an augmented optimization scheme (Basir et al., 2021).

Thermodynamic and analytic relationships: Physical relationships such as the ideal gas law, Clausius–Clapeyron, or stoichiometric constraints are encoded in the loss (e.g., to ensure that postprocessed temperature and humidity predictions respect $p=\rho R_d T$ and $q \leq q_\text{sat}$, or that predicted CT numbers are consistent with material density) (Zanetta et al., 2022, Chang et al., 2022).

Equality and inequality constraints: State or output variables are bounded within physically plausible ranges via slack-variable penalties or projection operators (e.g., to ensure temperature remains within operational bounds, or Reynolds stresses are realizable) (Drgona et al., 2020, Golder et al., 5 Dec 2025, Chu et al., 2024).

Statistical and generative constraints: When surrogate models are designed to produce distributions (e.g., for uncertainty quantification), loss functions are constructed as reverse Kullback-Leibler divergence with respect to a Boltzmann–Gibbs measure constructed from the physics loss (Zhu et al., 2019).

3. Applications Across Domains

Physics-constrained deep learning has been developed for and applied to a wide variety of scientific and engineering problems, including:

• Computational Fluid Dynamics: Surrogate modeling of fluid flows governed by Navier–Stokes, enabling mesh-free, high-dimensional predictors across parameter spaces (e.g., Reynolds number, geometry parameters) with negligible online evaluation cost (McDevitt et al., 2022, Sun et al., 2019).
• Hydrology and Soil Physics: Modeling soil moisture dynamics under physical flow laws (Richards’ equation), with joint data and physics constraints yielding field-scale moisture reconstructions from sparse sensor data (Xie et al., 2024, 2403.07228).
• Multiphase Porous Media Flow: Fast emulation of high-fidelity compositional simulators for multiphase flows in heterogeneous 3D reservoirs, using CNN-based architectures with spatial smoothers reflecting local continuity (Yan et al., 2021).
• Turbulence Model Uncertainty Quantification: Integration of realizable eigenspace perturbation frameworks with CNN-based marker functions to adaptively quantify and reduce uncertainty in RANS turbulence predictions (Chu et al., 4 Sep 2025, Chu et al., 2024).
• Building and Energy Systems: Structured deep state-space or gray-box models of building thermal dynamics and distributed “virtual battery” energy storage systems, ensuring long-term stability and physical parameter recovery (Drgona et al., 2020, Qi et al., 8 Mar 2026).
• Medical and Imaging Inverse Problems: Deep learning for mass density estimation in medical imaging constrained by differentiable physical models linking image data to material properties (Chang et al., 2022), or for cardiac inverse problem regularization with electrophysiological PDEs (Xie et al., 2021).
• Geophysical Forecasting and Scientific Discovery: Learning bounded, interpretable embeddings for partially observed dynamical systems using neural ODEs with Lyapunov-type energy constraints (Ouala et al., 2022), as well as physics-informed generative models for surrogate modeling and uncertainty propagation (Zhu et al., 2019, Liu et al., 2023).
• Physics-constrained Active Learning: Coupling physics-regularized surrogates with adaptive sensor placement strategies to minimize posterior uncertainty, quantified by physics residuals (2403.07228).

4. Quantitative Performance and Regime Limitations

Physics-constrained deep learning surrogates have demonstrated strong, sometimes state-of-the-art, quantitative performance across diverse benchmarks:

• In incompressible cavity flows, mesh-free surrogates achieve stream-function errors <1% for $u(x)$0, and <5% up to $u(x)$1 in the absence of data; with sparse data augmentation, accurate predictions are achieved up to $u(x)$2 (McDevitt et al., 2022).
• For soil moisture estimation, physics-constrained surrogates trained with full-batch Adam optimization achieve relative water-content errors below 0.1%, outperforming unconstrained DNNs and physics-only PDE solvers under sensor scarcity (Xie et al., 2024).
• In turbulence model UQ, CNN-guided physics-constrained corrections reduce kinetic-energy profile errors by 1–2 orders of magnitude over baseline RANS, producing sharp, well-calibrated uncertainty envelopes (Chu et al., 4 Sep 2025, Chu et al., 2024).
• For multiparametric imaging inversion, inclusion of differentiable physics constraints yields mean percentage errors below 1% for mass density in soft-tissue surrogates, dramatically reducing systematic biases versus unconstrained approaches (Chang et al., 2022).

Nevertheless, limitations persist. For instance, purely physics-constrained models may struggle to resolve small-scale features (e.g., tertiary vortices in high-Reynolds flows) and performance deteriorates as flow parameters leave the training regime (McDevitt et al., 2022). For DAEs, hard projection layers may require careful Taylor expansion/linearization and may be bypassed only when learned states are already near-physical (Golder et al., 5 Dec 2025). For data-scarce or ill-posed settings, physics constraints serve as strong regularizers but may not fully compensate for missing data in highly heterogeneous or out-of-distribution regimes.

5. Architectural Innovations and Training Strategies

Physics-constrained deep learning has generated a diverse design space of network architectures and optimization methods:

• Stream function and output masking: Incorporate physically invariant variable transforms at the very last layer to identically ensure incompressibility or boundary conditions (McDevitt et al., 2022, Sun et al., 2019).
• Differentiable physics layers: Embed ODE/PDE integrators (or non-trainable update rules derived from system dynamics) as differentiable modules, so physical consistency is exact or controllable (e.g., in VB-NET for virtual battery dynamics) (Qi et al., 8 Mar 2026).
• Solver-inspired smoothers: Mimic key numerical stencil operations as post-processing layers to improve continuity or local conservation (e.g., 3×3 kernels for local pressure averaging in porous media flow) (Yan et al., 2021).
• Augmented Lagrangian and KKT projections: Employ projection via solution of constraints at each forward pass (DAE-HardNet, PECANN), granting orders of magnitude lower physics-violation residual than penalty-based alternatives (Golder et al., 5 Dec 2025, Basir et al., 2021).
• Multi-task learning and structured separation: Separate encoding of global “shared” drivers (such as meteorology) from “private” local fingerprints, aiding transfer learning and cold-start adaptability (e.g., VB-NET) (Qi et al., 8 Mar 2026).
• Active learning integration: Sequential sensor selection guided by physics residual maps and space-filling scores, minimizes resource cost for field monitoring without compromising physical fidelity (2403.07228).

Training strategies are context-dependent but often exploit hybrid schedules (Adam followed by L-BFGS), adaptive sampling (e.g., resampling points weighted by physics residual), and hyperparameter optimization (e.g., Bayesian search for data/physics weights).

6. Broader Implications, Generalization, and Future Directions

Physics-constrained deep learning combines data-driven flexibility with principled imposition of physical law, resulting in physically trustworthy, interpretable, and often data-efficient surrogates. Such models are particularly robust when labeled data is scarce or expensive, as physical losses can act as inductive priors that promote generalization even in out-of-distribution settings (Zanetta et al., 2022, Zhu et al., 2019).

Major implications include:

• Rapid UQ and design optimization: Surrogates can be evaluated for thousands of parameter samples in milliseconds, enabling uncertainty quantification, inverse design, and real-time control (Zhu et al., 2019, McDevitt et al., 2022, Yan et al., 2021).
• Interpretability and parameter recovery: Many models reveal interpretable mappings between learned parameters and physical quantities (e.g., capacity scaling in virtual battery models, affinity with thermodynamic laws) (Qi et al., 8 Mar 2026).
• Improved trust and downstream applicability: Physically valid surrogates ensure predictions are feasible for use in safety-critical or regulatory settings (e.g., clinical radiotherapy planning, grid energy management) (Chang et al., 2022, Qi et al., 8 Mar 2026).
• Blueprint for hybridization and modular extension: Hard constraints, physics-inspired smoothing, differentiable solvers, and data-driven pattern detection can all be integrated in modular, extensible ways across disparate scientific domains (Yan et al., 2021, Qi et al., 8 Mar 2026).

Key future directions include hybrid models coupling physics-based solvers for critical regions, adaptive residual point refinement, tensor-valued corrective architectures for turbulence, and further development of scalable, global optimization algorithms for highly constrained training in PDE-dominated systems.

7. State-of-the-Art Frameworks and Quantitative Overview

The following table synthesizes select physics-constrained deep learning frameworks, their key characteristics, and representative domains, illustrating the breadth of contemporary science applications:

Framework / Paper	Constraint Mechanism	Core Domain / Application
Hard BC/stream function (McDevitt et al., 2022Sun et al., 2019)	Output transformation	Incompressible flow surrogates
DAE-HardNet (Golder et al., 5 Dec 2025)	KKT-based projection	Hard DAE/PDE constraint, parameter estimation
VB-NET (Qi et al., 8 Mar 2026)	Differentiable physics layer, isomorphic mapping	Energy systems, virtual batteries
PECANN (Basir et al., 2021)	Augmented Lagrangian, equality constraints	Multi-fidelity, inverse/forward PDE
Physics smoothing (Yan et al., 2021)	Local convolutional smoother	Multiphase flow, porous media
Thermodynamic loss (Zanetta et al., 2022Chang et al., 2022)	Analytic physical penalty	Weather postprocessing, medical imaging
Eigenspace perturbation + CNN (Chu et al., 4 Sep 2025Chu et al., 2024)	Realizability + marker function	Turbulence model UQ
Neural ODE + Lyapunov (Ouala et al., 2022)	Structure / energy-based constraints	Partial observation, geophysical flow
PINF (Liu et al., 2023)	Characteristic ODE, change of variable	Fokker-Planck equations, density flows
P-DL active learning (2403.07228)	Residual-based sampling	Soil field monitoring

This scope highlights the maturation of physics-constrained deep learning as a unifying paradigm for embedding domain knowledge in neural models across computational science and engineering.