Conservation-Law Generative Networks

• Conservation-law-respecting generative networks are architectures that embed physical invariants to guarantee outputs remain on a physically admissible manifold.
• They employ various mechanisms such as soft penalties, projection methods, and hard-coded parameterizations to enforce both deterministic and statistical conservation laws.
• Empirical evaluations demonstrate improved prediction reliability and exact conservation metrics across applications like turbulence modeling, climate emulation, and probabilistic density generation.

Searching arXiv for recent and foundational papers on conservation-law-respecting generative networks and related physics-constrained generative modeling. arXiv search results for "conservation law respecting generative networks physics constrained generative models diffusion GAN PINN" retrieved and matched to the key papers used below, including (Baez et al., 2024, Yang et al., 2019, Tretiak et al., 2022, Hua et al., 23 Jun 2025, Liu et al., 2024, Zhang et al., 2023, Beucler et al., 2019, Ji et al., 3 Jan 2025, Zhu et al., 21 May 2026), and the broader Perspective (Banad et al., 8 Jun 2026). Conservation-law-respecting generative networks are generative architectures whose parameterization or dynamics are designed so that the relevant physical invariants hold automatically during generation, or are enforced so stringently that generated states remain on, or are driven toward, a physically admissible manifold. In the taxonomy of physics-informed generative AI, they are grouped with structure-preserving generative networks and placed in the hard-constrained generation category, where outputs are parameterized to satisfy constraints by construction rather than screened only after the fact (Banad et al., 8 Jun 2026). Across recent work, the conserved structure may be conservation of mass, conservation of charge, conservation of energy, momentum conservation, continuity laws, reaction-diffusion structure, Hamiltonian or conservative dynamics, or the Fokker–Planck conservation law for probability densities (Baez et al., 2024, Tretiak et al., 2022, Hua et al., 23 Jun 2025).

1. Conceptual scope and taxonomy

A central distinction in this literature is between deterministic constraints and statistical constraints. Deterministic constraints are properties that each individual sample should satisfy, including mass conservation, divergence-free velocity fields, exact geometric relations, and PDE residuals. Statistical constraints describe properties of an ensemble of samples, such as energy spectra, correlation functions, and probability distributions of increments. The physics-constrained GAN framework of Wu et al. is explicitly about deterministic yet imprecise constraints, written in the generic form

$\mathcal{H}[\mathbf{u}] \le 0,$

often with $\mathcal{H}$ chosen to be nonnegative so that this effectively means $\mathcal{H}[\mathbf{u}] = 0$ (Yang et al., 2019).

The same work also formalizes the idea of a tolerance band around the exact manifold: $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$ This makes the admissible set a band rather than a single exact surface. Physically, this is useful when the physics is only known approximately, when reduced-order models introduce residual error, or when numerical discretization prevents exact satisfaction (Yang et al., 2019).

At the opposite end of the spectrum, the 2026 semiconductor-manufacturing Perspective treats conservation-law-respecting generative networks as the clearest example of hard-constrained generation, emphasizing the structural principle that the generator does not search all outputs but only the physically admissible subset. The same source contrasts this with post-hoc filtering and states: “Post-hoc filtering does not improve the model. The next sample from the same prior has the same probability of failure. The model has not learned that the rejected region is infeasible; it has learned nothing, because the filter operates outside the gradient path” (Banad et al., 8 Jun 2026).

This taxonomy implies three recurrent design positions. Some models bias generation toward admissible states through penalties. Some project unconstrained predictions back to an invariant manifold. Others encode the law directly into the architecture so that conservation holds by construction. A plausible implication is that the field is organized less by model family—GAN, PINN, diffusion, or neural integrator—than by the location of the physical law in the computational graph.

2. Enforcement mechanisms: penalties, projections, and by-construction parameterizations

Soft-constraint methods attach a conservation residual to an otherwise standard generative objective. In the physics-constrained GAN formulation, the min–max game

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$

is modified to

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$

with the physical term applied only to the generator. In the climate-emulator setting, the same pattern appears as

$${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$

where ${\cal P}$ is the residual of a linear conservation system encoded by a sparse $4 \times (304+218)$ matrix (Yang et al., 2019, Beucler et al., 2019). These constructions are flexible and easy to add to existing models, but the law is only encouraged, not guaranteed.

Projection methods instead treat the conservation manifold as a geometric object and map predictions back to it. In PINN-Proj, the conserved quantity is momentum,

$$c(t) = \int_X u(x,t)\,dx \approx \sum_{x\in X} u(x,t)\,\Delta x,$$

and the projected state is

$\mathcal{H}$0

Integrating the projected state over the domain gives exactly $\mathcal{H}$1, so momentum is enforced by construction rather than by an added penalty (Baez et al., 2024). ConCerNet uses the same principle in phase space. If $\mathcal{H}$2 is the learned invariant and $\mathcal{H}$3, then the projected dynamics are

$\mathcal{H}$4

which guarantees

$\mathcal{H}$5

and therefore preserves the learned invariant along the rollout (Zhang et al., 2023).

By-construction parameterizations go further by restricting the model class itself. In PhyGAN for incompressible turbulence, the generator does not emit velocity directly; it emits a vector potential $\mathcal{H}$6, and a non-trainable physics layer computes

$\mathcal{H}$7

Since $\mathcal{H}$8, mass conservation is automatic (Tretiak et al., 2022). In the climate emulator, the architecture outputs only $\mathcal{H}$9 unconstrained variables and reconstructs the remaining $\mathcal{H}[\mathbf{u}] = 0$0 from the conservation equations so that

$\mathcal{H}[\mathbf{u}] = 0$1

holds to numerical precision (Beucler et al., 2019). In the continuous-time probability-setting of Neural Conservation Laws, the density and flux are parameterized so that

$\mathcal{H}[\mathbf{u}] = 0$2

holds automatically (Hua et al., 23 Jun 2025).

Mechanism	Representative formulation	Representative source
Soft penalty	$\mathcal{H}[\mathbf{u}] = 0$3	(Yang et al., 2019)
Projection	$\mathcal{H}[\mathbf{u}] = 0$4	(Baez et al., 2024)
Tangent-space projection	$\mathcal{H}[\mathbf{u}] = 0$5	(Zhang et al., 2023)
Hard architectural embedding	$\mathcal{H}[\mathbf{u}] = 0$6	(Tretiak et al., 2022)
Hard density-flux parameterization	$\mathcal{H}[\mathbf{u}] = 0$7	(Hua et al., 23 Jun 2025)

The recurring contrast is therefore not merely between “physics-informed” and “data-driven,” but between conservation as a penalty, conservation as a projection, and conservation as an invariant of the forward map.

3. Architectural realizations across model families

In turbulence generation, conservation is embedded through differential structure. PhyGAN introduces two hard-constraint variants: finite-difference embedding and spectral embedding. Both are backpropagation-compatible physics layers placed inside the generator. The finite-difference version computes partial derivatives with second-order central differences in the interior and one-sided second-order stencils at boundaries; the spectral version differentiates in Fourier space through multiplication by $\mathcal{H}[\mathbf{u}] = 0$8. The result is a DCGAN-like generator whose outputs are divergence-free by construction (Tretiak et al., 2022).

For autoregressive spatiotemporal evolution, the entropy-stable conservative flux form neural network (CFN) internalizes the update rule of a hyperbolic conservation law rather than directly predicting the next state. The target PDE is written as

$\mathcal{H}[\mathbf{u}] = 0$9

and the neural model learns the numerical flux inside a conservative update: $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$0 Because the update is in flux-difference form, interior fluxes cancel telescopically and only boundary fluxes contribute to the total change. The 2024 entropy-stable CFN incorporates the Kurganov–Tadmor scheme, minmod slope limiting, and TVDRK3 time stepping, and is explicitly designed to maintain conservation, non-oscillatory shock structure, and entropy stability over long rollouts (Liu et al., 2024).

A different architectural route starts from symmetry rather than flux balance. The neural Lagrangian integrator of Cranmer et al. represents the Lagrangian as

$$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$1

where the first layer $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$2 is a symmetry-enforcing layer that outputs invariant scalars. Noether’s theorem then yields exact conserved charges. For rotational symmetry in three dimensions, the inputs $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$3, $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$4, and $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$5 force the learned Lagrangian to be rotationally invariant, and angular momentum is exactly conserved by construction (Müller, 2022).

The climate-emulator literature provides a linear-constraint version of the same principle. The architecture-constrained network outputs $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$6 unconstrained values and computes the remaining $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$7 from enthalpy conservation, mass conservation, terrestrial radiation conservation, and solar radiation conservation. This construction is specialized to linear laws representable as a sparse constraint system, but it shows that architectural enforcement can be exact even when the network is otherwise a standard multilayer perceptron (Beucler et al., 2019).

These examples show that conservation-law-respecting generation is not a single architectural recipe. The invariant may be induced by a curl, a flux difference, a symmetry layer, or a constraint-recovery block. The common property is that the network’s output geometry is restricted before training is asked to discover anything.

4. Probability conservation, diffusion processes, and score-based generation

Probability conservation introduces a broader notion of a conservation-law-respecting generator. In the simulation-free framework of “Neural Conservation Laws,” the learned object is not only a drift field but a coupled pair $$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$8 satisfying

$$C_{\text{phys}}^\varepsilon = \mathbb{E}_{z\sim p_z(Z)} \left[\max\left(\mathcal{H}(G(Z)),\varepsilon^2\right)\right].$$9

The drift is then recovered by

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$0

so that the SDE

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$1

has marginals $$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$2. The method further enforces

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$3

as hard constraints through explicit autoregressive or factorized density parameterizations, thereby making maximum-likelihood training possible (Hua et al., 23 Jun 2025).

A technically significant issue in this construction is the spurious flux phenomenon. With the naive choice $$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$4, the flux

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$5

can remain nonzero as $$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$6, causing the implied drift

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$7

to blow up or behave badly. The remedy is a carefully designed divergence-free field $$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$8 that cancels the asymptotic flux while preserving the density. For factorized densities, the velocity simplifies to

$$\min_{\mathbf{G}}\max_{\mathbf{D}} V(\mathbf{D},\mathbf{G})$$9

which the paper notes is a gradient field (Hua et al., 23 Jun 2025).

Score-based generation offers a complementary route in which the prior is learned first and the physical law is enforced during inference. The physics-informed generative solver of 2026 trains a score model with

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$0

where $$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$1 penalizes residuals of the Score Fokker–Planck Equation. The reverse-time update is then modified by physical residual gradients. Writing the physical residual energy as

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$2

the PI-ISS sampler adds

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$3

to each reverse step, together with observation-consistency and patch-continuity terms (Zhu et al., 21 May 2026).

In acoustics, the enforced laws include the wave equation,

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$4

the momentum equation,

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$5

and the continuity equation,

$$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$6

This is still a generative model, but conservation enters at inference time as a guided projection toward the PDE manifold rather than as an architectural identity (Zhu et al., 21 May 2026).

These two lines of work broaden the notion of conservation beyond mass, energy, or momentum of a physical state variable. In probability-path formulations, the conserved object is the density itself; in score-based inverse problems, the law governs the admissibility of the sampled field.

5. Empirical record and characteristic trade-offs

The empirical literature consistently reports that exact or near-exact conservation is attainable and often improves predictive reliability. In PINN-Proj, the abstract states that the method substantially outperformed PINN in conserving momentum and lowered prediction error by three to four orders of magnitude from the best benchmark tested, while also performing marginally better on state prediction across the advection equation, viscous Burgers’ equation, and the Korteweg–De Vries equation (Baez et al., 2024). The detailed conservation errors reported in the paper are $$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$7, $$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$8, and $$V_C(\mathbf{D},\mathbf{G}) = V(\mathbf{D},\mathbf{G}) + \lambda C_{\text{phys}},$$9 for advection, Burgers, and KdV, respectively, compared with $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$0, $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$1, and $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$2 for the soft-constraint baseline (Baez et al., 2024).

For 3D turbulence, PhyGAN shows that hard incompressibility can be embedded with very small overheads. The finite-difference hard constraint reaches divergence variance at machine precision and is reported as over nine orders of magnitude improvement over Vanilla, with about $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$3 runtime increase per epoch. The spectral hard constraint reaches mean divergence about $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$4 with about $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$5 runtime increase per epoch, although the paper notes that it hurts the learning of some turbulence statistics, especially in the $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$6-$${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$7 diagnostics at inertial scales (Tretiak et al., 2022).

In climate emulation, architecture constraints enforce linear conservation laws to essentially exact precision. The architecture-constrained network NNA yields $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$8 on the in-distribution $${\cal L}(\alpha)=\alpha {\cal P}(x,y_{\mathrm{NN}}) + (1-\alpha)\mathrm{MSE}(y,y_{\mathrm{NN}}),$$9K validation set and ${\cal P}$0 on the out-of-distribution ${\cal P}$1K set, while all constrained networks generalize better than the unconstrained network under warming (Beucler et al., 2019).

ConCerNet reports analogous gains when the invariant is discovered rather than supplied. For the ideal spring-mass system, the baseline neural network has coordinate MSE ${\cal P}$2 and conservation violation ${\cal P}$3, whereas ConCerNet reports ${\cal P}$4 and ${\cal P}$5, respectively. For chemical kinetics, the learned invariant achieves ${\cal P}$6 against the exact conservation law (Zhang et al., 2023).

Setting	Conservation result	Additional observation
PINN-Proj on Advection/Burgers/KdV	${\cal P}$7	Marginally better state prediction across three PDE datasets (Baez et al., 2024)
PhyGAN for 3D turbulence	Mean divergence about ${\cal P}$8 (finite difference) and about ${\cal P}$9 (spectral)	Runtime increase about $4 \times (304+218)$0 and $4 \times (304+218)$1 per epoch (Tretiak et al., 2022)
Architecture-constrained climate emulator	$4 \times (304+218)$2 on $4 \times (304+218)$3K	Constrained networks generalize better under $4 \times (304+218)$4K warming (Beucler et al., 2019)
ConCerNet on ideal spring-mass	Conservation violation $4 \times (304+218)$5	Coordinate MSE reduced to $4 \times (304+218)$6 (Zhang et al., 2023)

The trade-offs are equally consistent. Projection adds computational cost and may introduce numerical variation due to repeated integration (Baez et al., 2024). Hard constraints can reduce model expressiveness, and the spectral hard constraint in turbulence, while best for divergence, can degrade some flow statistics (Tretiak et al., 2022). Very soft imprecise constraints may introduce fluctuations because the penalty is only intermittently active (Yang et al., 2019). In semiconductor applications, the literature further argues that evaluation should move away from FID-like metrics toward physics-fidelity benchmarks measuring the fraction of samples passing independent solvers, magnitude of constraint violation, and distance to known feasible or printable designs (Banad et al., 8 Jun 2026).

6. Invariant discovery, verification, and broader research directions

Not all relevant methods enforce a known conservation law directly. Some aim to discover the invariant first and then use it as a trustworthy modeling component. ConservNet learns a scalar $4 \times (304+218)$7 from grouped trajectories using the noise-variance loss

$4 \times (304+218)$8

thereby recovering hidden invariants from synthetic systems, Lotka–Volterra, the Kepler problem, and a real double pendulum trajectory. The paper explicitly states that ConservNet is not itself a generative model, but that the learned invariant could serve as a constraint, latent coordinate, or regularizer for downstream physics-aware generation (Ha et al., 2021).

NGCG takes the same discovery problem into a neural-symbolic regime designed to avoid false positives. Its latent objective is

$4 \times (304+218)$9

combined with a strict constancy gate $$c(t) = \int_X u(x,t)\,dx \approx \sum_{x\in X} u(x,t)\,\Delta x,$$0 and a diversity filter requiring $$c(t) = \int_X u(x,t)\,dx \approx \sum_{x\in X} u(x,t)\,\Delta x,$$1. On a benchmark of nine systems, the method reports DR = 1.00, FDR = 0.00, F1 = 1.00 on all four systems with true conservation laws and correctly outputs no law on all five systems without invariants (Ray, 20 Mar 2026). This verifies a key point in the broader literature: conservation-law-respecting generation depends not only on enforcing a candidate invariant, but on verifying that the invariant is genuine.

The 2026 Perspective situates these developments inside a wider integration agenda between generative models and physics-based simulators. It identifies four patterns: constrained generative process planning, physics-governed synthetic data generation, simulator-conditioned inverse design, and multimodal foundation models. It also identifies three research horizons: near-term physics-fidelity benchmarks, medium-term differentiable simulator infrastructure, and long-term multimodal foundation models pretrained jointly on text, layout/netlists, and simulation outputs (Banad et al., 8 Jun 2026).

A plausible synthesis is that conservation-law-respecting generative networks are becoming a unifying design principle rather than a niche architecture class. In one branch, the law is known and built into the generator by curl layers, flux form, symmetry, or explicit density-flux parameterization. In another, the law is learned from trajectories and then enforced by projection. In both branches, the governing idea is the same: a learned model for a physical system is most trustworthy when generation is restricted to the admissible manifold instead of being corrected only after the fact.