Physics-Informed GANs: Constrained Adversarial Models

• Physics-informed GANs are generative adversarial models that integrate physical laws such as differential equations and conservation principles to produce physically admissible outputs.
• They modify the adversarial training by incorporating physics-based losses in the generator and/or discriminator, steering the generated samples closer to realistic physics.
• Applications span fluid mechanics, stochastic PDEs, and biomechanics, where enforcing physics constraints improves simulation fidelity and prediction accuracy.

A physics-informed generative adversarial network is a generative adversarial framework in which the adversarial game is constrained, guided, or reweighted by known physical structure, such as governing differential equations, conservation laws, constitutive relations, boundary conditions, simulator outputs, or physics-derived summary scores. In the literature, the term does not denote a single canonical architecture. Instead, it covers several closely related constructions: generator losses augmented with residual penalties from governing equations; discriminators that receive physics consistency scores; adversarial schemes in which an external physical model labels generated samples as acceptable or unacceptable; and hybrid systems in which surrogate physics models, adaptive samplers, or latent-space conditioning enforce physically meaningful behavior. Across these variants, the shared objective is to move generation away from purely data-distribution matching toward samples that lie closer to a manifold of physically admissible solutions (Yang et al., 2018).

1. Conceptual foundations and historical development

The central motivation for physics-informed GANs is that ordinary GANs can generate samples that are plausible in data space yet inconsistent with the underlying physical system. This concern appears across fluid mechanics, stochastic differential equations, mechanics, geophysics, traffic modeling, precipitation nowcasting, wind super-resolution, porous media, plant remote sensing, and biomechanics. In each case, the inadequacy of purely visual or statistical realism is emphasized: turbulent fields may violate incompressibility or pressure relations, multispectral plant imagery may have unrealistic band covariation, precipitation forecasts may violate moisture conservation, and stochastic PDE samples may not satisfy governing operators (Subramaniam et al., 2020).

A major early formulation is the stochastic-differential-equation PI-GAN, which embeds the governing operator into the architecture through automatic differentiation and uses Wasserstein GAN with gradient penalty for stability. In that framework, two feed-forward generators represent stochastic fields such as $u(x;\omega)$ and $k(x;\omega)$, while induced networks are obtained by applying the differential operator and boundary operator to the generator outputs, allowing forward, inverse, and mixed stochastic problems to be handled within one adversarial formulation (Yang et al., 2018).

Subsequent work broadened the concept in several directions rather than converging to a single standard. Some papers retained the PINN-like idea of adding a residual term to the generator objective, as in inverse elastic-modulus estimation and turbulence enrichment (Warner et al., 2020). Others relocated physics from the generator loss to the discriminator or an external judge, as in Physics-Informed Discriminator GAN for precipitation nowcasting and physics-guided GANs for black-box or non-differentiable physical models (Yin et al., 2024). Still others used auxiliary discriminators or pretrained surrogate models to evaluate structural efficiency or other physical objectives during generation (Lourenço et al., 17 Apr 2026).

This diversity suggests that “physics-informed” in the GAN literature is broader than the residual-penalty interpretation inherited from PINNs. A plausible implication is that the field is best understood as a family of adversarial learning schemes in which physics enters wherever it can most effectively regularize generation: the generator, the discriminator, the latent space, the training sampler, or an external simulation loop.

2. Core architectural patterns

A recurrent pattern is the residual-augmented generator. In “Turbulence Enrichment using Physics-informed Generative Adversarial Networks” (Subramaniam et al., 2020), the GAN-based model TEGAN adapts SRGAN to turbulence super-resolution. The generator is a deep residual network with convolutional layers and batch normalization inside each residual block, and the discriminator is a deep CNN ending in fully connected layers for binary real/fake classification. Its generator loss explicitly mixes reconstruction, enstrophy sensitivity, governing-equation residuals, and adversarial learning: $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$

$$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$

Here $\mathcal{L}_\mathrm{content}$ includes both $\mathcal{L}_\mathrm{MSE}$ and an enstrophy-sensitive term, while $\mathcal{L}_\mathrm{physics}$ penalizes continuity and pressure-Poisson residuals.

A second pattern is the physics-informed discriminator. In precipitation nowcasting, PID-GAN uses a VQ-GAN and Transformer as the generator and a temporal discriminator that is augmented with physics consistency scores derived from a moisture-conservation residual (Yin et al., 2024). The discriminator loss is written as

$$\mathcal{L}_D(\phi) = - \frac{1}{N} \sum_{i=1}^{N} \log D(x_i, \eta_i) - \frac{1}{N} \sum_{i=1}^{N} \log \big(1 - D(\hat{x}_i, \hat{\eta}_i)\big),$$

so the discriminator judges both sequence realism and physical plausibility.

A third pattern is the external-physics referee. In PG-GAN, the physical model remains outside the computational graph. A generated sample is placed in

$$\mathcal{R}_{\varepsilon} = \left\{ x \mid P(x \mid \theta) \leq \varepsilon \right\}$$

or

$$\mathcal{F}_{\varepsilon} = \left\{ x \mid P(x \mid \theta) > \varepsilon \right\},$$

depending on whether the residual $k(x;\omega)$0 lies below a threshold. The discriminator then learns from this physically acceptable/unacceptable partition, and the generator is optimized only against the unphysical subset (Yonekura, 2023).

A fourth pattern is the surrogate-critic architecture. In the design of funicular shells, a modified DCGAN named AD-DCGAN uses three discriminative components: $k(x;\omega)$1, $k(x;\omega)$2, and a pretrained PB-PUNet surrogate acting as $k(x;\omega)$3. The auxiliary branch predicts physical response fields so that the membrane factor can be evaluated, thereby penalizing geometries dominated by bending (Lourenço et al., 17 Apr 2026).

A fifth pattern uses the GAN not to generate physical states directly but to generate informative training points. PhyTF-GAN combines a decoder-only Transformer with a GAN-based adaptive sampler. The generator maps Gaussian noise to candidate spatiotemporal points, the discriminator distinguishes high-residual points from generated ones, and accepted generated points are added to the PDE training set (Zhang et al., 15 Jul 2025). This is still a physics-informed GAN, but its object of generation is the collocation distribution rather than the solution field.

3. Mechanisms for embedding physics

The most direct mechanism is residual penalization derived from governing equations. In incompressible turbulence enrichment, the relevant constraints are

$k(x;\omega)$4

and the pressure Poisson relation

$k(x;\omega)$5

The corresponding losses are residual-based mean-squared penalties computed from derivatives of the generated output, and the paper emphasizes that this improves the model’s ability to generate data that satisfy the governing equations (Subramaniam et al., 2020).

A related mechanism is PDE-constrained inverse generation. In elastic-modulus estimation, the PI-GAN uses two generators, one for displacement and one for stiffness, and derives equilibrium and boundary-residual networks by automatic differentiation. The total generator objective is

$k(x;\omega)$6

so the unobserved modulus field is inferred indirectly through measured deformation data and the mechanics constraints (Warner et al., 2020).

Another mechanism operates through physics-derived scores rather than direct residual minimization. In precipitation nowcasting, a residual $k(x;\omega)$7 is mapped to

$k(x;\omega)$8

and these scores are appended to the discriminator input. The physical rule is a moisture conservation equation, simplified using horizontal winds at 10 m and 100 m because vertical wind is difficult to measure reliably (Yin et al., 2024).

Physics can also shape the latent space. PlantPlotGAN uses an extension of DCGAN with one generator, two discriminators, an optimizer, and a spectral regularization module. Its physical prior is the empirical relationship between red-edge and NIR reflectance,

$k(x;\omega)$9

and a second discriminator compares spectral profiles to penalize spectral inconsistency between real and synthetic multispectral imagery (Lopes et al., 2023).

A still looser but practically important mechanism is simulator-in-the-loop conditioning. In 3D porous media generation, a pretrained 3D WGAN-GP first learns realistic sandstone microstructures, after which a pore-network model computes porosity, permeability, specific surface area, Euler characteristic, and pore/throat sizes. The latent vector is then steered through gradual Gaussian deformation,

$$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$0

until the generated sample matches a target physical response within tolerance (Ren et al., 2024). This is physics-informed without requiring the simulator to be differentiable.

These mechanisms are often conflated, but the literature distinguishes them sharply. Residual-based PI-GANs resemble PINNs in the way physics is encoded; discriminator-based models make physical plausibility part of adversarial judgment; physics-guided models can treat the physical model as an external judge; and simulator-guided models may never insert the governing equations into the GAN graph at all. The commonality is not the exact location of the constraint, but the use of physics to reshape the adversarial learning signal.

4. Representative application domains

Fluid mechanics is among the clearest application areas. TEGAN reconstructs $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$1 turbulent fields from $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$2 coarse inputs in forced, incompressible, homogeneous isotropic turbulence, with velocity and pressure as output channels. The paper reports improved recovery of finer-scale velocity and pressure features, energy spectra closer to DNS than interpolation, closely matching longitudinal and transverse two-point correlations, improved third-order statistics relative to interpolation, and more realistic $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$3-$$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$4 morphology (Subramaniam et al., 2020). Related subgrid work in reactive flows treats LES closure as an inverse-filter problem: PIESRGAN reconstructs high-resolution turbulent structure from filtered fields so that subgrid terms can be evaluated from the reconstruction (Bode et al., 2019). Wind-field super-resolution extends this paradigm to atmospheric data with a pixel-wise self-attention SR-GAN that upsamples 900 m wind outputs to 100 m resolution while regularizing the attention map to capture vertical convection (Kurihana et al., 2023).

Stochastic differential equations and uncertainty quantification form another substantial branch. The original PI-GAN for SDEs uses WGAN-GP and automatic differentiation to solve forward, inverse, and mixed stochastic problems from limited scattered measurements (Yang et al., 2018). WGAN-PINNs extend this to uncertainty quantification in PDEs with uncertain initial and boundary data, using GroupSort critics to satisfy 1-Lipschitz constraints required by Wasserstein theory (Gao et al., 2021). PI-VEGAN introduces a variational encoder to infer latent variables from the actual measurements before adversarial training (Gao et al., 2023), and PI-GEA moves matching into a lower-dimensional latent feature space using an encoder updated alternately with the generator (Gao et al., 2023).

Mechanics and structural design supply several distinct examples. Inverse elastic-modulus estimation learns the probability distribution of a spatially varying stiffness field from deformation measurements without direct modulus labels (Warner et al., 2020). A separate line integrates Kirchdoerfer–Ortiz data-driven mechanics with GANs by making the generator a PINN over displacement and stress while the discriminator uses the closest strain–stress data point to assess authenticity (Ciftci et al., 2023). In shell design, AD-DCGAN uses the membrane factor

$$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$5

as the core physics criterion to bias generation toward membrane-dominated funicular shells (Lourenço et al., 17 Apr 2026).

Earth-system and environmental forecasting provide discriminator-centric formulations. PID-GAN for precipitation nowcasting combines a VQ-GAN, an autoregressive Transformer, and a temporal discriminator receiving physics consistency scores from moisture conservation, and it is explicitly motivated by the difficulty of extreme-event prediction for both numerical weather prediction and purely data-driven deep generative models (Yin et al., 2024). In geophysical imaging, FWIGAN replaces the usual neural generator with the acoustic wave-equation operator itself and trains a WGAN-GP critic on simulated versus observed shot gathers, turning full waveform inversion into an unsupervised adversarial distribution-matching problem (Yang et al., 2021).

Additional application areas show that “physics” in these models is domain specific rather than tied to PDEs alone. PlantPlotGAN constrains multispectral UAV imagery by red-edge/NIR relationships and FFT-based spectral regularization to improve early yellow-rust detection (Lopes et al., 2023). TrafficFlowGAN combines a normalizing-flow generator, a convolutional discriminator, and PIDL regularization for uncertainty quantification in traffic state estimation (Mo et al., 2022). A biomechanics model integrates Lagrange’s equation of motion and an inverse dynamic muscle model into a GAN with policy-gradient-style updates for low-shot estimation of muscle force and joint kinematics from sEMG (Shi et al., 2023).

5. Empirical behavior and evaluation practice

Evaluation protocols in this literature are notably heterogeneous because the success criterion depends on whether the target is a physical field, a stochastic distribution, a design geometry, or a downstream decision task. In turbulence enrichment, performance is evaluated not only by field-level visualization but by velocity energy spectra, longitudinal and transverse two-point correlations, triple correlations, and $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$6-$$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$7 diagrams (Subramaniam et al., 2020). The reported dev and test losses show that TEGAN keeps content loss similar to TEResNet while reducing physics loss by more than 10%, with dev physics improving from $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$8 to $$\mathcal{L}_\mathrm{GAN} = \left( 1 - \lambda_\mathrm{A} \right) \mathcal{L}_\mathrm{resnet} + \lambda_\mathrm{A} \mathcal{L}_\mathrm{adversarial},$$9 and test physics from $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$0 to $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$1 (Subramaniam et al., 2020).

In precipitation nowcasting, evaluation shifts to forecasting metrics and rare-event skill. PID-GAN is assessed using PCC, MAE, MSE, CSI, FAR, FSS, and precision-recall AUC for extreme-event detection. The paper states that removing the physics constraint in PID-GAN(-P) reduces extreme-event AUC by 6.17%, indicating that the physics component materially improves detection of extremes (Yin et al., 2024).

For stochastic PDEs, distributional fidelity is central. WGAN-PINNs report means, uncertainty bands, histograms at selected points, relative $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$2 errors, and empirical Wasserstein-type losses, and emphasize that stronger discriminators generally improve performance in accordance with theory (Gao et al., 2021). PI-VEGAN and PI-GEA compare means, standard deviations, covariance spectra, Wasserstein or MMD-based discrepancies, and relative $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$3 errors across forward, inverse, and mixed problems (Gao et al., 2023). The original PI-GAN paper also highlights generator overfitting in addition to discriminator overfitting, an observation it states had previously been reported mainly for the discriminator (Yang et al., 2018).

In plant remote sensing, image realism is evaluated with Fréchet Inception Distance, Chi-square, Intersection Coefficient, Bhattacharyya Coefficient, and mean spectral-profile comparison, because visual plausibility alone is insufficient when downstream prediction depends on vegetation indices (Lopes et al., 2023). The paper reports improvements in downstream disease classification when real and synthetic imagery are combined: for example, XGBoost improves from $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$4 to $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$5 F1 and from $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$6 to $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$7 accuracy (Lopes et al., 2023).

In porous media, evaluation includes visual comparison of slices, box plots of property distributions, Pearson correlation matrices, scatter plots, normalized RMSE for constrained generation, and runtime per conditioned sample. The method reports normalized RMSE values of $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$8 for porosity, $$\mathcal{L}_\mathrm{resnet} = \left( 1 - \lambda_\mathrm{P} \right) \mathcal{L}_\mathrm{content} + \lambda_\mathrm{P} \mathcal{L}_\mathrm{physics}.$$9 for permeability, $\mathcal{L}_\mathrm{content}$0 for mean pore size, and $\mathcal{L}_\mathrm{content}$1 for mean throat size, while preserving the porosity–permeability correlation at approximately $\mathcal{L}_\mathrm{content}$2 compared with $\mathcal{L}_\mathrm{content}$3 in real data (Ren et al., 2024).

A recurrent empirical pattern is that adversarial physics constraints tend to improve either high-frequency realism, typical residual statistics, or out-of-distribution plausibility, but not always every metric simultaneously. This suggests that model selection in physics-informed GANs is intrinsically multi-objective: fidelity to data, fidelity to governing physics, stability of training, and usefulness for downstream scientific or engineering tasks are often partially competing objectives rather than a single scalar target.

6. Distinctions, limitations, and open directions

A common misconception is that a physics-informed GAN is simply a GAN with a PDE residual appended to the generator loss. The literature is broader. PG-GAN was proposed precisely because PINN-style residual minimization requires the physical equations to be embedded in the differentiable model, which can be restrictive when the physics comes from black-box, commercial, or non-differentiable solvers (Yonekura, 2023). PID-GAN, by contrast, injects physics through discriminator inputs rather than generator penalties (Yin et al., 2024). Porous-media conditioning uses post-training latent optimization through a pore-network simulator rather than end-to-end adversarial differentiation through the simulator (Ren et al., 2024).

Training stability is a pervasive difficulty. TEGAN notes that excessive physics weighting can lead to a trivial zero-field local minimum, so $\mathcal{L}_\mathrm{content}$4 must be carefully balanced, and it explicitly suggests WGAN-GP as a possible route to improved stability (Subramaniam et al., 2020). PG-GAN requires pretraining with an ordinary GAN because $\mathcal{L}_\mathrm{content}$5 may be empty at the start of training; it also relies on threshold scheduling to avoid signal collapse (Yonekura, 2023). GA-PINN improves standard PINNs by adversarially exploiting a few exact solution samples, yet its associated point-weighting method can interfere with adversarial dynamics for difficult PDEs such as Burgers and Schrödinger (Li et al., 2022). PI-VEGAN improves stability relative to PI-WGAN, but at the price of more parameters and more computation because of the encoder (Gao et al., 2023).

Another limitation is domain specificity. PlantPlotGAN only enforces covariance between sequential bands, specifically red-edge and NIR, and the extension to combinatorial covariance across all multispectral bands is identified as open (Lopes et al., 2023). The wind super-resolution model is described as preliminary and uses a relatively lightweight physics prior centered on attention-map regularization for vertical convection rather than full PDE constraints (Kurihana et al., 2023). In porous media, the output size is fixed at $\mathcal{L}_\mathrm{content}$6, larger-scale heterogeneity is not captured, and the current scale may be too small for robust estimation of more complex flow functions such as relative permeability (Ren et al., 2024). WGAN-PINNs explicitly acknowledge a theoretical vacancy: the theory bounds the training loss but does not guarantee that uncertainty propagated into the interior of the domain is correct (Gao et al., 2021).

Open directions recur across papers. Physics-aware discriminators are repeatedly proposed as a route to stronger guidance. TEGAN suggests broader turbulence training distributions and physics-aware discriminator objectives (Subramaniam et al., 2020). PlantPlotGAN notes that a per-channel discriminator could improve quality and that StyleGAN-style approaches may help if richer class-conditional multispectral data become available (Lopes et al., 2023). AD-DCGAN shows one route to such richer discrimination through multiple branches, including shell, mask, and auxiliary physics critics (Lourenço et al., 17 Apr 2026). PhyTF-GAN points toward adversarial training that acts on the sampling distribution rather than only on solution generation, particularly for multiscale, time-dependent PDE systems in which residuals are highly localized and temporal causality matters (Zhang et al., 15 Jul 2025).

Taken together, these works indicate that the defining feature of a physics-informed GAN is not a fixed architecture but a design principle: the adversarial learning loop is modified so that the generated outputs, generated sampling points, or latent codes are judged not only by empirical realism but also by whether they satisfy physically meaningful structure. This suggests a unifying interpretation of the field: physics-informed GANs are adversarial models for constrained generation, where the constraints may come from equations, simulators, conservation laws, constitutive data, or domain-specific invariants, and where the most effective place to impose those constraints remains application dependent.