Normalized Conservation Law Violations

• Normalized conservation law violations are dimensionless metrics that quantify the accuracy of numerical and operator-learning methods in preserving invariants like mass, momentum, and energy.
• They are computed by scaling the difference between discrete and initial conserved quantities, enabling robust comparisons across classical discretizations and modern neural approaches.
• Research indicates that strategies such as adaptive correction and energy-preserving schemes can reduce violation metrics to machine precision, enhancing simulation fidelity.

Normalized conservation law violations quantify the extent to which numerical, data-driven, or operator-learning schemes fail to preserve fundamental invariants (such as mass, momentum, energy) mandated by the governing partial differential equations. These violations are typically assessed relative to appropriate scaling quantities—most commonly the initial value of the conserved quantity—yielding dimensionless error metrics that enable robust algorithmic comparisons, highlight physical fidelity, and guide model or method selection. Research on normalized conservation law violations systematically characterizes such errors across classical discretizations, Galerkin and finite-volume schemes, and modern operator-learning architectures, providing well-established frameworks for diagnosis and remediation.

1. Formal Definition and Normalization Schemes

Let $C(t)$ denote a continuous conserved quantity (e.g., mass, energy) and $C_h(t)$ its discrete or learned analog at time $t$. The absolute violation is defined as

$\Delta C(t) = C_h(t) - C_h(0).$

To enable consistent comparisons and to neutralize units, normalized (or relative) conservation law violations are formed by scaling the absolute error by the initial value: $$\Delta C_{\rm norm}(t) = \frac{\Delta C(t)}{|C_h(0)|}.$$ This convention is dominant in the numerical analysis and machine learning for PDEs literature (Charnyi et al., 2016, Liu et al., 30 May 2025).

For vector-valued and higher-moment invariants, such as kinetic energy or $L^2$-norm, analogous normalization is employed:

• Linear invariants (e.g., mass, momentum):

$$\varepsilon_{\rm lin} = \frac{|M(U) - m_0|}{|m_0|}, \quad M(U) = \sum_{i=1}^N U_i \Delta x, \quad m_0 = \sum_{i=1}^N U_{gt,i} \Delta x$$

• Quadratic invariants (e.g., energy, $L^2$-norm):

$$\varepsilon_{\rm quad} = \frac{|N(U) - c_0|}{|c_0|}, \quad N(U) = \sum_{i=1}^N U_i^2 \Delta x, \quad c_0 = \sum_{i=1}^N U_{gt,i}^2 \Delta x$$

where $U$ is the numerical or predicted solution and $C_h(t)$0 is the ground truth (Liu et al., 30 May 2025).

Alternative normalization schemes, such as scaling by domain volume $C_h(t)$1 or mean density, may be additionally used to facilitate cross-system comparison (Charnyi et al., 2016). No evidence supports normalization strategies involving time or viscosity scaling for conservation law metrics.

2. Conservation Law Violations in Discretizations

Discretizations for PDEs—finite element, finite volume, spectral, and operator-learning methods—vary greatly in their conservation properties. For incompressible Navier–Stokes, mixed finite element discretizations that weakly enforce divergence-free velocity lead to violations in energy, momentum, and helicity conservation (Charnyi et al., 2016). The canonical balance laws and their normalized violation forms include:

• Energy: $C_h(t)$2,
• Linear momentum: $C_h(t)$3,
• Angular momentum, helicity, enstrophy, and total vorticity: analogous expressions.

For example, in the Gresho vortex test, standard convective or conservative forms display large $C_h(t)$4 or $C_h(t)$5 over time, while structure-preserving schemes such as EMAC drastically reduce normalized violations to machine precision. The size and drift of normalized violations serve as direct diagnostics of spurious dissipation or generation of invariants and provide a quantitative basis for method selection (Charnyi et al., 2016).

3. Conservation Law Violations in Iterative and Operator Learning Methods

Iterative solvers and operator learning approaches introduce distinct patterns of normalized conservation law violations. Explicit Runge–Kutta pseudo-time iterations, when used to accelerate implicit finite-volume updates, result in a modified conservation law with a flux rescaled by a method-dependent factor $C_h(t)$6: $C_h(t)$7 with $C_h(t)$8 the Runge–Kutta stability function and $C_h(t)$9. Normalized conservation violations are thus controlled by how closely $t$0 approaches unity; if not corrected, significant drift accumulates (Birken et al., 2021).

A "zero-root" strategy—explicitly selecting step sizes so that $t$1—restores $t$2 and eliminates the systematic source of normalized violation, driving the corresponding normalized residuals to zero. Empirical comparisons (1D Burgers, 2D Euler vortex) confirm an order of magnitude improvement in normalized residual metrics for such renormalized schemes (Birken et al., 2021).

4. Normalized Violation Metrics in Operator Learning

Neural operator approaches—including Fourier Neural Operators (FNOs), DeepONets, and related frameworks—require new mechanisms to monitor and enforce conservation. Standard FNOs exhibit substantial normalized conservation law violations in both mass ($t$3) and norm ($t$4). Several approaches have been proposed to reduce these violations via:

• Loss function penalization,
• Hard projection of network outputs,
• Adaptive correction frameworks.

An effective strategy constructs a learnable correction matrix $t$5 satisfying

$t$6

guaranteeing that the post-corrected solution strictly inherits the conservation law ($t$7). Empirically, adaptive correction reduces normalized violation metrics to machine precision while at least preserving predictive accuracy relative to unconstrained or projection-based methods (Liu et al., 30 May 2025).

Representative normalized violation results for FNO-based settings (Liu et al., 30 May 2025):

Method	Normalized Mass Violation $t$8 (%)	Normalized Norm Violation $t$9 (%)
FNO (raw)	$\Delta C(t) = C_h(t) - C_h(0).$0	$\Delta C(t) = C_h(t) - C_h(0).$1
Projection	$\Delta C(t) = C_h(t) - C_h(0).$2	$\Delta C(t) = C_h(t) - C_h(0).$3
Adaptive	$\Delta C(t) = C_h(t) - C_h(0).$4	$\Delta C(t) = C_h(t) - C_h(0).$5

5. Diagnosis and Best Practices

Best practice dictates the tabulation or plotting of normalized conservation law violations for any method under study:

• For each invariant $\Delta C(t) = C_h(t) - C_h(0).$6, present $\Delta C(t) = C_h(t) - C_h(0).$7 as a function of time or training epoch,
• Benchmark across mesh resolutions, network architectures, or PDE regimes,
• Prefer methods or post-processing strategies yielding violations matching machine precision,
• For operator learning, adaptive correction modules offer a modular, extensible mechanism that is model-architecture-agnostic, does not degrade loss minima, and directly drive all normalized error metrics to zero (Liu et al., 30 May 2025, Charnyi et al., 2016).

For classical discretizations, energy-preserving formulations such as EMAC or use of divergence-free function spaces are recommended when long-term fidelity of invariants is required. For iterative solvers, enforcing the zero-root strategy on step sizes robustly corrects the main structural error affecting normalized law violations (Birken et al., 2021).

6. Broader Implications and Extensions

Normalized conservation law violations serve as model-agnostic, hyperparameter-free indicators of physical fidelity. They are critical in the assessment of:

• Stability and long-term integration performance,
• Uncertainty quantification in stochastic or learned models,
• Comparison of operator-learning and classical solvers,
• Applicability to systems-of-interest in turbulence, wave dynamics, and control.

Recent advances enable the adaptive correction and enforcement of multiple, possibly nonlinear, conservation laws in neural architectures via a learnable correction map, and provide a foundation for future extensions to higher-order or coupled invariants (Liu et al., 30 May 2025). The normalized conservation law violation metrics thus remain central diagnostics and methodological design criteria in contemporary computational physical sciences.