Viscosity-Regularized Losses

• Viscosity-regularized losses are training objectives derived from variational PDE formulations that incorporate dissipation terms to stabilize neural network training.
• They are applied in fluid simulations and neural signed distance functions to enforce adherence to physical laws and mitigate ill-posed optimization problems.
• Empirical benchmarks demonstrate that these regularizers improve stability and performance across heterogeneous viscosity regimes and complex geometric tasks.

Viscosity-regularized losses are a class of physically- and mathematically-inspired training objectives that enhance the stability, generalization, and physical fidelity of neural networks in scientific computing and geometry learning tasks. Originating in the variational formulations of PDE solvers, these losses incorporate dissipation-like or parabolic regularization—often directly derived from the underlying continuous operators—resulting in improved compliance with the governing equations, mitigating ill-posedness of certain objectives, and providing essential inductive bias for neural approximations of physical phenomena and geometric quantities. Canonical examples include variational losses encoding viscous dissipation in data-driven fluid solvers and viscosity-inspired terms stabilizing the Eikonal loss in neural signed distance functions.

1. Variational and PDE Motivation

The conceptual core of viscosity-regularized losses is the principle of minimizing an energy functional whose Euler–Lagrange equation recovers the desired PDE solution under appropriate conditions. In fluid simulation, the implicit-viscosity variational formulation seeks a velocity field $u$ (post-viscosity) minimizing

$$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$

where $\rho$ is density, $\mu$ dynamic viscosity, $\Delta t$ timestep, $u^s$ pre-viscosity velocity, and $\epsilon(u)$ the rate-of-strain tensor. The minimizer is the physically consistent solution to the viscous Stokes substep, and this variational form seeds the architecture of the associated regularizer in data-driven settings (Park et al., 2024).

In geometry learning, especially for neural SDFs, the Eikonal equation $\|\nabla u(x)\|_2 = 1$ defines the signed distance property, but its loss-driven minimization is known to be non-elliptic and ill-posed. Introducing a vanishing-viscosity regularization as $$\|\nabla u_\varepsilon(x)\|_2 = 1 + \varepsilon\Delta u_\varepsilon(x)$$ selects the unique viscosity solution by penalizing deviations from the desired gradient norm while adding Laplacian-based diffusion for stability and uniqueness (Krishnan et al., 1 Jul 2025).

2. Discrete and Neural Loss Formulations

In modern neural PDE solvers and implicit representation learning, the variational principles are discretized and incorporated as explicit loss terms.

For viscosity-regularized fluid solvers,

$$L_v = \underbrace{\frac{1}{n(n+1)}\sum_{i,j}\rho[u_{i-\frac{1}{2},j} - u^s_{i-\frac{1}{2},j}]^2 + \cdots}_{\text{momentum memory terms}} + \underbrace{\frac{1}{n^2 2\Delta t}\sum_{i,j}\mu\|\epsilon(u_{i,j})\|_F^2}_{\text{viscous dissipation}}$$

is computed per-batch on a symmetric MAC grid, ensuring consistent tensor shapes and facilitating automatic differentiation (Park et al., 2024).

For neural SDFs, the ViscoReg loss is

$$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$0

where both gradient and Laplacian are obtained via automatic differentiation. The viscosity parameter $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$1 is typically annealed to zero over training, facilitating a path to the correct solution while initially providing robust regularization (Krishnan et al., 1 Jul 2025).

3. Integration with Supervised and Physics-Inspired Objectives

Viscosity-regularized losses typically supplement supervised or conventional physics-driven objectives. In fluid simulation, the total objective is

$$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$2

where $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$3 is the mean squared error to ground-truth viscous velocity change, and $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$4 encodes the variational regularization. Empirically, $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$5 is essential when training over datasets with heterogeneous viscosity coefficients, ensuring that the network generalizes across different dissipation regimes and does not merely interpolate between training samples (Park et al., 2024).

In neural SDF learning, ViscoReg is included alongside on-surface ("manifold") and off-surface penalties,

$$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$6

with $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$7 coefficients tuned according to empirical validation. The viscosity component both stabilizes the dynamics of training and enforces correct geometric regularity (Krishnan et al., 1 Jul 2025).

4. Stability and Theoretical Analysis

A core motivation for viscosity-based regularization is stabilization of ill-posed or non-parabolic PDEs encountered in loss-driven training. In the context of neural SDF learning, the pure Eikonal loss leads to gradient flows governed by forward–backward parabolic PDEs, which are unstable and prone to amplifying high-frequency artifacts. The inclusion of a vanishing-viscosity term yields a fourth-order (genuinely diffusive) PDE for the training dynamics: $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$8 where the Laplacian term dissipates high-frequency error, ensuring robustness and convergence to the unique viscosity solution as $$E[u] = \int_\Omega \rho\|u-u^s\|^2\,dA + 2\Delta t \int_\Omega \mu\|\epsilon(u)\|_F^2\,dA$$9. This stabilizing effect is verified analytically and aligns with best practices in numerical PDEs (Krishnan et al., 1 Jul 2025).

5. Practical Implementation and Training Considerations

Modern implementations leverage automatic differentiation for efficient computation of gradients, Laplacians, and Frobenius norms. In fluid simulation, all velocity and derivative fields are embedded into a symmetric (2n+1)×(2n+1) MAC grid, so loss terms can be constructed with consistent tensor operations. In neural SDFs, SIREN-style MLPs are preferred to ensure second derivative continuity, and loss weights for each component are set according to grid searches over validation data. The viscosity coefficient ($\rho$0 in SDFs or $\rho$1 in fluids) can be constant or spatially varying, and is often annealed or randomized during training (Park et al., 2024, Krishnan et al., 1 Jul 2025).

Pseudocode for viscosity-regularized losses reflects standard supervised learning pipelines, with explicit computation of regularization terms via batch sampling, forward passes for neural prediction, loss evaluation via elementwise norms or derivatives, and update via stochastic optimization (Adam). For SDF learning, the computational overhead for viscosity terms is negligible relative to total training time (Krishnan et al., 1 Jul 2025).

6. Empirical Effects and Benchmarks

Viscosity-regularized losses have demonstrated significant improvements on benchmarks spanning physical simulation and geometric learning.

In data-driven fluid solvers, $\rho$2 is required to generalize across varying viscosity distributions: training solely with $\rho$3 on multi-viscosity data fails to capture correct dissipation, while the regularized objective matches unseen intermediate and mixed-viscosity scenarios (e.g., accurate velocity spread in scenes with both high and low $\rho$4 regions) (Park et al., 2024).

In neural SDFs, ViscoReg outperforms or matches prior state-of-the-art methods such as SIREN, DiGS, and StEik across 2D fractals, 3D shape benchmarks, and real indoor scene reconstructions. Key metrics include Chamfer distance, Hausdorff distance, and IoU. For instance, on the ShapeNet benchmark, ViscoReg (quadratic) achieves a squared-Chamfer of $\rho$5 and IoU of 0.959, outperforming competitors (Krishnan et al., 1 Jul 2025).

7. Limitations and Extensions

The main additional hyper-parameter introduced by viscosity-based losses is the viscosity coefficient ($\rho$6 or $\rho$7) and its scheduling. Linear decay or stepwise annealing are generally robust, though more adaptive or spatially-varying schemes remain open research directions. While current theory rigorously explains stabilization in the infinite-dimensional (continuum) limit, discrete optimization behavior in high-dimensional parameter spaces is not fully characterized. Although second-order derivatives are now computable in deep learning frameworks via autodiff, computational overhead remains a practical consideration in extremely large-scale models. Extensions include exploring finite-difference Laplacian approximations for efficiency and incorporating adaptive viscosity related to local loss or PDE residuals (Krishnan et al., 1 Jul 2025).