Exterior-Embedded Conservation Framework (ECF)

• ECF is a universal data-driven framework that enforces conservation laws by correcting the spectral zero mode in neural operator predictions for time-dependent PDEs.
• It integrates a Fourier-domain correction mechanism that replaces the predicted zero frequency with the true conserved quantity, ensuring exact conservation.
• ECF enhances model performance by reducing L2-error across diverse PDE scenarios while offering modular integration with various neural operator architectures.

to=arxiv_search.search 天天中彩票公司 大发游戏官网_json {"^^^^^^^^5query5 Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^^^^^", "^^^^^^^^5max_results5 5} to=arxiv_search.search аҧсынџьjson {"^^^^^^^^5query5 Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws^^^^^^^^5\5 "^^^^^^^^5max_results5 ^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5query5^^^^} The Exterior-Embedded Conservation Framework (ECF) is a universal conserving framework for data-driven neural operators that is designed to enforce conservation laws strictly in predictions for time-dependent partial differential equations (PDEs) governed by conservation laws such as conservation of mass, energy, or matter. It is introduced in "An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws" as a plug-and-play wrapper around a base neural operator, with conservation enforced by extracting the conserved quantity from the input state and reinserting it into the predicted state through a Fourier-domain correction (&&&^^^^5query5^^^^&&&). The framework is motivated by two observations: existing neural operators fail to satisfy conservation properties, which leads to degraded model performance and limited generalizability, and distinct PDE problems generally require different optimal neural network architectures, underscoring the limitations of specialized models in generalizing across diverse problem domains (&&&^^^^5query5^^^^&&&).

^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^. Mathematical setting and conservation target

The framework is formulated for time-dependent PDEs in conservation-law form on a spatial domain PRESERVED_PLACEHOLDER_^^^^5query5^^^^ over PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^: PRESERVED_PLACEHOLDER_^^^^5max_results5^^^^ with boundary conditions

PRESERVED_PLACEHOLDER_^^^^5query5^^^^

and initial data

PRESERVED_PLACEHOLDER_^^^^5\5^^^^

The associated global conserved quantity is

$Q(t)=\int_\Omega u(x,t)\,dx.$

Strict conservation, in the absence of boundary or source leakage, implies

$\frac{dQ}{dt}=0.$

Equivalently, one may write $\partial_t u=\mathcal F(u)$ with the requirement that $Q(u(t))\equiv \text{const}$ (&&&^^^^5query5^^^^&&&).

Within this setting, ECF does not replace the base neural operator’s approximation of the full state evolution. Instead, it isolates the conserved scalar content of the field and constrains the predicted state so that the global invariant matches the input state exactly. This suggests that the framework targets the invariant subspace associated with the zero Fourier mode rather than the full spectral content of the solution.

^^^^5max_results5^^^^. Exterior-embedding architecture

ECF wraps any base neural operator $\mathcal N$ and enforces PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5query5^^^^-conservation through a Fourier-domain correction. The architecture has three parts: a conserved-quantity encoder PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^, a base neural operator PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5max_results5^^^^, and a conserved-quantity decoder PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5query5^^^^ (&&&^^^^5query5^^^^&&&).

The conserved-quantity encoder PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5\5^^^^ takes as input the current state PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws55^^^^, sampled on an PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^6-dimensional grid over PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^7, and applies a truncated discrete Fourier transform: PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^8 Its output is the zero mode

PRESERVED_PLACEHOLDER_^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^9

The base neural operator PRESERVED_PLACEHOLDER_^^^^5max_results5query5^^^^ maps the current state to a predicted next state,

PRESERVED_PLACEHOLDER_^^^^5max_results5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^

The conserved-quantity decoder PRESERVED_PLACEHOLDER_^^^^5max_results5max_results5^^^^ computes the Fourier spectrum of the prediction,

PRESERVED_PLACEHOLDER_^^^^5max_results5query5^^^^

replaces the predicted zero mode with the true one,

PRESERVED_PLACEHOLDER_^^^^5max_results5\5^^^^

and reconstructs the corrected field by inverse transform: PRESERVED_PLACEHOLDER_^^^^5max_results55^^^^

A concise way to characterize the mechanism is that ECF preserves all nonzero Fourier modes generated by the base operator while overwriting the zero mode with the conserved quantity extracted from the input. Because the global integral over PRESERVED_PLACEHOLDER_^^^^5max_results5^^^^6 is represented entirely by the zero mode, the correction directly targets the conservation defect.

^^^^5query5^^^^. Theoretical guarantees

The paper states two central theoretical results for the corrected prediction PRESERVED_PLACEHOLDER_^^^^5max_results5^^^^7 (&&&^^^^5query5^^^^&&&).

Theorem ^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^ (Exact Conservation) establishes that

PRESERVED_PLACEHOLDER_^^^^5max_results5^^^^8

so

PRESERVED_PLACEHOLDER_^^^^5max_results5^^^^9

The proof sketch is spectral: all nonzero Fourier modes integrate to zero over PRESERVED_PLACEHOLDER_^^^^5query5query5^^^^, and only the zero mode contributes to the integral.

Theorem ^^^^5max_results5^^^^ (Error Reduction) states that if PRESERVED_PLACEHOLDER_^^^^5query5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^ is the uncorrected output and PRESERVED_PLACEHOLDER_^^^^5query5max_results5^^^^ is the true next state, then

PRESERVED_PLACEHOLDER_^^^^5query5query5^^^^

Replacement of the erroneous zero mode can only reduce, or leave unchanged, the PRESERVED_PLACEHOLDER_^^^^5query5\5^^^^-error.

These results give ECF a specific theoretical profile. Exact conservation is algebraic rather than asymptotic: it is enforced at the level of each corrected prediction. The error-reduction result is similarly mode-specific: correcting the conserved quantity cannot worsen the PRESERVED_PLACEHOLDER_^^^^5query55^^^^ discrepancy because the correction removes error in the zero mode while leaving the remaining modes unchanged. The paper further states that, since the architecture enforces conservation laws, it theoretically proves that it enhances model performance (&&&^^^^5query5^^^^&&&).

^^^^5\5^^^^. Integration with neural operators and training paradigms

A central design choice is modularity. ECF is described as a plug-and-play wrapper around any operator PRESERVED_PLACEHOLDER_^^^^5query5^^^^6, including FNO, UNO, CNO, Transolver, and U-Net, without altering the internal architecture of the base model (&&&^^^^5query5^^^^&&&). This directly addresses the observation that different PDE problems generally require different optimal neural network architectures.

Two training paradigms are defined.

In the Integrated mode, denoted PRESERVED_PLACEHOLDER_^^^^5query5^^^^7, training is end-to-end with the corrected prediction inside the loss: PRESERVED_PLACEHOLDER_^^^^5query5^^^^8

In the Staged mode, denoted PRESERVED_PLACEHOLDER_^^^^5query5^^^^9, the base operator is first trained alone with

PRESERVED_PLACEHOLDER_^^^^5\5query5^^^^

then frozen, with the correction applied only at test time (&&&^^^^5query5^^^^&&&).

The reported optimization setup uses AdamW with learning rate PRESERVED_PLACEHOLDER_^^^^5\5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^ and default weight decay, batch size PRESERVED_PLACEHOLDER_^^^^5\5max_results5^^^^, PRESERVED_PLACEHOLDER_^^^^5\5query5^^^^ epochs, and results averaged over PRESERVED_PLACEHOLDER_^^^^5\5\5^^^^ runs. Training uses mean-absolute-error (MAE), or MSE if preferred, while evaluation is by RMSE. In the integrated mode the correction is differentiable; in the staged mode it is purely post-hoc (&&&^^^^5query5^^^^&&&).

The two modes imply different deployment regimes. PRESERVED_PLACEHOLDER_^^^^5\55^^^^ couples the conservation mechanism to representation learning, whereas PRESERVED_PLACEHOLDER_^^^^5\5^^^^6 preserves the original training pipeline and adds conservation only at inference. A plausible implication is that the staged variant is useful when retraining a pretrained neural operator is undesirable.

5. Experimental validation

The framework is evaluated on multiple conservation-law-constrained PDE scenarios on PRESERVED_PLACEHOLDER_^^^^5\5^^^^7, using grid sizes from PRESERVED_PLACEHOLDER_^^^^5\5^^^^8 to PRESERVED_PLACEHOLDER_^^^^5\5^^^^9, periodic or Neumann boundary conditions, and $Q(t)=\int_\Omega u(x,t)\,dx.$^^^^5query5^^^^ output snapshots (&&&^^^^5query5^^^^&&&).

The benchmark set comprises the following cases:

Scenario	Conserved quantity
AC-DW, AC-FH	$Q(t)=\int_\Omega u(x,t)\,dx.$^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^
Heat (adiabatic diffusion)	$Q(t)=\int_\Omega u(x,t)\,dx.$^^^^5max_results5^^^^
Water (shallow water Eqn)	$Q(t)=\int_\Omega u(x,t)\,dx.$^^^^5query5^^^^
Diff (pure diffusion)	$Q(t)=\int_\Omega u(x,t)\,dx.$^^^^5\5^^^^
CD (convection–diffusion)	$Q(t)=\int_\Omega u(x,t)\,dx.$5

The paper reports representative RMSE improvements in Table ^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^. For UNO on AC-DW, RMSE decreases from $Q(t)=\int_\Omega u(x,t)\,dx.$6 to $Q(t)=\int_\Omega u(x,t)\,dx.$7, corresponding to $Q(t)=\int_\Omega u(x,t)\,dx.$8. For FNO on Diff, RMSE decreases from $Q(t)=\int_\Omega u(x,t)\,dx.$9 to $\frac{dQ}{dt}=0.$^^^^5query5^^^^, corresponding to $\frac{dQ}{dt}=0.$^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^ (&&&^^^^5query5^^^^&&&).

The conservation error results are more categorical. Table ^^^^5query5^^^^ reports that conservation error drops from $\frac{dQ}{dt}=0.$^^^^5max_results5^^^^ to machine-precision $\frac{dQ}{dt}=0.$^^^^5query5^^^^ in every case. The provided examples are: $\frac{dQ}{dt}=0.$^^^^5\5^^^^

$\frac{dQ}{dt}=0.$5

The computational overhead is reported as $\frac{dQ}{dt}=0.$6 for $\frac{dQ}{dt}=0.$7, while $\frac{dQ}{dt}=0.$8 adds no training cost (&&&^^^^5query5^^^^&&&).

Taken together, these results position ECF as a method that changes the conservation profile of neural-operator predictions from approximate to strict, while also improving predictive accuracy on the tested benchmarks. The experiments are consistent with the theoretical statement that replacing an erroneous zero mode should not increase $\frac{dQ}{dt}=0.$9-error.

6. Scope, limitations, and relation to structure-preserving numerics

The framework is presented as universally enforcing exact conservation with provable error reduction and minimal overhead. Within the two training paradigms, $\partial_t u=\mathcal F(u)$^^^^5query5^^^^ typically yields the best accuracy, while $\partial_t u=\mathcal F(u)$^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^ offers a modular, zero-retraining option (&&&^^^^5query5^^^^&&&).

Its principal stated limitation is geometric and spectral: the current Fourier-based correction assumes periodic or simple-geometry boundary conditions, and extending the approach to complex geometries or nonuniform meshes is left for future work (&&&^^^^5query5^^^^&&&). This suggests that the present formulation is most natural when the conserved quantity can be robustly identified with a Fourier zero mode on a regular grid.

Within the broader literature on conservation-preserving computation, the ^^^^5max_results5query5max_results55^^^^ neural-operator ECF occupies a distinct methodological position. Finite element exterior calculus for the rotating shallow-water equations uses exact subcomplexes and discrete differential forms to obtain mass-consistent potential vorticity together with mass and energy conservation (&&&^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws55^^^^&&&). Dual-field FEEC discretization of port-Hamiltonian systems preserves energy balance and, for Maxwell, keeps the magnetic and electric fields divergence free at the discrete level (&&&^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^6&&&). FEEC combined with symplectic Runge–Kutta methods yields a local multisymplectic conservation law for time-dependent Hamiltonian PDEs (&&&^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^7&&&). Discrete exterior calculus for incompressible Euler and Navier–Stokes on prismatic Delaunay–Voronoi meshes enforces exact discrete mass, energy, and circulation conservation and leads to a selection principle excluding dissipative Euler limits (&&&^^^^5(Dong et al., 20 Nov 2025) Exterior-Embedded Conservation Framework An Exterior-Embedding Neural Operator Framework for Preserving Conservation Laws5^^^^8&&&). By contrast, ECF for neural operators does not build conservation into the discretization of the PDE itself; it enforces conservation at the prediction level by correcting the spectral zero mode of a learned operator (&&&^^^^5query5^^^^&&&).

This distinction is consequential. FEEC- and DEC-based methods derive conservation from exact algebraic or variational structure in the discretized equations, whereas ECF derives exact conservation from a wrapper architecture that can be integrated with learned operators without altering their internal design. A plausible implication is that ECF is aimed less at replacing structure-preserving solvers than at making data-driven neural operators compatible with conservation-law constraints across heterogeneous PDE domains.