betaFeedback

WorksComputer Science

Hamiltonian Two-Way Coupling of Nonlinear Waves and 3D Flows

Abstract

Simulating large-scale free-surface water by coupling a localized 3D fluid solver to a cheaper 2D surface model has long faced a mismatch in wave dynamics: efficient 2D wave models used in graphics are typically either linear or non-dispersive. These models are fast, simple, and accurate for calm, small-amplitude seas, but coupling them with strongly nonlinear 3D solvers produces visible reflections and artifacts at the 2D–3D interface. We address this problem by introducing a nonlinear and dispersive 2D wave model based on the canonical Zakharov formulation. Its Hamiltonian structure, in which the surface elevation and surface potential form a canonical pair (ηη, ψψ) governed by the wave energy, enables a canonically consistent two-way coupling scheme, allowing information to pass smoothly across the 2D–3D interface. Our 2D solver reduces mean wave-height error by 1.7–5×\times over SWE, BEM, and Airy baselines while running more than 103×10^3\times faster than BEM; it achieves greater nonlinear accuracy and coupling fidelity than SWE and Airy, with minor losses in speed and stability. Coupling it with a 3D Navier–Stokes solver yields a full system that suppresses visible seam artifacts across a range of experiments, including dispersion-matching and Kelvin-wake tests, and runs over 4×\times faster than a pure GPU NB-FLIP simulation on the same domain.

A seaplane landing on the ocean surface, shown in near view and far view, with near and far views of a battleship, two boats, and a submarine

Figure 1. A seaplane landing on the ocean surface, shown in near view (top-left) and far view (top-right). Near views of a battleship navigating through heavy waves (middle-left), two boats chasing each other (middle-center), and a submarine rising to the ocean surface (middle-right), with their corresponding far views shown in the bottom row.

Authors’ Contact Information: Sinan Wang, [email protected], Georgia Institute of Technology, USA; Ruicheng Wang, [email protected], Georgia Institute of Technology, USA; Taiyuan Zhang, [email protected], Dartmouth College, USA; Fan Feng, [email protected], Dartmouth College, USA; Jinjin He, [email protected], Georgia Institute of Technology, USA; Yuchen Sun, [email protected], Georgia Institute of Technology, USA; Zhiqi Li, [email protected], Georgia Institute of Technology, USA; Bo Zhu, [email protected], Georgia Institute of Technology, USA.

This work is licensed under a Creative Commons Attribution 4.0 International License.

© 2026 Copyright held by the owner/author(s).

ACM 1557-7368/2026/12-ART179

https://doi.org/10.1145/3842540

CCS Concepts: • Computing methodologies → Physical simulation.

Additional Key Words and Phrases: Fluid simulation, wave simulation

ACM Reference Format:

Sinan Wang, Ruicheng Wang, Taiyuan Zhang, Fan Feng, Jinjin He, Yuchen Sun, Zhiqi Li, and Bo Zhu. 2026. Hamiltonian Two-Way Coupling of Nonlinear Waves and 3D Flows. ACM Trans. Graph. 45, 6, Article 179 (December 2026), 20 pages. https://doi.org/10.1145/3842540

Introduction

Simulating large-scale free-surface flow remains a long-standing computational bottleneck in graphics, whether the scene is a seaplane splashing down on the open ocean, a battleship plowing through heavy seas, or a submarine breaking the ocean surface. A fully three-dimensional incompressible Navier–Stokes (NS) simulation captures every regime of interest, including splashes, breaking waves, and fluid–solid coupling, but its memory and compute costs grow prohibitive once the domain spans hundreds of meters. A practical remedy is to confine the expensive 3D solver to a small embedding region and model the surrounding open ocean with a much cheaper 2D surface theory [15, 31, 59, 66, 72] in order to propagate far-field waves efficiently. The success of this 2D–3D decomposition rests on the quality of the 2D model and the fidelity of the coupling between the two solvers.

Within this 2D–3D decomposition, a key obstacle to practical, high-fidelity coupling is that graphics currently lacks an efficient 2D wave model that is simultaneously nonlinear and dispersive: none retains the wave–wave interactions whose strength grows with wave steepness (nonlinearity) while allowing waves of different wavelengths to travel at their correct, wavelength-dependent speeds (dispersion). The two wave theories that dominate the literature both bypass the nonlinear surface dynamics. Airy wave theory [1, 10, 58, 70] linearizes the free-surface boundary conditions around the flat state, retaining the correct dispersion ω2=gktanh⁡(kh)\omega^2 = gk \tanh(kh) but discarding every nonlinear interaction that gives rise to scenarios like the steepening of Stokes wavetrains. The discarded terms scale with wave steepness, so for the gentle ambient swell that Airy solvers typically animate, the linearization costs little and buys great speed and simplicity. The shallow-water equations (SWE) [14, 37, 38] carry nonlinear depth-averaged advection but collapse the dispersion relation to ω=gh k\omega= \sqrt{gh}\,k, removing the very dispersive structure that distinguishes a real ocean from a sloshing tank, a sacrifice that is acceptable in the transport-dominated shallow flows SWE was built for, but not in open water. Lagrangian alternatives that have gained traction in graphics, e.g., wave particles [81], wave packets [35], wavefront tracking [34], and water-surface wavelets [33], all transport their primitives along linear ray paths and inherit the same linearization. Each of these trade-offs is defensible in its intended regime. Once strongly nonlinear waves dominate, however, these models lose accuracy, whereas our method makes the opposite trade-off, sacrificing a little speed and stability in exchange for greater nonlinear accuracy and coupling fidelity.

Coupling a localized 3D NS solver to such a linear 2D wave model creates precisely such a strongly nonlinear setting. The 3D solver is strongly nonlinear by construction, so waves reaching the embedding-box boundary exhibit nonlinear effects such as crest steepening and amplitude-dependent dispersion. A linear 2D model cannot represent any of these, so the two solvers disagree on the very content of the surface state at the interface. The disagreement shows up as spurious reflections at the seam and visible surface artifacts. We therefore identify the absence of a nonlinear and dispersive 2D wave model in graphics as a key obstacle to a 2D–3D coupling pipeline that is at once practical, accurate, and largely free of interface artifacts. The one prior nonlinear route, the boundary-element method (BEM) [31], recovers the nonlinear free-surface potential flow at the PDE level, but discretizes it on a Lagrangian surface mesh, and pays for its nonlinear accuracy with per-step surface tracking/remeshing, dense O(N3)O(N^3) boundary interactions, and high implementation complexity.

We address this limitation by introducing a 2D Eulerian wave model that is at once nonlinear and dispersive, and we show that its underlying Hamiltonian structure is what makes the two-way coupling to a 3D NS solver natural. The wave model evolves the canonical Zakharov pair [82]: the surface elevation η\eta and the surface value ψ\psi of the velocity potential. Here Hamiltonian means that (η,ψ)(\eta,\psi) form a canonical pair in the sense of classical mechanics: they evolve like position and momentum under the total wave energy, the Hamiltonian (Section 3). We advance the Zakharov system using the Craig–Sulem expansion of the Dirichlet–Neumann operator (DNO) G(η)G(\eta) [16]. This reduces every nonlinear correction to FFTs and pointwise products on a regular grid at O(N2log⁡N)O(N^2 \log N) cost. In practice, we retain this expansion through second or third order (Section 3). The same DNO machinery enables two-way coupling: it lifts (η,ψ)(\eta,\psi) to depth-resolved boundary velocities for the 3D solver and, in reverse, reconstitutes a canonical wave state after relaxing η\eta toward the 3D surface.

On standalone 2D wave-propagation benchmarks, our wave model alone lowers the mean wave-height error by 1.7× (deep water) to 2.4× (shallow water) over the strongest linear baseline (Airy with dispersion kernels) and by 3 to 5× over SWE and BEM, while running more than 10310^3× faster than BEM; the price of this accuracy is a modestly more expensive and slightly less stable wave step than the linear baselines, a few milliseconds per substep in absolute terms (Table 3). Once coupled to the 3D NS solver, a dispersion-matching test in which a wave packet radiates from the 3D box across the seam and a single-boat experiment in which a hull radiates a Kelvin wake through the seam together show that our coupled pipeline yields symmetric wave fields with minimal seam artifacts, whereas prior coupling schemes exhibit varying degrees of asymmetric propagation, spurious reflection, or visible 2D–3D inconsistency (Section 8); on the single-boat scene, our solver also runs more than 4× faster than a GPU NB-FLIP simulation of the same domain, i.e., the bounding box of the entire visible water region. We further run a suite of large-scale scenarios, a battleship in heavy seas, a surfacing submarine, two chasing boats, and a seaplane landing on the open ocean, in which the 2D–3D interface remains largely free of visible artifacts.

Contributions. Our main contributions are:

  1. A nonlinear and dispersive Eulerian wave model for graphics. We evolve the canonical Zakharov pair (η,ψ)(\eta,\psi) under the exact water-wave Hamiltonian on a structured grid at FFT cost, capturing nonlinear surface phenomena such as the steepening of Stokes wavetrains. This retains the nonlinear modeling capability of the prior BEM approach while reducing computational complexity from O(N3)O(N^3) to O(N2log⁡N)O(N^2 \log N) without requiring per-step surface tracking or remeshing.

  1. A canonically consistent two-way coupling between (η,ψ)(\eta,\psi) and a 3D NS solver. Both directions build on the DNO expansion: the forward direction harmonically extends (η,ψ)(\eta,\psi) to depth-resolved bulk velocities on the embedding-box faces, while the

reverse direction relaxes the wave model’s η\eta toward the 3D free-surface elevation and reconstitutes ψ\psi from the relaxed η\eta and the wave’s own surface velocity through an inverse DNO solve, keeping the wave state canonical at every substep.

(3) A continuously tunable nonlinear wave model. A scalar perturbation parameter ε∈[0,1]\varepsilon\in[0,1] modulates every nonlinear correction in the truncated DNO and the Bernoulli equation for stability, parametrizing a continuous family of solvers that degenerates to linear Airy theory at ε=0\varepsilon= 0 and recovers the nonlinear (truncated) Zakharov system at ε=1\varepsilon= 1.

2D simulation

SWE. The SWE, a simplified form of the Navier–Stokes equations, originated in engineering physics [17, 39, 65] and were first introduced to computer graphics by [37]. [38] introduced an implicit semi-Lagrangian scheme to improve stability and support larger time steps. [14] extended SWE to handle arbitrary slopes and depths. Later works enhanced SWE by adding surface tension [74], reducing numerical dissipation while ensuring stability via implicit Newmark integration [3], and improving efficiency using a collocated grid and a simplified SWE formulation [40].

Airy Wave Theory. Airy wave theory [1, 6, ?], popularized in computer graphics by [69], synthesizes ocean height fields by evolving wave spectra according to dispersion relations. Some implementations augment linearly evolved waves with geometric displacement or post-sharpening [66, 69]; we distinguish this geometric nonlinearity from the dynamical nonlinearity in our wave evolution. [29] further improved spectral realism using an advanced empirical model, and proposed an artist-friendly parameterization scheme to facilitate intuitive control over ocean wave elongation. Beyond Tessendorf’s formulation, [64] proposed a compact FFT-based fluid solver widely used in 2D surface simulations. [53] fit artist-created wave fields to spectral parameters for consistent high-frequency synthesis. Later, [70] presented the eWave model, which provides a time integration in both Fourier and convolution forms, enabling more flexible boundary treatments. Dispersion-kernel approaches [10] handle solid boundaries via convolution and shadow masks. More recently, [58] introduced a fundamental solution-based method using Green’s functions to capture wave–obstacle interactions in open domains without grids.

High-Order Spectral Method. Starting from the canonical Hamiltonian formulation of [82], the high-order spectral (HOS) method was independently introduced by [20] and [79] to evolve nonlinear surface waves on a flat reference grid via a Taylor expansion of the Dirichlet–Neumann operator, with [16] casting the expansion in the explicit operator form used in modern HOS solvers. Such interior-to-boundary reductions via Dirichlet–Neumann (Steklov–Poincaré) operators also appear in graphics, including extrinsic shape analysis [78], boundary-first conformal flattening [57], and boundary-only physics-based skinning [25]. Subsequent work extended HOS to directional sea states [4], variable bathymetry [26], and large-scale open-ocean simulation [21]. We adopt the HOS expansion as the engine of our 2D wave solver.

Other Wave Simulation. A variety of approaches have explored alternative physical representations to better capture dispersion, boundary interactions, and visual detail. These include wave particles [81], wave packets [35], wavefront tracking [34], and water surface wavelets [33]. Hybrid models combine different regimes, such as SWE with Airy theory [36], or augment height fields with particles for breaking waves [71]. Others enhance detail via a post-process, like wave curves on deforming surfaces [61].

3D simulation

Early works in 3D fluid simulation for graphics [23, 24, 63] primarily adopt Eulerian frameworks, but often suffer from numerical dissipation. Pure Lagrangian approaches, such as smoothed particle hydrodynamics (SPH) [32, 48], require costly neighbor searches and struggle to enforce incompressibility. Hybrid Lagrangian-Eulerian methods such as Particle-in-Cell (PIC) [27] alleviate some issues but still exhibit dissipation. To address this, the Fluid-Implicit Particle (FLIP) method was introduced to graphics by [85], building on [8], significantly reducing dissipation. To improve efficiency, Narrow Band FLIP (NB-FLIP) [22] and extended NB-FLIP [56] were later proposed. Furthermore, several approaches [2, 45, 52, 84] have employed adaptive techniques to resolve fine-scale fluid features while maintaining computational efficiency for large-scale scenes. Recently, flow-map-based methods [11, 12, 13, 19, 28, 41, 42, 43, 50, 49, 67, 68, 73, 76, 77, 83] and the Coadjoint Orbit FLIP [49] have reduced numerical dissipation and better preserved vortices.

Coupled 2D and 3D simulation

[72] coupled shallow water with a 3D solver using the Lattice Boltzmann method. [51] improved art-directability by introducing guide shapes to constrain high-resolution liquid simulations to low-resolution inputs. [15] combined SWE, SPH, and a 3D Eulerian solver, enabling efficient simulation but lacking accurate wave dispersion. [31] addressed this by coupling BEM [18, 30] with FLIP for more realistic ocean–ship interaction, though BEM’s mesh operations are computationally expensive and limit scalability. The fluxed animated boundary (FAB) method [66] allows artistic control but only supports one-way coupling (the 3D simulation cannot affect the “GPU-native” means that the 2D solver advances a height field stored on a uniform structured grid, making it well suited for GPU implementation; “Mesh-op-free” means the 2D solver requires no per-step mesh operations on an unstructured surface mesh (mesh construction, tracking, or remeshing) during simulation, in contrast to BEM-style methods that maintain a Lagrangian surface mesh; “Warmup-free” means the 2D solver is usable from t=0t = 0 without first accumulating a temporal sampling window; “Nonlinear” means nonlinear wave dynamics rather than geometric sharpening; “Dispersive” means the 2D wave model reproduces the correct deep-water dispersion relation, so waves of different wavelengths propagate at different speeds; “Hamiltonian” means the 2D solver evolves the canonical Zakharov pair (η,ψ)(\eta,\psi) under the water-wave Hamiltonian (Section 3).

2D simulation). Perfectly Matched Layers (PML)-based methods [5, 7, 62] reduce reflection at the interface via non-reflecting boundaries and allow localized re-simulation, but are also one-way. More recently, Schreck and Wojtan [59] coupled a 2D fundamental solution solver [58] with a 3D FLIP simulation, using dispersion matching to reduce coupling artifacts. Nonetheless, its underlying wave model remains linear Airy theory and therefore does not capture nonlinear wave effects. Table 1 summarizes how our method compares to representative prior coupled 2D–3D water solvers. To our knowledge, ours is the first to combine nonlinear and dispersive surface-wave dynamics with two-way coupling and ocean-scale efficiency in a single framework.

MethodHamiltonianTwo-wayNonlinearGPU-nativeMesh-op-freeWarmup-freeDispersive
Thürey et al. [72]✗✓✓✓✓✓✗
FAB [66]✗✗✗✓✓✓✓
Chentanez et al. [15]✗✓✓✓✓✓✗
Schreck and Wojtan [59]✗✓✗✓✗✗✓
Huang et al. [31]✗✓✓✗✗✓✓
Ours✓✓✓✓✓✓✓

Table 1. Comparison with prior 2D–3D coupled water-simulation methods.

Hamiltonian Spectral Wave Solver

We model the open-ocean portion of the domain as the free surface of an incompressible, irrotational, inviscid fluid of finite depth −d≤y≤η(x,z,t)-d \le y \le\eta(x,z,t) under gravity gg. Throughout this section we let x=(x,z)\mathbf{x}=(x,z) denote the horizontal coordinates, yy the vertical coordinate, η(x,t)\eta(\mathbf{x},t) the surface elevation, Φ(x,y,t)\Phi(\mathbf{x},y,t) the bulk velocity potential (u=∇Φ\mathbf{u}=\nabla\Phi, ΔΦ=0\Delta\Phi=0), and ψ(x,t)=Φ(x,η(x,t),t)\psi(\mathbf{x},t)=\Phi(\mathbf{x},\eta(\mathbf{x},t),t) the velocity potential evaluated on the moving free surface. Our full method couples the 2D wave solver developed in this section to a localized 3D NS solver (Sections 4 and 5); to keep the two components clearly distinguished, we refer to the former as the wave model and to the latter as the NS solver: when we speak of the “2D (wave) side” and the “3D (NS) side”, we mean these two subdomains and their respective solvers. The dimension labels describe our default configuration; the same construction applies verbatim in one dimension lower, where a 1D wave model embeds a 2D NS solver, as in some of our validation experiments (Section 8). Symbol conventions are summarized in Table 2.

NotationTypeDefinition
η\etascalarfree-surface elevation
ψ\psiscalarsurface velocity potential Φ∣surface\Phi|_{\mathrm{surface}}
Φ\Phiscalarbulk velocity potential, u=∇Φ\mathbf{u}=\nabla\Phi
G(η)G(\eta)operatorDirichlet–Neumann operator (DNO)
G0G_0operatorDNO at the flat state η=0\eta=0
B0B_0operatorFourier multiplier with symbol ∣k∣2\lvert k\rvert^2
k\mathbf{k}vectorhorizontal wave vector
ω\omegascalarwave angular frequency, ω2=gG0\omega^2=gG_0
u\mathbf{u}vectorvelocity, components (u,v,w)(u,v,w)
ε\varepsilonscalarnonlinear-amplitude parameter
ddscalarstill-water depth
σ(x)\sigma(\mathbf{x})scalarrelaxation source rate
ϕ\phiscalar3D level-set field
ttscalartime

Table 2. Summary of the main symbols and notations.

Canonical Zakharov formulation

Following Zakharov [82], the water-wave problem admits a Hamiltonian formulation in which the canonical pair is exactly (η,ψ)(\eta,\psi). The fluid Hamiltonian is the sum of kinetic and potential energy, expressed entirely in surface variables,

H[η,ψ]=12∫ψG(η)ψ dx+12g∫η2 dx(1)H[\eta,\psi] = \frac{1}{2}\int\psi G(\eta)\psi\,\mathrm{d}x + \frac{1}{2}g\int\eta^2\,\mathrm{d}x \tag*{(1)}

and the equations of motion are

∂tη=δHδψ,∂tψ=−δHδη.(2)\partial_t\eta= \frac{\delta H}{\delta\psi}, \qquad\partial_t\psi= -\frac{\delta H}{\delta\eta}. \tag*{(2)}

The single nontrivial object in this formulation is the Dirichlet–Neumann operator (DNO) G(η)G(\eta), defined by

G(η)ψ(x)=1+∣∇η∣2 ∂nΦ∣y=η.(3)G(\eta)\psi(\mathbf{x}) = \sqrt{1+\lvert\nabla\eta\rvert^2}\,\partial_n\Phi\big|_{y=\eta}. \tag*{(3)}

i.e., the operator that, for each prescribed Dirichlet datum ψ\psi on the moving surface, returns the normal derivative of the harmonic extension Φ\Phi into the fluid. Knowing G(η)G(\eta) collapses the original three-dimensional Laplace problem to two surface fields (η,ψ)(\eta,\psi) on a flat horizontal grid.

Carrying out the variational derivatives in (2) yields the canonical Zakharov system (Zakharov [82], recast by Craig and Sulem [16]),

∂tη=G(η)ψ,(4)\partial_t\eta= G(\eta)\psi, \tag*{(4)}
∂tψ=−gη−12∣∇ψ∣2+(G(η)ψ+∇η⋅∇ψ)22(1+∣∇η∣2).(5)\partial_t\psi= -g\eta- \frac{1}{2}\lvert\nabla\psi\rvert^2 + \frac{\left(G(\eta)\psi+\nabla\eta\cdot\nabla\psi\right)^2}{2\left(1+\lvert\nabla\eta\rvert^2\right)}. \tag*{(5)}

(4) is the kinematic free-surface boundary condition expressed through the DNO; (5) is the exact Bernoulli equation for the surface potential. The system is closed: a numerical method that can evaluate G(η)ψG(\eta)\psi for arbitrary (η,ψ)(\eta,\psi) on a regular grid yields a nonlinear ocean solver. Throughout the paper, “Hamiltonian” and “canonical” refer to this Zakharov state (η,ψ)(\eta,\psi) and its energy structure. This structure is not a prerequisite for high-quality coupling in general; its value here is that (η,ψ)(\eta,\psi) completely describes the irrotational surface state, so both directions of the 2D–3D exchange reduce to well-defined maps on these two fields: the 2D-to-3D direction (Section 4) lifts (η,ψ)(\eta,\psi) to a depth-resolved bulk velocity, and

the 3D-to-2D direction (Section 5) folds the 3D surface state back into a valid pair.

Spectral expansion of the Dirichlet–Neumann operator

A direct evaluation of G(η)ψG(\eta)\psi requires solving Laplace’s equation in the irregular fluid domain at every step. Craig and Sulem [16] observed that, for η\eta small relative to the characteristic wavelength 1/k1/k, G(η)G(\eta) admits a convergent Taylor expansion in homogeneous powers of η\eta that involves no Laplace solve at all. We adopt this expansion, known in ocean engineering as the High-Order Spectral (HOS) method [20], as the engine of our wave solver. Let G0G_0 denote the operator at the flat reference state η≡0\eta\equiv0. Its symbol on a horizontal Fourier mode eik⋅xe^{i\mathbf{k}\cdot\mathbf{x}} is

G0(k)=∣k∣tanh⁡(∣k∣d),(6)G_0(\mathbf{k}) = \lvert\mathbf{k}\rvert\tanh(\lvert\mathbf{k}\rvert d), \tag*{(6)}

which is exactly the linear water-wave dispersion operator and reproduces ω2(k)=gG0(k)\omega^2(\mathbf{k}) = gG_0(\mathbf{k}) at leading order [1]. Writing G(η)=∑j≥0Gj(η)G(\eta) = \sum_{j \ge0} G_j(\eta) with GjG_j homogeneous of degree jj in η\eta, the recursion of Craig and Sulem [16] gives, in spectral form,

G1(η)ψ=−G0(ηG0ψ)−∇⋅(η∇ψ),(7)G_1(\eta)\psi= -G_0(\eta G_0\psi) - \nabla\cdot(\eta\nabla\psi), \tag*{(7)}
G2(η)ψ=G0(ηG0(ηG0ψ))−12G0(η2B0ψ)−ηB0(ηG0ψ)+12η2G0(B0ψ)+∣∇η∣2G0ψ,(8)\begin{aligned} G_2(\eta)\psi={}& G_0(\eta G_0(\eta G_0\psi)) - \frac{1}{2}G_0(\eta^2 B_0\psi) \\ &{}- \eta B_0(\eta G_0\psi) + \frac{1}{2}\eta^2 G_0(B_0\psi) + \lvert\nabla\eta\rvert^2 G_0\psi, \tag*{(8)} \end{aligned}

where B0B_0 is the homogeneous Fourier multiplier with symbol ∣k∣\lvert\mathbf{k}\rvert (the second vertical derivative of the harmonic extension at the flat state). All operators in (7)–(8) are concatenations of Fourier multipliers (G0G_0, B0B_0, ∇\nabla) with pointwise products of η\eta, ∇η\nabla\eta, ∇ψ\nabla\psi, and are therefore evaluated with FFTs and physical-space multiplications. To improve stability in coupled simulations with steep 3D-driven surface features, we modulate every nonlinear term by a scalar amplitude parameter ε∈[0,1]\varepsilon\in[0,1],

G(η)≈G0+εG1(η)+ε2G2(η),(9)G(\eta) \approx G_0 + \varepsilon G_1(\eta) + \varepsilon^2G_2(\eta), \tag*{(9)}

and read ε=1\varepsilon= 1 as the fully nonlinear model and ε=0\varepsilon= 0 as the linear Airy limit. The dynamic equation (5) is rewritten consistently as

∂tψ=−gη+ε[−12∣∇ψ∣2+12(1+ε2∣∇η∣2)W2],(10)\partial_t\psi= -g\eta+ \varepsilon\left[-\frac{1}{2}\lvert\nabla\psi\rvert^2 + \frac{1}{2}(1+\varepsilon^2\lvert\nabla\eta\rvert^2)W^2\right], \tag*{(10)}

with

W=G0ψ+εG1(η)ψ+ε2G2(η)ψ+ε∇η⋅∇ψ1+ε2∣∇η∣2.(11)W = \frac{G_0\psi+\varepsilon G_1(\eta)\psi+\varepsilon^2G_2(\eta)\psi+\varepsilon\nabla\eta\cdot\nabla\psi}{1+\varepsilon^2\lvert\nabla\eta\rvert^2}. \tag*{(11)}

Throughout the paper we refer to the truncation in (9), which retains the three terms G0G_0, εG1\varepsilon G_1, ε2G2\varepsilon^2G_2, as the Order-3 (or HOS-3) scheme; the Order-2 (HOS-2) variant referenced in our experiments retains only G0G_0 and εG1\varepsilon G_1.

We verify the effectiveness of this ε\varepsilon modulation in Section 8: the wave-propagation sweep against a high-resolution 3D Eulerian reference (Figure 14) and the Stokes-wave sweep against the analytical second-order Stokes solution (Figure 15) both show the height error decreasing monotonically as ε\varepsilon grows from 0 to 1, with the drop between ε=0\varepsilon= 0 and ε≈0.1\varepsilon\approx0.1 already accounting for a large fraction of the linear-model error. In other words, even a small ε\varepsilon already yields a significant improvement over the linear case. Conversely, in scenarios where strong forcing pushes the surface close to the breaking threshold and the higher-order DNO terms can excite numerical instabilities, we deliberately keep ε\varepsilon small to preserve solver stability: ε=0.3\varepsilon= 0.3 for the submarine, ε=0.2\varepsilon= 0.2 for the battleship in heavy waves, and ε=0.2\varepsilon= 0.2 for the pond.

Figure 14. Wave height error in the shallow water wave propagation experiment, plotted against the 3D Eulerian reference as a function of the nonlinearity parameter ε\varepsilon, for both Order 2 and Order 3 HOS truncations. Hollow black circles mark the linear case (ε=0\varepsilon= 0); hollow green circles mark the fully nonlinear case (ε=1\varepsilon= 1). The error decreases monotonically as ε\varepsilon grows, confirming that the higher-order HOS terms are correctly oriented; even a small ε\varepsilon already removes a large fraction of the linear-model error. Order 2 and Order 3 are nearly indistinguishable for ε≲0.6\varepsilon\lesssim0.6 and diverge beyond, with Order 3 attaining a noticeably lower error at ε=1\varepsilon= 1.

Figure 15. Stokes wave propagation error against the analytical second-order Stokes solution as a function of the nonlinearity parameter ε\varepsilon. Four curves correspond to the two HOS truncation orders (O2, O3) and the two Stokes initial conditions parameterized by (ka,kh)(ka,kh). Hollow black circles mark the linear endpoint (ε=0\varepsilon= 0); hollow green circles mark the fully nonlinear endpoint (ε=1\varepsilon= 1).

Boundary conditions via even extension

The FFT machinery above is intrinsically periodic; closed boundaries (a tank wall, a coastline, the perimeter of a pond) require additional care. We support reflecting and replicating boundaries through an even extension of the wave domain. Given the physical state (η,ψ)(\eta,\psi) on a rectangle of resolution (M,N)(M,N), we form

η~(x)=η(Rx),ψ~(x)=ψ(Rx),(12)\widetilde{\eta}(\mathbf{x}) = \eta(R\mathbf{x}), \qquad\widetilde{\psi}(\mathbf{x}) = \psi(R\mathbf{x}), \tag*{(12)}

on a doubled (2M,2N)(2M,2N) grid, where RR denotes the appropriate reflection across each boundary. The operators G0G_0, G1G_1, and G2G_2 are then constructed on the doubled grid and applied to η~\widetilde{\eta}, ψ~\widetilde{\psi}; after the time step we crop the result back to (M,N)(M,N). The construction makes the discrete normal derivatives at the boundary vanish identically (Neumann condition) for reflecting walls, while a replicate variant approximates outflow. Because the extension only doubles the grid, the FFT cost increases only by a constant factor of four, and the boundary treatment carries no asymptotic overhead.

Time integration: exact-linear Adams–Bashforth

Following Craig and Sulem 2, we split the system into an exactly integrated linear part and an AB2-advanced nonlinear part:

∂t(ηψ)=L(ηψ)+N(η,ψ),(13)\partial_t \begin{pmatrix} \eta\\ \psi \end{pmatrix} = \mathcal{L} \begin{pmatrix} \eta\\ \psi \end{pmatrix} + \mathcal{N}(\eta,\psi), \tag*{(13)}

where the linear part L=(0G0−g0)\mathcal{L}=\begin{pmatrix}0&G_0\\-g&0\end{pmatrix} is diagonal in Fourier space with eigenfrequency ω(k)=gG0(k)\omega(\mathbf{k})=\sqrt{gG_0(\mathbf{k})}, and the nonlinear forcing N\mathcal{N} collects all terms in ε\varepsilon. The linear part is stiff at high wavenumber, i.e., ω(k)\omega(\mathbf{k}) scales like ∣k∣\sqrt{\lvert\mathbf{k}\rvert} in deep water, so an explicit step on the full RHS would require Δt≲1/ω(kNyq)\Delta t \lesssim1/\omega(k_{\mathrm{Nyq}}). We avoid the stiffness by integrating L\mathcal{L} exactly. Per Fourier mode, the linear flow is the rotation

\begin{pmatrix} \hat{\eta}\\ \hat{\psi} \end{pmatrix}_{\mathrm{lin}}(t+\Delta t) = \begin{pmatrix} \cos(\omega\Delta t) & \frac{G_0}{\omega}\sin(\omega\Delta t)\\ -\frac{g}{\omega}\sin(\omega\Delta t) & \cos(\omega\Delta t) \end{pmatrix} \begin{pmatrix} \hat{\eta}\\ \hat{\psi} \end{pmatrix}(t) \tag*{(14)}\end{pmatrix}

with the limit η^→η\hat{\eta}\to\eta, ψ^→ψ^−gΔt η^\hat{\psi}\to\hat{\psi}-g\Delta t\,\hat{\eta} at k=0\mathbf{k}=0. The nonlinear forcing N^n=(F^nη,F^nψ)\hat{\mathcal{N}}_n=(\hat{F}^{\eta}_n,\hat{F}^{\psi}_n) is then advanced through the Duhamel response of (14), i.e., the variation-of-constants formula, which propagates the accumulated nonlinear forcing with the exact linear flow: setting F^nq∗=32F^nq−12F^n−1q\hat{F}^{q*}_n=\frac{3}{2}\hat{F}^{q}_n-\frac{1}{2}\hat{F}^{q}_{n-1}, q∈{η,ψ}q\in\{\eta,\psi\} (an Adams–Bashforth-2 extrapolation), we update

η^n+1=cos⁡(ωΔt)η^n+G0ωsin⁡(ωΔt)ψ^n+sin⁡(ωΔt)ωF^nη∗+1−cos⁡(ωΔt)gF^nψ∗(15)\hat{\eta}_{n+1} = \cos(\omega\Delta t)\hat{\eta}_n + \frac{G_0}{\omega}\sin(\omega\Delta t)\hat{\psi}_n + \frac{\sin(\omega\Delta t)}{\omega}\hat{F}^{\eta*}_n + \frac{1-\cos(\omega\Delta t)}{g}\hat{F}^{\psi*}_n \tag*{(15)}
ψ^n+1=−gωsin⁡(ωΔt)η^n+cos⁡(ωΔt)ψ^n−1−cos⁡(ωΔt)G0F^nη∗+sin⁡(ωΔt)ωF^nψ∗(16)\hat{\psi}_{n+1} = -\frac{g}{\omega}\sin(\omega\Delta t)\hat{\eta}_n + \cos(\omega\Delta t)\hat{\psi}_n - \frac{1-\cos(\omega\Delta t)}{G_0}\hat{F}^{\eta*}_n + \frac{\sin(\omega\Delta t)}{\omega}\hat{F}^{\psi*}_n \tag*{(16)}

The first step (and any step in which Δt\Delta t changes) bootstraps with F^0∗=N0\hat{F}^{*}_0=\mathcal{N}_0, which reduces (15)–(16) to a first-order integrating-factor Euler step. The k=0k=0 singularities cancel analytically: sin⁡(ωΔt)/ω→Δt\sin(\omega\Delta t)/\omega\to\Delta t and (1−cos⁡(ωΔt))/G0→12gΔt2(1-\cos(\omega\Delta t))/G_0\to\frac{1}{2}g\Delta t^2.

The scheme has three concrete advantages for our coupled setting. First, (14) integrates the linear dispersion exactly per Fourier mode, so the surface step is non-dissipative on linear waves regardless of Δt\Delta t. Second, the AB2 extrapolation only touches the nonlinear forcing, which is bounded; the time-step constraint reduces from ω(kNyq)Δt<1\omega(k_{\mathrm{Nyq}})\Delta t<1 to a much milder CFL on N\mathcal{N}. Third, the same exponential-integrator strategy extends to the implicit relaxation source we add in Section 5: that source is linear in η−ηtarget\eta-\eta^{\mathrm{target}} with a position-dependent rate σ(x)\sigma(x), so it admits its own closed-form exponential decay (34), which we compose with the wave step by operator splitting (Step 8 of Alg. 2). Each half is therefore integrated exactly over Δt\Delta t, and the time-step constraint never tightens with σmax⁡\sigma_{\max}. After each step we apply the spectral state filter of Craig and Sulem (1993),

A(k)=∏α∈{x,y,z}[18(5+4cos⁡(π∣kα∣kNyqα)−cos⁡(2π∣kα∣kNyqα))](17)A(k)=\prod_{\alpha\in\{x,y,z\}}\left[\frac{1}{8}\left(5+4\cos\left(\pi\frac{\lvert k_\alpha\rvert}{k^\alpha_{\mathrm{Nyq}}}\right)-\cos\left(2\pi\frac{\lvert k_\alpha\rvert}{k^\alpha_{\mathrm{Nyq}}}\right)\right)\right] \tag*{(17)}

to (η,ψ)(\eta,\psi). The filter is unity at low wavenumbers, smoothly suppresses the highest modes, and removes residual high-frequency noise without affecting the resolved spectrum.

State recovery: forward and inverse DNO maps

The wave solver evolves the canonical pair (η,ψ)(\eta,\psi), but the surrounding pipeline routinely needs to translate between this canonical state and the physically observable surface vertical velocity v=G(η)ψv=G(\eta)\psi. The forward direction (η,ψ)↦v(\eta,\psi)\mapsto v is the right-hand side of the kinematic Hamilton equation (4), evaluated every substep by the time integrator. The inverse direction (η,v)↦ψ(\eta,v)\mapsto\psi appears in two places. First, in the 3D-to-2D relaxation of Section 5, the canonical conjugate ψ\psi is reconstituted from the post-relaxation pair (η,v)(\eta,v) to keep the wave state self-consistent across the relaxation step. Second, initial conditions and analytically prescribed sea states are naturally specified in physical (η,v)(\eta,v) form and must be lifted back to the canonical pair before the time integrator can advance them. We therefore expose both maps as utilities.

Forward map. The forward map is a single DNO application,

v=G0ψ+εG1(η)ψ+ε2G2(η)ψ,(18)v = G_0\psi+\varepsilon G_1(\eta)\psi+\varepsilon^2G_2(\eta)\psi, \tag*{(18)}

evaluated by the same de-aliased FFT pipeline (Section 6.1).

Inverse map. The inverse map, needed whenever the wave state is initialized from a physical (η,v)(\eta,v) pair or when the relaxation source modifies η\eta relative to ψ\psi, is obtained by fixed-point iteration on

ψ(n+1)=G0−1(v−εG1(η)ψ(n)−ε2G2(η)ψ(n)),(19)\psi^{(n+1)}=G_0^{-1}\left(v-\varepsilon G_1(\eta)\psi^{(n)}-\varepsilon^2G_2(\eta)\psi^{(n)}\right), \tag*{(19)}

seeded with ψ(0)=G0−1v\psi^{(0)}=G_0^{-1}v. Because εG1\varepsilon G_1 and ε2G2\varepsilon^2G_2 are subdominant relative to G0G_0 in the parameter regimes of interest, the iteration is contractive and converges in a handful of passes; we use eight iterations throughout. Each iteration evaluates the nonlinear DNO correction once and applies G0−1G_0^{-1} through pointwise division in Fourier space, yielding O(N2log⁡N)O(N^2\log N) cost per iteration.

2D-to-3D Coupling: Bulk Velocity from the Canonical Surface State

The 3D NS solver lives in a localized embedding box Ω3D⊂R3\Omega_{\mathrm{3D}}\subset\mathbb{R}^3 surrounding the area of interest (a hull, a falling object, an outflow); see Figure 5 for the overall layout of the coupled system. Its lateral and bottom faces sit inside the open ocean simulated by the wave model of Section 3, and at every substep the 3D solver requires a velocity boundary condition on the cells that lie outside the air phase but on the box perimeter. We build that boundary condition by reconstructing the bulk velocity field implied by the canonical surface state (η,ψ)(\eta,\psi) through finite depth, depth layer by depth layer; the left panel of Figure 4 illustrates this pipeline, from the HOS lift of (η,ψ)(\eta,\psi) to Φ∣y=0\Phi|_{y=0} through the harmonic continuation down to the depth layers from which u=∇Φu=\nabla\Phi is read off. The reverse half of the loop, in which the 3D surface state is folded back into the canonical pair through a relaxation source on the wave-side η\eta equation, is deferred to Section 5; in this section we treat the relaxation as a black box and focus solely on lifting (η,ψ)(\eta,\psi) to a depth-resolved bulk velocity on the FAB ring.

Figure 4. Two-way coupling between the 2D wave solver and the 3D NS solver. Left, 2D-to-3D (Section 4): (η,ψ)(\eta,\psi) is lifted to y=0y=0 via the Craig–Sulem (HOS) Taylor expansion and propagated to depth layers by finite-depth harmonic kernels under Dirichlet top and Neumann no-flux bottom ∂yϕ∣y=−d=0\partial_y\phi|_{y=-d}=0. The FAB ring on the box faces takes its level set BC from the 2D height field via ϕ=y−η\phi=y-\eta and its velocity BC from u=∇ϕu=\nabla\phi. Right, 3D-to-2D (Section 5): an exponential-decay source on the box footprint relaxes the wave-side η\eta toward the 3D surface η3D\eta_{3D} implicitly inside the relaxation zone, while outside the box the Zakharov equation ∂tη=G(η)ψ\partial_t\eta=G(\eta)\psi evolves the ocean freely.

Overview of the two-way coupled method showing localized 3D NB-FLIP solvers inside embedding boxes and a 2D nonlinear wave solver

Figure 5. Overview of our two-way coupled method: a localized 3D NB-FLIP solver inside each embedding box (black outline) resolves the local 3D flow, and a 2D nonlinear wave solver evolves the surrounding ocean on a structured grid. Multiple 3D regions can coexist and interact through the shared wave field (bottom).

The FAB layer: coupling boundary of the 3D box

Following the FAB method of Stomakhin and Selle [66], we mark the outermost band of cells on the lateral and bottom faces of Ω3D\Omega_{\mathrm{3D}} as a FAB layer of width hFABh_{\mathrm{FAB}} (four cells in all our runs); the band is drawn in the left panel of Figure 4. FAB cells are excluded from the 3D pressure projection and instead carry Dirichlet data prescribed by the wave model each substep: the level set is taken directly from the canonical height field, ϕ(x,y)=y−η(x,t)\phi(x,y)=y-\eta(x,t), with FLIP particles reseeded inside the band to maintain narrow-band density, and the cell-centered velocity is read off the layered bulk reconstruction derived in the remainder of this section.

Bulk velocity potential from the canonical surface pair

Recall that u=∇Φu=\nabla\Phi and Φ\Phi is harmonic in the fluid. Inside the strip −d≤y≤η(x,t)-d\leq y\leq\eta(x,t), Φ\Phi is uniquely determined by its surface trace ψ(x,t)=Φ∣y=η\psi(x,t)=\Phi|_{y=\eta} and the bottom no-flux condition ∂yΦ∣y=−d=0\partial_y\Phi|_{y=-d}=0. Reconstructing Φ\Phi from ψ\psi on the moving surface is exactly the elliptic inverse to the Dirichlet–Neumann map of Section 3.1, and the same Craig–Sulem expansion that produces G(η)ψG(\eta)\psi produces, with one fewer FFT, the boundary value of Φ\Phi on the flat reference plane y=0y=0. Specifically, writing Φ∣y=0=∑j≥0Φ(j)∣y=0\Phi|_{y=0}=\left.\sum_{j\geq0}\Phi^{(j)}\right|_{y=0} as a Taylor series in homogeneous powers of η\eta, the recursion of Craig and Sulem [16] gives

Φ(0)∣y=0=ψ,(20)\left.\Phi^{(0)}\right|_{y=0}=\psi, \tag*{(20)}
Φ(1)∣y=0=−ηG0ψ,(21)\left.\Phi^{(1)}\right|_{y=0}=-\eta G_0\psi, \tag*{(21)}
Φ(2)∣y=0=−ηG0Φ(1)∣y=0−12η2B0ψ.(22)\left.\Phi^{(2)}\right|_{y=0}=-\eta G_0\Phi^{(1)}\big|_{y=0}-\frac{1}{2}\eta^2B_0\psi. \tag*{(22)}

We retain the expansion through first order, Φ∣y=0≈ψ−εηG0ψ\Phi|_{y=0}\approx\psi-\varepsilon\eta G_0\psi, computed via 3/2-rule de-aliasing [54] exactly as in the time step. The second-order term (22) contains B0B_0, whose ∣k∣2|k|^2 symbol amplifies any high-wavenumber forcing that the relaxation source (Section 5) injects into the wave state when the 3D box transmits a hull depression or a localized splash; we therefore truncate at first order to keep the boundary condition robust under arbitrary 3D forcing.

Comparison of four 2D–3D coupling methods in the dispersion matching test

Figure 6. Comparison of four 2D–3D coupling methods in the dispersion matching test, all driven by the same 3D NB-FLIP solver. The method of Chentanez et al. [2015] exhibits clear left–right asymmetry in the 2D wave field. The method of Schreck and Wojtan [2022], run with its dispersion-law rescaling factor recalibrated to our 3D solver configuration (d=0.8d = 0.8; see Section 8), removes most of the left–right asymmetry, but its 2D and 3D regions remain less consistent than ours. FAB [Stokman and Selle 2017] combined with the Airy Dispersion Kernel (DK) wave model [Canabal et al. 2016] shows no visible asymmetry, but produces visibly deeper waves inside the 3D region than in the surrounding 2D region. Our method produces a wave field that is nearly indistinguishable across the 2D and 3D regions, while exhibiting smoother waveforms in the 3D region. The wave-solver times are 2.3 ms, 27.5 ms, 3.7 ms, and 13.3 ms per substep for Chentanez et al. [2015], Schreck and Wojtan [2022], FAB + Airy DK, and our method, respectively. Our wave model uses ε=1\varepsilon= 1.

Finite-depth harmonic kernels

Once Φ∣y=0\left.\Phi\right|_{y=0} is in hand, the bulk potential at every depth follows from the harmonic extension into a flat strip with rigid bottom. The unique solution to ΔΦ=0\Delta\Phi= 0 on x∈R2\boldsymbol{x} \in\mathbb{R}^{2}, −d≤y≤0-d \le y \le0, with prescribed Φ∣y=0\left.\Phi\right|_{y=0} and ∂yΦ∣y=−d=0\left.\partial_y\Phi\right|_{y=-d} = 0, separates per Fourier mode as

Φ^(k,y)=cosh⁡(∣k∣(y+d))cosh⁡(∣k∣d)Φ^(k,0).(23)\widehat{\Phi}(\boldsymbol{k},y) = \frac{\cosh\left(|\boldsymbol{k}|(y+d)\right)}{\cosh\left(|\boldsymbol{k}|d\right)}\widehat{\Phi}(\boldsymbol{k},0). \tag*{(23)}

For a query depth yi∈[0,d]y_i \in[0,d] measured downward from the flat reference (so the actual coordinate is y=−yiy=-y_i), define

Ci(k)=cosh⁡(∣k∣(d−yi))cosh⁡(∣k∣d),(24)C_i(\boldsymbol{k}) = \frac{\cosh\left(|\boldsymbol{k}|(d-y_i)\right)}{\cosh\left(|\boldsymbol{k}|d\right)}, \tag*{(24)}
Si(k)=∣k∣sinh⁡(∣k∣(d−yi))cosh⁡(∣k∣d).(25)S_i(\boldsymbol{k}) = |\boldsymbol{k}|\frac{\sinh\left(|\boldsymbol{k}|(d-y_i)\right)}{\cosh\left(|\boldsymbol{k}|d\right)}. \tag*{(25)}

The two kernels carry the horizontal continuation factor and the vertical-derivative factor at depth yiy_i, respectively. Their limits encode the physical boundary conditions: at the surface (yi=0y_i = 0) we recover C0=1C_0 = 1 and S0=G0S_0 = G_0, so the leading-order vertical surface velocity is exactly G0ψG_0\psi; at the bottom (yi=dy_i = d) we recover Cd=1/cosh⁡(∣k∣d)C_d = 1/\cosh(|\boldsymbol{k}|d) and Sd=0S_d = 0, enforcing the no-flux condition. In the deep-water limit ∣k∣d→∞|\boldsymbol{k}|d \to\infty, both kernels collapse to e−∣k∣yie^{-|\boldsymbol{k}|y_i} (with ∣k∣|\boldsymbol{k}| multiplied for SiS_i), recovering the exponential attenuation used by linear deep-water coupling schemes; our formulation generalizes this result to finite depth.

For numerical robustness on float32 GPU arithmetic, we rewrite (24)–(25) in their stable exponential form,

Ci=e−∣k∣yi+e−∣k∣(2d−yi)1+e−2∣k∣d,Si=∣k∣e−∣k∣yi−e−∣k∣(2d−yi)1+e−2∣k∣d,(26)C_i = \frac{e^{-|\boldsymbol{k}|y_i} + e^{-|\boldsymbol{k}|(2d-y_i)}}{1 + e^{-2|\boldsymbol{k}|d}}, \qquad S_i = |\boldsymbol{k}|\frac{e^{-|\boldsymbol{k}|y_i} - e^{-|\boldsymbol{k}|(2d-y_i)}}{1 + e^{-2|\boldsymbol{k}|d}}, \tag*{(26)}

which avoids overflow when ∣k∣d|\boldsymbol{k}|d is large at high wavenumbers in shallow water. At k=0k = 0 we set Ci=1C_i = 1, Si=0S_i = 0 to suppress the spurious uniform-mode flow.

Velocity reconstruction at each depth layer

We sample the embedding box at LdL_d depth layers, yi=iΔydy_i = i\Delta y_d for i=0,…,Ld−1i = 0,\ldots,L_d-1, with the layer spacing Δyd\Delta y_d chosen to match the box height. At each layer we obtain horizontal and vertical velocities from the spectral Φ∣y=0\left.\Phi\right|_{y=0} by

u^i(k)=ikxCi(k)Φ^∣y=0(k),(27)\widehat{u}_i(\boldsymbol{k}) = i k_x C_i(\boldsymbol{k}) \left.\widehat{\Phi}\right|_{y=0}(\boldsymbol{k}), \tag*{(27)}
w^i(k)=ikzCi(k)Φ^∣y=0(k),(28)\widehat{w}_i(\boldsymbol{k}) = i k_z C_i(\boldsymbol{k}) \left.\widehat{\Phi}\right|_{y=0}(\boldsymbol{k}), \tag*{(28)}
v^i(k)=Si(k)Φ^∣y=0(k),(29)\widehat{v}_i(\boldsymbol{k}) = S_i(\boldsymbol{k}) \left.\widehat{\Phi}\right|_{y=0}(\boldsymbol{k}), \tag*{(29)}

followed by inverse FFT and, when reflecting boundaries are active, cropping back to the physical (M,N)(M,N) grid. The cost is LdL_d FFT pairs per substep, all on the same surface grid. We then trilinearly interpolate the layered velocity field onto the 3D box’s lateral and bottom face cells, i.e., the FAB ring of cells that the NB-FLIP solver consumes as a velocity boundary condition during pressure projection. The layered reconstruction is summarized in Algorithm 1.

5 3D-to-2D Coupling: Implicit Relaxation in Canonical Variables

The reverse half of the coupling closes the loop: at every substep, the 3D NS solution inside the embedding box must be transferred back to the wave model so that hull depressions, splashes, and other forced features born in the 3D regime appear smoothly in the open-ocean surface field. The right panel of Figure 4 sketches this direction: the 3D solver advances (u,ϕ)(\boldsymbol{u},\phi) inside the box and exposes its surface elevation η3D\eta_{3D}, which is then folded back into the wave-side η\eta through a relaxation zone supported on the box footprint. Two physical incompatibilities make this nontrivial. First, the wave Algorithm 1 ComputeVolumeVelFrom2D (bulk velocity from (η,ψ)(\eta,\psi) via the Craig–Sulem extension and finite-depth kernels)

Input: surface state (η,ψ)(\eta,\psi) on wave grid, kernels {Ci,Si}i=0Ld−1\{C_i,S_i\}_{i=0}^{L_d-1}, padding mode Pad.

Output: layered velocities {(ui,vi,wi)}i=0Ld−1\{(u_i,v_i,w_i)\}_{i=0}^{L_d-1}.

  1. if Pad then

  2. (η,ψ)←EvenExtend⁡(η,ψ)(\eta,\psi) \leftarrow\operatorname{EvenExtend}(\eta,\psi) ▸ Sec. 3.3

  3. end if

  4. η^,ψ^←FFT⁡(η),FFT⁡(ψ)\hat{\eta},\hat{\psi} \leftarrow\operatorname{FFT}(\eta),\operatorname{FFT}(\psi)

  5. Φ^∣y=0←ψ^−εη⋅G0ψ^\left.\hat{\Phi}\right|_{y=0} \leftarrow\hat{\psi}-\varepsilon\widehat{\eta\cdot G_0\psi} ▸ (21)

  6. Φ^∣y=0←Φ^∣y=0⋅taper⁡couple\left.\hat{\Phi}\right|_{y=0} \leftarrow\left.\hat{\Phi}\right|_{y=0}\cdot\operatorname{taper}_{\mathrm{couple}} ▸ Sec. 6.2

  7. for i=0,…,Ld−1i=0,\ldots,L_d-1 do

  8. u^i,w^i←ikxCiΦ^∣y=0,ikzCiΦ^∣y=0\hat{u}_i,\hat{w}_i \leftarrow\mathrm{i}k_x C_i\left.\hat{\Phi}\right|_{y=0},\mathrm{i}k_z C_i\left.\hat{\Phi}\right|_{y=0}

  9. v^i←SiΦ^∣y=0\hat{v}_i \leftarrow S_i\left.\hat{\Phi}\right|_{y=0}

  10. (ui,vi,wi)←IFFT⁡(u^i,v^i,w^i)(u_i,v_i,w_i) \leftarrow\operatorname{IFFT}(\hat{u}_i,\hat{v}_i,\hat{w}_i)

  11. if Pad then (ui,vi,wi)←Crop⁡(ui,vi,wi)(u_i,v_i,w_i) \leftarrow\operatorname{Crop}(u_i,v_i,w_i)

  12. end if

  13. end for

model is canonical in (η,ψ)(\eta,\psi) but the 3D NS solver tracks a level set ϕ\phi and a velocity field u\mathbf{u}; there is no canonical ψ\psi directly available on the 3D side, so the wave model must reconstitute it. Second, naive overwriting of the wave state inside the box would create a discontinuity at the box edge that radiates spurious reflections back into the open ocean. We resolve the first issue by transferring only the surface elevation η\eta from the 3D solver and reconstructing the canonical ψ\psi on the wave side through an inverse DNO solve ψ=G(η)−1v\psi=G(\eta)^{-1}v, which converts the pair (η,v)(\eta,v) back into the canonical conjugate, so that no fluid-derived source acts on the dynamic equation for ψ\psi. The second issue is handled by formulating the data transfer as an implicit relaxation source on the Zakharov right-hand side that decays exponentially through a smooth transition band.

Kinematic targets from the 3D state

Let ϕ3D(x,y,t)\phi^{3D}(\mathbf{x},y,t) be the 3D level set at the current substep, with yy the vertical axis and (x,z)(x,z) the horizontal plane of the wave grid. We construct the elevation target ηtarget(x,z)\eta^{\mathrm{target}}(x,z) on the wave grid by a column-wise scan of ϕ3D\phi^{3D}, following the strategy of [15]. For each wave-grid column (i,k)(i,k), we sweep the 3D cells from the bottom upward and count consecutive air/solid cells; once the running count reaches a small threshold nairn_{\mathrm{air}} (we use nair=3n_{\mathrm{air}}=3), we identify the cell nairn_{\mathrm{air}} steps below as the surface cell j⋆j^\star, and set

ηi,ktarget=(j⋆+12+oy)Δx−ϕi,j⋆,k3D,(30)\eta^{\mathrm{target}}_{i,k}=\left(j^\star+\frac{1}{2}+o_y\right)\Delta x-\phi^{3D}_{i,j^\star,k}, \tag*{(30)}

i.e., the yy-coordinate of the surface cell center, corrected by the signed-distance value at that cell as a sub-cell offset (oyo_y is the box origin offset in yy). To keep this target compatible with the wave grid, we then apply a few iterations of FAB-band padding plus a small Laplace smoothing to ηtarget\eta^{\mathrm{target}}; this damps the high-frequency staircase noise that column-scanned level sets exhibit.

Relaxation source on the Zakharov system

With the elevation target ηtarget\eta^{\mathrm{target}} in hand, we drive the wave state toward the 3D state by augmenting the Zakharov system with a relaxation source on the canonical equations,

∂tη=G(η)ψ−σ(x)(η−ηtarget),(31)\partial_t\eta=G(\eta)\psi-\sigma(x)(\eta-\eta^{\mathrm{target}}), \tag*{(31)}
∂tψ=−gη+Nψ,(32)\partial_t\psi=-g\eta+\mathcal{N}_\psi, \tag*{(32)}

in which σ(x)≥0\sigma(x)\ge0 is a spatially graded relaxation rate (units 1/s1/\mathrm{s}) supported on the embedding-box footprint. We choose

σ(x)=σmax⁡ρ(dist⁡(x,∂Ω3D)wb)1x∈Ω3D,(33)\sigma(x)=\sigma_{\max}\rho\left(\frac{\operatorname{dist}(x,\partial\Omega_{3D})}{w_b}\right)\mathbf{1}_{x\in\Omega_{3D}}, \tag*{(33)}

where ∂Ω3D\partial\Omega_{3D} is the box edge, wbw_b is the transition-band width, ρ:[0,1]→[0,1]\rho:[0,1]\to[0,1] is a smooth ramp (ρ(0)=0,ρ(1)=1\rho(0)=0,\rho(1)=1), and 1x∈Ω3D\mathbf{1}_{x\in\Omega_{3D}} is the indicator of the box footprint, which restricts the support of σ\sigma to the box interior. Outside the box, σ(x)≡0\sigma(x)\equiv0 and the wave evolves freely. In the transition band σ\sigma rises from 0 to σmax⁡\sigma_{\max}, and in the box interior σ=σmax⁡\sigma=\sigma_{\max} pulls the wave state toward the 3D state at rate σmax⁡\sigma_{\max}. Here, the rate σ\sigma is a property of position, not of the wave field, so the source enters (31) linearly; this is what makes it amenable to exact integration.

The source acts on the kinematic equation (31) alone because (η,ψ)(\eta,\psi) are canonically conjugate rather than two independent fields: the wave Hamiltonian already commits ψ\psi’s evolution to η\eta through the dynamic equation (32), so pinning η\eta to ηtarget\eta^{\mathrm{target}} already determines what ψ\psi must do, and adding an independent ψtarget\psi^{\mathrm{target}} source would risk over-determining the pair and breaking the symplectic identity v=G(η)ψ=∂tηv=G(\eta)\psi=\partial_t\eta inside the box. In practice such a fluid-derived target carries the NS solver’s numerical viscosity and discretization artifacts, so it generally disagrees with the conjugate the wave Hamiltonian produces, and the mismatch radiates from the box edge. We validate this choice empirically in Figure 17: additionally driving ψ\psi toward a Helmholtz-decomposed target extracted from the 3D surface velocities produces visibly degraded results.

Figure 17. Ablation on relaxing ψ\psi on the Crown Splash setup. The right panel additionally relaxes ψ\psi toward a target obtained from a Helmholtz decomposition of the 3D surface velocity field, while the left panel is our method. Relaxing ψ\psi produces an obvious abnormal depression in the 3D region.

Implicit integration of the relaxation source

A naive explicit treatment ηn+1=ηn−σΔt(ηn−ηtarget)\eta_{n+1} = \eta_n - \sigma\Delta t \left(\eta_n - \eta^{\mathrm{target}}\right) becomes unstable for σmax⁡Δt>2\sigma_{\max}\Delta t > 2, which would force a uselessly small Δt\Delta t when the box must absorb a violent feature. Because the source is linear in the deviation η−ηtarget\eta- \eta^{\mathrm{target}}, the relaxation has a closed-form solution: holding ηtarget\eta^{\mathrm{target}} frozen across one substep, the relaxation flow is

η(t+Δt)=ηtarget+(η(t)−ηtarget)e−σ(x)Δt,(34)\eta(t + \Delta t) = \eta^{\mathrm{target}} + \left(\eta(t) - \eta^{\mathrm{target}}\right)e^{-\sigma(\mathbf{x})\Delta t}, \tag*{(34)}

which is unconditionally stable, monotonically convergent, and exact at every σ\sigma. We compose this analytic decay with the wave step of Section 3.4 via operator splitting: first advance (η,ψ)(\eta,\psi) by Δt\Delta t under the wave Hamiltonian, then apply (34) pointwise.

(34) updates η\eta but says nothing about the wave’s surface vertical velocity v=G(η)ψv = G(\eta)\psi, and we deliberately do not introduce a vtargetv^{\mathrm{target}} to drive it with. At every substep we preserve the wave’s own vv across the η\eta update (the value v=G(ηstep)ψstepv = G(\eta_{\mathrm{step}})\psi_{\mathrm{step}} carried out of the Hamiltonian substep) and recompute the canonical conjugate from the post-relaxation pair (ηrelaxed,v)(\eta^{\mathrm{relaxed}},v) via an inverse DNO solve,

ψrelaxed=[G(ηrelaxed)]−1v,(35)\psi^{\mathrm{relaxed}} = \left[G\left(\eta^{\mathrm{relaxed}}\right)\right]^{-1}v, \tag*{(35)}

see (19). The triple (η,ψ,v)(\eta,\psi,v) is therefore self-consistent at the end of each substep.

We do not introduce a vtargetv^{\mathrm{target}} because the information vv carries is already encoded in the time history of ηtarget\eta^{\mathrm{target}}: the canonical relation v=G(η)ψv = G(\eta)\psi together with the unmodified Hamilton evolution of ψ\psi means that pinning η\eta to ηtarget\eta^{\mathrm{target}} across substeps automatically pulls the wave’s vv toward ∂tηtarget\partial_t\eta^{\mathrm{target}}, so a separate vv target would supply little independent information. The single-boat experiment in Figure 12 also supports this choice: variant (c), which additionally relaxes η\eta toward a finite-difference target built from the 3D level set, introduces spurious reflections behind the stern, while variant (e), which relaxes η\eta alone, lets the outgoing wake pass through cleanly.

Single-boat wake comparison across seven methods

Figure 12. Single-boat wake comparison across seven methods. (a) Chentanez et al. [2015] (the white box marks the common 3D simulation region): the wake is qualitatively wrong, as the shallow water equations cannot resolve the dispersive Kelvin pattern. (b) Airy DK [Canabal et al. 2016] coupled via the FAB scheme [Stomakhin and Selle 2017], (c) coupling through height and velocity relaxation, and (d) our wave model combined with the coupling scheme of Chentanez et al. [2015] all reproduce the feather-like wake [Rabaud and Moisy 2013] but exhibit clear spurious reflections behind the stern (highlighted by the red boxes). (e) Our full method reproduces the feather-like Kelvin wake with minimal spurious reflection at the 2D–3D interface. (f) 3D NB-FLIP reference: the trailing wave at the far right is noticeably damped due to numerical dissipation of the 3D solver. (g) Pure 3D Eulerian level-set simulation. The variants built on our wave model, (c), (d), and (e), use ε=1\varepsilon= 1.

Implementation Details

Our method is implemented in Warp [47] with GPU acceleration. In 2D simulation, we utilize PyTorch’s FFT module and padding functions to handle spectral operations and boundary conditions efficiently. Our 2D solver is compatible with standard ocean spectrum initializations. In some of our experiments, the initial height and velocity are generated from a TMA spectrum with Donelan–Banner directional spreading, following [29]. For 3D simulation, we implement NB-FLIP [22] with weak two-way solid–fluid coupling following [9]. Solid objects are represented in voxelized cells. The simulation runs on staggered grids. Pressure projection is solved with a CUDA-based unsmoothed aggregated Algebraic Multigrid Preconditioned Conjugate Gradient solver [60, 68]. Whitewater (spray, foam, and bubbles) shown in our results is simulated as a post-process using Houdini’s built-in whitewater solver, driven by the velocity and surface fields produced by our simulation. The fine-scale ripples and shimmering visible on the ocean surface in our rendered results are small ocean-spectrum displacements added in Houdini at render time for visual richness; they are disabled in all quantitative comparisons.

Masking residual coupling artifacts. While our coupling suppresses most interface artifacts, it does not eliminate them entirely. In our most demanding scene, the battleship in heavy seas (Figure 11), the accompanying animation shows wavefronts reflecting from the rectangular boundary of the 3D box and propagating into the 2D domain. In production settings, such residual artifacts can be effectively obscured using two standard render-time layers. First, the ocean-spectrum displacements described above can be amplified to overlay additional procedural spectral detail on the simulated surface. Second, foam and whitewater can be emitted to further obscure the remaining artifacts.

Battleship sailing through heavy waves under weak two-way solid–fluid coupling

Figure 11. Battleship sailing through heavy waves under weak two-way solid–fluid coupling. The 2D wave domain measures 120 m×60 m120\,\mathrm{m} \times60\,\mathrm{m} and embeds a 24 m×12 m×12 m24\,\mathrm{m} \times12\,\mathrm{m} \times12\,\mathrm{m} 3D fluid box around the hull. Columns correspond to three time instants (t=6.25t = 6.25, 10.0010.00, 12.25 s12.25\,\mathrm{s}); the rows show, from top to bottom, the far, mid, and near views of the same scene. The simulation remains stable under the near-breaking background spectrum, and the 2D and 3D regions stay consistent in the frames shown; residual boundary artifacts visible in the accompanying animation are discussed in Section 8. The wave model uses ε=0.2\varepsilon= 0.2.

De-aliased nonlinear evaluation

Because every GjG_j for j≥1j \ge1 is built from quadratic and higher-order pointwise products in η\eta, ∇η\nabla\eta, ∇ψ\nabla\psi, G0ψG_0\psi, and B0ψB_0\psi, naive spectral evaluation aliases high-wavenumber content back into the resolved band and corrupts the nonlinear corrections within a few steps. We follow Orszag’s 3/2-rule [54]: every quadratic product in (7), (8), and (10) is computed by zero-padding both factors from (M,N)(M,N) to (32M,32N)(\frac{3}{2}M,\frac{3}{2}N), multiplying in physical space on the padded grid, and truncating the result back to (M,N)(M,N). The padded grid is large enough that the highest aliased mode ∣k∣+∣k′∣=2∣k∣max⁡\lvert\mathbf{k}\rvert+\lvert\mathbf{k}'\rvert=2\lvert\mathbf{k}\rvert_{\max} falls outside the retained band of the truncated product, so the surviving modes are exact. After de-aliasing we apply a smooth super-Gaussian taper exp⁡[−(∣k∣/kc)8]\exp[-(\lvert\mathbf{k}\rvert/k_c)^8] with kc=23kNyqk_c=\frac{2}{3}k_{\mathrm{Nyq}} to all nonlinear products. The taper is unity through most of the resolved band and prevents the high-wavenumber tail of G2G_2, whose symbol grows like ∣k∣2\lvert\mathbf{k}\rvert^2, from amplifying numerical noise injected by the coupling boundary.

Spectral bandlimiting of the boundary condition

The kernels in (27)–(29) act as spectral derivatives, so they amplify any high-wavenumber content in Φ^∣y=0\hat{\Phi}|_{y=0} linearly in ∣k∣\lvert\mathbf{k}\rvert. That tail is also the least reliable part of the wave field, since the truncated Craig–Sulem expansion loses accuracy near the wave grid’s Nyquist. Feeding it unaltered into the 3D solver injects near-grid-scale forcing that the pressure projection then has to fight. We therefore multiply Φ^∣y=0\hat{\Phi}|_{y=0} by a separable super-Gaussian taper exp⁡[−(∣kα∣/(12kNyqα))8]\exp\left[-\left(\lvert k_\alpha\rvert/\left(\frac{1}{2}k_{\mathrm{Nyq}}^\alpha\right)\right)^8\right] for α∈{x,z}\alpha\in\{x,z\} before applying (27)–(29). The taper is unity through the lower half of the spectrum and decays smoothly to suppress the upper half, thereby preserving the resolved nonlinear wave content while keeping the 3D pressure solve free of injected high-wavenumber noise.

Time Integration

At each substep the 3D solver advances by Δt3D\Delta t_{3D}, set by an outer CFL on the 3D advection and solid motion. The 2D wave solver uses its own step Δt2D≤Δt3D\Delta t_{2D}\le\Delta t_{3D}: in most of our scenes we simply take Δt2D=Δt3D\Delta t_{2D}=\Delta t_{3D}, but when finer resolution of the surface dynamics is desired we sub-cycle the wave solver with a smaller Δt2D\Delta t_{2D} within one 3D step. The main algorithm is summarized below and outlined in Alg. 2. In the enumerated description that follows, the parenthetical labels refer to the corresponding line numbers of Alg. 2.

(1) Compute Time Step (Step 1). Based on the CFL condition on the 3D velocity field and the solid motion, we compute the substep size Δt\Delta t.

(2) Advect 3D (Step 2). We perform the 3D level-set advection and velocity transport as in standard NB-FLIP, producing the tentative state (ϕ3Dn+1,u3Dn+1)(\phi_{3D}^{n+1},\mathbf{u}_{3D}^{n+1}). (3) Update Solid (Step 3). We advance the rigid body by Δt\Delta t using its current linear and angular velocities and refresh the solid–fluid information carried by the 3D grid (solid SDF samples, solid velocity, and cell-type tags).

(4) Prepare Targets (Steps 4–5). We build ηtarget\eta^{\mathrm{target}} from ϕ~3Dn+1\tilde{\phi}_{3D}^{n+1} by the column-wise scan: sweep each column upward until nair=3n_{\mathrm{air}} = 3 consecutive air/solid cells appear, then read the surface from ϕ~3Dn+1\tilde{\phi}_{3D}^{n+1} at the cell nairn_{\mathrm{air}} steps below ((30), Section 5.1), followed by FAB-band padding and Laplace smoothing.

(5) Advance 2D Simulation (Step 6). We advance the canonical wave state (η,ψ)(\eta,\psi) under the Zakharov Hamiltonian via the integrating-factor AB2 scheme (15)–(16) (Section 3), producing the post-advance state (η~2Dn+1,ψ~2Dn+1)(\tilde{\eta}_{2D}^{n+1},\tilde{\psi}_{2D}^{n+1}).

(6) Forward DNO Map (Step 7). We evaluate the post-advance surface vertical velocity v=G(ψ~2Dn+1)v = G(\tilde{\psi}_{2D}^{n+1}) via the forward DNO map (18) (Section 3.5).

(7) Apply Relaxation (Step 8). We nudge η~2Dn+1\tilde{\eta}_{2D}^{n+1} toward ηtarget\eta^{\mathrm{target}} via the exact exponential decay (34) on η\eta alone, pointwise (Section 5.3), producing the final height field η2Dn+1\eta_{2D}^{n+1}.

(8) Reconstitute ψ\psi (Step 9). We recompute ψ2Dn+1\psi_{2D}^{n+1} from (η2Dn+1,v)(\eta_{2D}^{n+1},v) through the inverse Dirichlet–Neumann map so that the canonical pair (η2Dn+1,ψ2Dn+1)(\eta_{2D}^{n+1},\psi_{2D}^{n+1}) is consistent with the relaxed height (Section 5.3).

(9) Embed FAB Height Field (Step 10). We set the level-set values of the 3D solver in the FAB ring from the post-relax 2D height field η2Dn+1\eta_{2D}^{n+1} via ϕ=y−η\phi= y - \eta (see Section 4.1); the interior of ϕ3D\phi_{3D} is left untouched.

(10) Compute Volume Velocity from 2D (Step 11). We reconstruct the bulk velocity at each depth layer from the canonical 2D state using the Craig–Sulem lift to y=0y = 0 and the finite-depth harmonic kernels, as described in Sections 4.2–4.4 and summarized in Alg. 1.

(11) Interpolate Velocity in FAB from Depth Layers (Step 12). We trilinearly interpolate the velocity within the FAB ring from the depth-layered velocities.

(12) Advance 3D (Step 13). We follow standard NB-FLIP procedures to advance the 3D simulation, using the FAB-ring velocities as boundary conditions during pressure projection.

Validation

Experimental Setup. All experiments are conducted on a desktop with an NVIDIA GeForce RTX 4090. Implementation details of the 2D and 3D solvers are presented in Section 6. Relevant parameters and runtime performance are summarized in Table 5. In the wave-propagation benchmark and the Stokes-wave accuracy study we report both HOS-2 and HOS-3, while all other experiments use HOS-2, since this level is sufficient in the vast majority of our scenes and Order-2 is roughly 2×2\times faster than Order-3 per step (Table 3).

Comparisons and Validation Studies

Wave Propagation. To evaluate the accuracy of our wave model and our padding strategy, we simulated wave propagation in a square 20 m×\times20 m tank under shallow (1 m) and deep (20 m) water conditions with reflective boundary conditions on all four sides.

Algorithm 2 One substep of our coupled solver

  1. Δt←\Delta t \leftarrow ComputeTimeStep() &nbsp;&nbsp;&nbsp;&nbsp;▹ outer 3D CFL

  2. ϕ~3Dn+1,u~3Dn+1←\tilde{\phi}_{3D}^{n+1}, \tilde{\mathbf{u}}_{3D}^{n+1} \leftarrow Advect3D(Δt,u3Dn)(\Delta t, \mathbf{u}_{3D}^{n})

  3. Sn+1←S^{n+1} \leftarrow UpdateSolid(Δt,Sn)(\Delta t, S^{n}) &nbsp;&nbsp;&nbsp;&nbsp;▹ rigid-body pose

  4. η~samp3D←\tilde{\eta}_{\mathrm{samp3D}} \leftarrow Sample3DHeight(ϕ~3Dn+1)(\tilde{\phi}_{3D}^{n+1}) &nbsp;&nbsp;&nbsp;&nbsp;▹ Sec. 5.1

  5. ηtarget←\eta^{\mathrm{target}} \leftarrow Smooth(η~samp3D)(\tilde{\eta}_{\mathrm{samp3D}}) &nbsp;&nbsp;&nbsp;&nbsp;▹ Laplace smoothing

  6. η~2Dn+1,ψ~2Dn+1←\tilde{\eta}_{2D}^{n+1}, \tilde{\psi}_{2D}^{n+1} \leftarrow Advance2D(Δt,η2Dn,ψ2Dn)(\Delta t, \eta_{2D}^{n}, \psi_{2D}^{n}) &nbsp;&nbsp;&nbsp;&nbsp;▹ Sec. 3

  7. v←v \leftarrow ForwardDNO(η~2Dn+1,ψ~2Dn+1)(\tilde{\eta}_{2D}^{n+1}, \tilde{\psi}_{2D}^{n+1}) &nbsp;&nbsp;&nbsp;&nbsp;▹ forward DNO, (18)

  8. η2Dn+1←\eta_{2D}^{n+1} \leftarrow ApplyRelaxation(Δt,η~2Dn+1,ηtarget)(\Delta t, \tilde{\eta}_{2D}^{n+1}, \eta^{\mathrm{target}}) &nbsp;&nbsp;&nbsp;&nbsp;▹ Sec. 5.3

  9. ψ2Dn+1←\psi_{2D}^{n+1} \leftarrow ReconstitutePsi(η~2Dn+1,v)(\tilde{\eta}_{2D}^{n+1}, v) &nbsp;&nbsp;&nbsp;&nbsp;▹ inverse DNO, (19)

  10. ϕ3Dn+1←\phi_{3D}^{n+1} \leftarrow EmbedFABHeight(η2Dn+1,ϕ~3Dn+1)(\eta_{2D}^{n+1}, \tilde{\phi}_{3D}^{n+1}) &nbsp;&nbsp;&nbsp;&nbsp;▹ ϕ=y−η\phi= y - \eta (FAB)

  11. udepthn+1←\mathbf{u}_{\mathrm{depth}}^{n+1} \leftarrow ComputeVolumeVelFrom2D(η2Dn+1,ψ2Dn+1)(\eta_{2D}^{n+1}, \psi_{2D}^{n+1}) &nbsp;&nbsp;&nbsp;&nbsp;▹ Alg. 1

  12. uFAB←\mathbf{u}_{\mathrm{FAB}} \leftarrow InterpDepthLayerFAB(udepthn+1)(\mathbf{u}_{\mathrm{depth}}^{n+1})

  13. u3Dn+1←\mathbf{u}_{3D}^{n+1} \leftarrow Advance3D(Δt,uFAB,u~3Dn+1,ϕ3Dn+1)(\Delta t, \mathbf{u}_{\mathrm{FAB}}, \tilde{\mathbf{u}}_{3D}^{n+1}, \phi_{3D}^{n+1})

In both cases the surface is initialized at rest (v=0v = 0) with a radially symmetric Gaussian height bump centered at the tank center, η0(x,z)=Aexp⁡(−r2/2σ2)\eta_0(x,z) = A\exp(-r^2/2\sigma^2) for r<Rr < R and zero otherwise, where r=(x−L/2)2+(z−L/2)2r = \sqrt{(x-L/2)^2 + (z-L/2)^2}, with width σ=1\sigma= 1 m and truncation radius R=6R = 6 m; the peak amplitude is A=1A = 1 m in the shallow setting and A=2A = 2 m in the deep setting. Each configuration is propagated for 2000 frames at Δt=0.01\Delta t = 0.01 s. As shown in Figure 13, we compared a total of four 2D wave models, i.e., the SWE [14] and the BEM [18] used in previous coupling methods [15, 31, 72], the Airy Dispersion Kernel (DK) model [10], and our nonlinear wave model, against a high-resolution 3D Eulerian level-set simulation that serves as the ground-truth reference. We found that our nonlinear wave model produces results most consistent with the 3D reference, and correctly captures wave reflection.

Wave propagation comparison across shallow and deep water conditions

Figure 13. Wave propagation comparison in a 20-meter-wide tank under shallow (depth = 1 m) and deep (depth = 20 m) water conditions. For each depth, the upper row shows the rendered free surface produced by SWE [14], BEM [18], Airy DK [10], our nonlinear wave model (ε=1\varepsilon= 1), and a high-resolution 3D Eulerian level-set simulation that serves as the ground-truth reference; the lower row shows the corresponding per-cell absolute height error against this 3D reference. Our wave model agrees most closely with the 3D reference in both regimes.

We quantify these differences in Table 4. For a fair comparison, every method’s output is resampled onto the same 128×128128 \times128 height field before being compared with the reference; for BEM, which represents the surface as a mesh, we cast vertical rays downward from this 128×128128 \times128 grid and take the first ray–mesh intersection as the per-cell height. For each frame we compute the absolute height error against the reference and reduce it to its spatial mean and max; we then average each of these quantities over the 2000 frames of the simulation, yielding the time-averaged mean and max errors reported in the table. Compared with SWE and BEM, our model reduces the time-averaged mean error by roughly 4×4\times in the shallow-water case and by 33–5×5\times in the deep-water case; even compared with the strongest baseline (Airy DK), it still cuts the mean error by about 2.4×2.4\times in the shallow-water case and 1.7×1.7\times in the deep-water case, with similar improvements on the max error. We also tested a strengthened Airy DK baseline that multiplies the dispersion relation by the amplitude-dependent factor (1+a2k2)(1+a^2k^2). The correction improves the Airy DK baseline (row “Airy DK + dispersion fix” in Table 4), lowering its mean error by roughly 10%10\% in both regimes, but the corrected baseline still incurs about 1.51.5–2.1×2.1\times our mean error. Runtime comparisons among SWE, BEM, Airy DK, and our Order 2 and 3 models are listed in Table 3; our wave model runs about 1000×1000\times faster than BEM (though it could be further optimized). While

MethodDeep Water Time (ms)Shallow Water Time (ms)
BEM8693.725973.8
SWE1.3551.375
Airy DK1.0811.157
Ours (Order 2)6.3526.167
Ours (Order 3)13.5213.43

Table 3. Runtime comparison of the wave propagation experiment.

MethodErrorShallowDeep
SWE [14]Mean0.0482020.096199
SWE [14]Max0.2109020.422074
BEM [18]Mean0.0517200.170493
BEM [18]Max0.2345080.466547
Airy DK [10]Mean0.0276660.054496
Airy DK [10]Max0.1261020.298801
Airy DK + dispersion fixMean0.0246550.049221
Airy DK + dispersion fixMax0.1139560.280982
Ours (Order 2)Mean0.0136070.032774
Ours (Order 2)Max0.0574690.174430
Ours (Order 3)Mean0.0115010.034247
Ours (Order 3)Max0.0518430.161604

Table 4. Wave height error comparison of the wave propagation experiment. For each method, the first row is the mean error and the second row is the max error, each obtained by computing the spatial mean (resp. max) of the absolute height error on a 128×128128 \times128 grid at every frame and then averaging over the 2000 frames of the simulation.

our model is more expensive than the linear-wave-based SWE and Airy DK solvers, it delivers the accuracy improvements reported in Table 4 and remains substantially cheaper than BEM. The 2D grid spacings used were Δx=L128,L128,L128\Delta x = \frac{L}{128}, \frac{L}{128}, \frac{L}{128}, and L64\frac{L}{64} for the SWE, Airy DK, our wave model, and BEM, respectively, matching the grid spacing ratios used in Huang et al. [31], while the 3D Eulerian reference uses a much finer grid spacing of Δx=L256\Delta x = \frac{L}{256}.

To further probe the role of the nonlinear correction, we sweep the nonlinearity parameter ϵ\epsilon from 0 (linear Airy) to 1 (fully nonlinear HOS) in the shallow-water setting, using both the Order 2 and Order 3 HOS truncations, and measure the wave-height error against the same 3D reference. As shown in Figure 14, the error decreases monotonically as ϵ\epsilon grows, confirming that the higher-order HOS terms are correctly oriented and pull the solver toward the 3D reference rather than away from it. The drop between ϵ=0\epsilon= 0 and ϵ=0.1\epsilon= 0.1 is already substantial, indicating that even a mild amount of nonlinearity removes a large fraction of the linear-model error. Order 2 and Order 3 are nearly indistinguishable for ϵ≲0.6\epsilon\lesssim0.6, but the two curves separate beyond this point, with Order 3 attaining a noticeably lower error at ϵ=1\epsilon= 1.

Stokes Wave Accuracy under Varying Nonlinearity. We assess the accuracy of our HOS wave solver against the analytical second-order Stokes wave (with the third-order frequency correction) on a periodic 1D domain of length L=2πL = 2\pi through two complementary tests. Explicitly, the reference is the classical second-order Stokes profile [?],

ηSt(x,t)=acos⁡θ+ka2(3−tanh⁡2(kh))4tanh⁡3(kh)cos⁡2θ\eta_{\mathrm{St}}(x,t) = a \cos\theta+ \frac{ka^{2}\left(3-\tanh^{2}(kh)\right)}{4\tanh^{3}(kh)}\cos2\theta

with θ=kx−ωt\theta= kx - \omega t, whose frequency follows the third-order Stokes dispersion relation [80]

ω2=gktanh⁡(kh)(1+δ)\omega^{2} = gk\tanh(kh)(1+\delta)

with

δ=(ka)2(9−10tanh⁡2(kh)+9tanh⁡4(kh))8tanh⁡4(kh).\delta= \frac{(ka)^{2}\left(9-10\tanh^{2}(kh)+9\tanh^{4}(kh)\right)}{8\tanh^{4}(kh)}.

The correction δ\delta is the amplitude-dependent Stokes frequency shift, and dropping it recovers the linear relation ω2=gktanh⁡(kh)\omega^{2} = gk\tanh(kh).

Long-time phase fidelity. We initialize four mild Stokes regimes, i.e., weakly nonlinear, non-breaking Stokes waves, with (ka,kh)∈{(0.08,1.0),(0.10,1.5),(0.12,1.8),(0.13,2.0)}(ka,kh) \in\{(0.08,1.0), (0.10,1.5), (0.12,1.8), (0.13,2.0)\} and propagate each for 18, 14, 12, and 10 linear periods, respectively. Figure 2 overlays the final-time surface elevation produced by the linear (ε=0\varepsilon= 0), intermediate (ε=0.5\varepsilon= 0.5), and fully nonlinear (ε=1\varepsilon= 1, HOS-3) solvers with the analytical Stokes profile. The linear solver (equivalent to [10]) progressively drifts out of phase across all four regimes because it lacks the amplitude-dependent Stokes frequency shift defined above, while the fully nonlinear HOS-3 solver remains phase-locked to the analytical curve. The intermediate ε=0.5\varepsilon= 0.5 curve lies between the two, but closer to the linear one.

Figure 2. Final-time surface elevation η(x)\eta(x) of the linear wave solver (ε=0\varepsilon= 0, blue, equivalent to [10]), our nonlinear High-Order Spectral (HOS)-3 solver with ε=0.5\varepsilon= 0.5 (red) and ε=1\varepsilon= 1 (green), and the analytical second-order Stokes wave (black dashed) in four mild Stokes regimes, i.e., weakly nonlinear, non-breaking Stokes waves. The linear solver drifts out of phase due to the missing amplitude-dependent frequency shift, while the fully nonlinear (ε=1\varepsilon= 1) HOS-3 solver stays phase-locked to the analytical wave. The ε=0.5\varepsilon= 0.5 curve lies between the two.

Error sweep over ε\varepsilon. To examine how the nonlinearity parameter ε\varepsilon controls accuracy, we sweep ε∈{0,0.1,…,1.0}\varepsilon\in\{0,0.1,\ldots,1.0\} across two HOS truncation orders (Order 2 and Order 3) and two distinct initial conditions with (ka,kh)∈{(0.08,1.0),(0.20,2.0)}(ka,kh) \in\{(0.08,1.0),(0.20,2.0)\}. We propagate each configuration for ten linear periods and compare against the analytical Stokes solution frame by frame. The spatio-temporal error field ∣ηsim−ηStokes∣/a|\eta_{\mathrm{sim}}-\eta_{\mathrm{Stokes}}|/a is reduced to its mean and max, as shown in Figure 15. The error decreases monotonically as ε\varepsilon grows from 0 to 1, confirming that the higher-order DNO terms steer the linear solver toward the analytical Stokes wave. For the milder initial condition, Order 2 and Order 3 are nearly indistinguishable, indicating that the second-order kinematic boundary condition already captures the relevant nonlinear physics; for the steeper case, Order 3 noticeably outperforms Order 2, justifying the inclusion of the third-order term G2G_{2} when handling steep waves.

Continuous Wavetrain Across the Coupling Interface. To stress-test the 2D–3D coupling under sustained periodic forcing, we drive a steady second-order Stokes wavetrain through an embedded Eulerian fluid box, as shown in Figure 3. A 1D wave domain of length 3 m and still-water depth 0.4 m hosts a 2 m×\times0.8 m fluid box centered at the origin. A wavemaker zone at the left edge of the wave domain softly relaxes (η,ψ)(\eta,\psi) toward the analytical Stokes solution with steepness ka=0.3ka = 0.3 and wavelength 0.8 m, generating a right-going train; the entire right side of the wave domain acts as a sponge that absorbs outgoing waves before they wrap around the periodic boundary. The fluid box is initialized with still water and exchanges (η,ψ)(\eta,\psi) with the 1D solver through the relaxation-zone protocol.

Figure 3. Continuous Stokes wavetrain crossing the 2D–3D coupling interface, comparing the linear wave model (ε=0\varepsilon= 0, top row of each column) with our fully nonlinear HOS model (ε=1\varepsilon= 1, bottom row). A 1D wave domain (x∈[−2,1]x \in[-2, 1] m, still-water depth 0.4 m) drives a second-order Stokes train of steepness ka=0.3ka = 0.3 and wavelength 0.8 m through an embedded 2D fluid box (x∈[−1,1]x \in[-1, 1] m, light blue fill); yellow shading marks the left-edge wavemaker zone. The red boxes outline the wave free-propagation region just outside the fluid box, where the 1D wave solver carries the train on its own. Inside the red boxes the linear surface develops visible kinks, is no longer smooth, and shows no crest–trough asymmetry, while the nonlinear surface remains smooth, exhibits the sharpened crests and flattened troughs of a Stokes train, and continues the surface shape inside the 2D fluid box almost exactly. The mismatch in the linear case reflects the incompatibility between the Airy wave model and the natively nonlinear 2D NS solver across the coupling interface at this steepness (ka=0.3ka = 0.3).

We use this setup to contrast the linear (ε=0\varepsilon= 0, equivalent to [10]) and fully nonlinear (ε=1\varepsilon= 1) wave models when both are coupled to the same 2D Eulerian fluid box. Figure 3 shows three representative snapshots at t≈2.65t \approx2.65 s, 3.75 s, 5.0 s (about 3.7, 5.2, and 7.0 wave periods). The red boxes mark the wave free-propagation region, where the 1D wave solver carries the train on its own. With the linear model this region develops visible kinks, the surface is no longer smooth, and the Stokes crest–trough asymmetry is entirely missing. The nonlinear model, by contrast, produces a smooth Stokes train whose shape inside the red boxes is continuous with the surface shape inside the 2D fluid box, with no visible discontinuity at the coupling interface. This indicates that the natively nonlinear 2D NS solver is incompatible with the 1D linear Airy wave model at this steepness (ka=0.3ka = 0.3): only the HOS wave model can match the 2D solver’s surface dynamics across the coupling; for gentler, nearly linear wavetrains the two models would agree and the cheaper Airy model would suffice. We note that this 1D setting has no transverse geometric spreading. In a full 2D wave domain, localized outgoing disturbances generally weaken as their wavefront expands, reducing nonlinear effects with distance.

Dispersion Matching Test. The 2D and 3D simulations may exhibit distinct numerical viscosity and dispersion characteristics. As noted by [59], such differences can potentially lead. to asymmetric wave propagation and spurious reflections at the coupling boundary. To investigate this, we conducted an experiment in a 5 m water tank with a depth of 1 m, where a wave packet with a peak amplitude of 0.2 m was initialized inside the 3D domain. The results are shown in Figure 16, demonstrating strong agreement between our 2D and 3D models: the wave maintained a highly symmetric shape throughout its propagation, and no artificial reflections were observed at the 2D–3D interface.

Dispersion matching test showing wave propagation from the 3D domain into 2D regions

Figure 16. Dispersion matching test. The wave propagates from the 3D domain (bounded by dotted lines) into 2D regions. The wave retains a symmetric shape throughout its propagation, demonstrating consistent dispersion relations between the 3D and 2D simulations under our two-way coupled scheme. The wave model uses ε=1\varepsilon= 1.

To highlight the importance of dispersion matching, we compared our method against three prior coupling schemes in Figure 6: [15], [59], and FAB [[66]] combined with the Airy DK wave model [[10]]. In this combined baseline, the outer 2D wave field is advanced with the Airy DK model [[10]], the 2D-to-3D direction of the coupling follows the FAB scheme [[66]], and the 3D-to-2D direction is handled with the procedure of [15]; since the 3D-to-2D step is essentially standard and shared across most coupling methods, we label this baseline simply as FAB + Airy DK. All methods are driven by the same 3D NB-FLIP solver under identical initial conditions. The method of [15] develops pronounced left–right asymmetry. For [59], we recalibrate their dispersion-law rescaling factor for our 3D solver configuration and find that d=0.8d = 0.8 gives the best result. This largely removes the left–right asymmetry, although the 2D and 3D regions remain visibly less consistent than with our method. We note that [59] report no asymmetry under their original simulation settings. The FAB + Airy DK combination shows no visible asymmetry, but produces visibly deeper waves inside the 3D region than in the surrounding 2D field. Our method yields a wave field that is virtually identical inside and outside the 3D region, with a smoother 3D region. In terms of runtime, the wave-solver cost in this comparison is 13.3 ms per substep for our method, 27.5 ms for [59], 3.7 ms for FAB + Airy DK, and 2.3 ms for [15].

Crown Splash. A water ball with a radius of 0.05 m falls from a height of 0.2 m into a tank measuring 1.2 m wide and 0.6 m deep, as shown in Figure 8. We compare our coupling scheme against a 3D NB-FLIP reference and [15]. We apply reflected padding to the boundaries of the 2D solver to enforce solid boundaries. Our coupling scheme produces results nearly identical to the NB-FLIP ground truth, enabling smooth wave transmission across the 2D–3D interface and accurate wave propagation throughout the whole domain.

Figure 8. Crown Splash. The crown splash scenario compares Chentanez et al. [15] (top row), ours (middle row), and the 3D NB-FLIP reference (bottom row). Our method produces results that closely match the NB-FLIP reference, faithfully preserving crown splash dynamics and enabling smooth wave transmission across the 2D–3D interface. In contrast, the method of Chentanez et al. [15] exhibits clearly incorrect wave propagation in the 2D region, as the shallow water equations cannot reproduce the dispersive behavior of the deeper-water reference. Our wave model uses ε=0.8\varepsilon= 0.8.

The method of [15], on the other hand, shows clearly incorrect wave propagation in the 2D region: since it relies on the shallow water equations, it fails to capture the dispersive wave dynamics.

ψ\psi Relaxation Ablation. On the same Crown Splash setup, we ablate the asymmetric source placement of Section 5.2 by additionally relaxing ψ\psi toward a Helmholtz-decomposed target extracted from the 3D surface velocity field. As shown in Figure 17, this symmetric variant produces spurious surface artifacts inside and around the embedding box.

Filter and Time-Integrator Ablation. We further ablate two components of the wave solver. First, disabling the low-pass state filter (17) causes the Crown Splash simulation to quickly become unstable. Second, replacing the IF-AB2 integrator (Section 3.4) with a first-order integrating-factor (exponential) Euler step increases the mean/max wave height error in the shallow-water wave-propagation experiment from 0.011501/0.051843 to 0.012752/0.053980 (cf. Table 4).

Single Boat. We simulate a 4.8 m-long boat moving at 4 m/s across a 72 m×36 m×36 m water domain; Figure 19 shows a rendering of this scene produced by our method with whitewater, where the feather-like Kelvin wake (“feather-like” follows the terminology of [55]) is carried continuously across the 2D–3D interface. To assess our coupling scheme, we benchmark it against six alternatives under identical conditions, as shown in Figure 12: (a) [15], (b) the Airy DK wave model [[10]], (c) a coupled via the FAB scheme [[66]], (d) our method with velocity relaxation, (e) our wave model combined with the coupling scheme of [15], (e) our full method, (f) a 3D NB-FLIP reference, and (g) a pure 3D Eulerian level-set reference. Method (a) produces a qualitatively wrong wake, as the SWE cannot resolve the dispersive Kelvin pattern. Methods (b), (c), and (d) reproduce the feather-like shape but exhibit clear spurious reflections behind the stern (highlighted by the red boxes in Figure 12). Our full method (e) reproduces the feather-like Kelvin wake with minimal spurious reflection at the 2D–3D interface. The NB-FLIP reference (f) closely matches our result, but its trailing wave at the far right is noticeably damped by the numerical dissipation of the 3D solver, while our coupled scheme preserves the wave because the 2D wave solver carries it without dissipation. The pure Eulerian level-set simulation (g) is included as an independent ground-truth reference. As reported in Table 5, our method also runs more than 4× faster than the pure GPU NB-FLIP simulation on this scene (85.0 ms vs. 363 ms per substep). We note that the feather-like appearance is a qualitative descriptor; the wake-angle study below provides the corresponding quantitative validation of the simulated wakes (Figure 18).

CaseRes 2DRes 3DΔx2D\Delta x_{2D} [m]Δx3D\Delta x_{3D} [m]Frame CountFrame Time Step [s]Simulation SubstepsTsubstepT_{\mathrm{substep}} [ms]CFLε\varepsilon
Dispersion Matching256×256256×64×5120.0200.00782001/80113387.80.251.0
Crown Splash128×128128×256×1280.00940.00633001/8066558.41.00.8
Single Boat (NB-FLIP)/1152×96×576/0.0632001/2012733631.0/
Single Boat576×288384×96×1920.130.0632001/20107385.01.01.0
Seaplane1024×256768×128×2560.1560.0472401/203951197.21.01.0
Battleship512×256512×256×2560.2340.0472401/203408160.81.00.2
Submarine512×256512×128×2560.1560.0473601/30258499.71.00.3
Two Chasing Boats512×256512×128×2560.350.0473001/303396268.81.01.0
Pond1024×1024512×256×5120.00200.00183001/3014447326.01.00.2

Table 5. Specifications and statistics of the simulations. “Simulation Substeps” refers to 3D substeps.

Figure 18. Wake angle vs. Froude Number (Fr=U/gLFr = U/\sqrt{gL}) compared with the Kelvin–Mach Transition Model [55]. All runs use ε=1\varepsilon= 1.

Specifically, according to [55], real-world ship wakes transition between the Kelvin angle and the Mach angle as the ship speed varies. We replicate this transition by simulating ship wakes at different speeds using the same experimental setting above. The resulting wake angles, illustrated in Figure 18, match the real-world data from [55] and correctly capture the Kelvin–Mach transition.

Large-Scale 3D Demonstrations

Submarine. As shown in Figure 9, to demonstrate the capability of our coupling scheme in handling complex fluid–solid interactions, we simulate a submarine submerging below the sea surface and gradually rising up. The simulation domain spans 80 m×40 m for the sea surface, with an embedded 3D region of 24 m×6 m×12 m that includes a 10 m-long submarine moving at 4.3 m/s. As the submarine surfaces, it leaves a visible wake on the 2D ocean surface, while detailed 3D water motion is captured around the hull. To suppress nonlinear instabilities triggered by the rapidly moving hull, we cap the nonlinear amplitude parameter (see Section 3) at ε=0.3\varepsilon= 0.3.

Surfacing submarine advancing toward the ocean surface with near-view insets

Figure 9. A surfacing submarine advancing forward and slowly rising to the ocean surface. In each panel, the main image is the far view and the inset is the near view of the same moment. The wave model uses ε=0.3\varepsilon= 0.3.

Seaplane. As shown in Figure 7, we simulate a seaplane landing on the ocean. The simulation domain spans 160 m×40 m for the sea surface, with an embedded 3D region of 36 m×6 m×12 m that includes a seaplane with a 10 m wingspan, pitched 9° nose-up and approaching at 8 m/s. Upon touchdown, the aircraft skims the surface at 10 m/s, leaving a long trailing wake that propagates smoothly across the 2D–3D interface.

Seaplane landing on the ocean

Figure 7. Seaplane landing on the ocean. A seaplane with a 10 m wingspan, pitched 9∘9^\circ nose-up, approaches at 8 m/s, touches down, and skims the surface at 10 m/s, leaving a long trailing wake that transmits smoothly across the 2D–3D interface. In each panel, the main image is the far view and the inset is the near view of the same moment. The wave model uses ε=1\varepsilon= 1.

Seaplane landing on the ocean

Figure 7. Seaplane landing on the ocean. A seaplane with a 10 m wingspan, pitched 9∘9^\circ nose-up, approaches at 8 m/s, touches down, and skims the surface at 10 m/s, leaving a long trailing wake that transmits smoothly across the 2D–3D interface. In each panel, the main image is the far view and the inset is the near view of the same moment. The wave model uses ε=1\varepsilon= 1.

Battleship in Heavy Waves. To further demonstrate the robustness of our coupling scheme, we simulate heavy ocean waves in a 120 m×60 m 2D domain and a 24 m×12 m×12 m 3D fluid domain. In this scenario, the battleship is simulated using weak two-way solid–fluid coupling. The results in Figure 11 show that our simulation remains stable even with heavy waves, and the 2D and 3D simulations remain visually consistent. Because the background spectrum drives the surface close to the breaking threshold, we further restrict the nonlinear amplitude parameter (see Section 3) to ε=0.2\varepsilon= 0.2 to keep the wave evolution stable. This scene is our most demanding stress test, and the coupling is not entirely artifact-free: in the accompanying animation, wavefronts reflected from the rectangular boundary of the 3D box can be seen propagating outward into the 2D domain. Because the wave model is restricted to ε=0.2\varepsilon= 0.2 for stability, it cannot fully absorb the strongly nonlinear surface content generated by the 3D solver under near-breaking forcing; the unabsorbed content instead radiates outward from the box boundary. Section 6 discusses practical mitigation strategies at render time, i.e., strengthening the ocean-spectrum displacement layer and emitting foam and whitewater around the box footprint.

A 4.8 m-long boat cruising across a water domain and producing a feather-like Kelvin wake

Figure 19. A 4.8 m-long boat cruising at 4 m/s across a 72 m×3 m×36 m water domain. The boat produces a clean feather-like Kelvin wake [55] whose long dispersive tail is carried far beyond the 3D fluid box by the 2D wave solver without visible discontinuity at the 2D–3D interface. The wave model uses ε=1\varepsilon= 1.

Two Chasing Boats. As shown in Figure 10, two boats with two independent 3D simulation domains are simulated on an ocean surface sized 180 m×90 m. The leading boat follows a curved path with one-way solid–fluid coupling. The trailing boat employs two-way solid–fluid coupling, responding dynamically to both the ambient waves and the wake generated by the leading boat.

Two chasing boats on the ocean surface, producing long trailing wakes

Figure 10. Two chasing boats on the ocean surface, producing long trailing wakes as they move through the water. The four images on the left show the near view, while the large image on the right shows the far view. The wave model uses ε=1\varepsilon= 1.

Pond. We simulate water falling into a shallow pond with a depth of 0.3 m and a 2D domain spanning 2 m×2 m, with the embedded 3D fluid box covering 0.9 m×0.45 m×0.9 m. The pond outline and the rocks scattered inside it are encoded as an irregular solid mask on the 2D wave grid, while a separate rock mesh is treated as a rigid obstacle within the 3D fluid box. Figure 20 illustrates falling flow that generates splashes and ripples that propagate across the surface and interact with the irregular solid boundaries. Before every spectral wave step, the solver fills the masked cells by iteratively averaging the height η\eta and vertical velocity vv from their fluid-side neighbors, producing a discrete zero-gradient extension of (η,v)(\eta, v) into the solid region. The subsequent FFT advance then sees this extension as replicated padding at the irregular boundary, so waves reflect off the pond walls and the rocks inside it at the cost of a few sweep iterations. Because the impact of the falling stream produces sharp local features that can excite high-wavenumber nonlinear instabilities, we cap the nonlinear amplitude parameter at ε=0.2\varepsilon= 0.2 to keep the wave evolution stable.

Water falling into a shallow pond with irregular boundaries

Figure 20. Water falling into a shallow pond with irregular boundaries. The wave model uses ε=0.2\varepsilon= 0.2.

Conclusions and Limitations

Conclusions. We have presented a two-way coupled water solver that pairs a nonlinear and dispersive 2D wave model with a localized 3D Navier–Stokes solver, addressing a long-standing structural mismatch in graphics where the 2D side of every prior coupling pipeline was either linear (Airy) or non-dispersive (SWE). Our solver makes the opposite trade-off to prior 2D wave models: whereas they buy speed and robustness by discarding nonlinearity, and remain the better choice when the surface stays gentle and nearly linear, we accept a modestly more expensive and less stable wave step in exchange for greater nonlinear accuracy and coupling fidelity that the strongly nonlinear scenes motivating a 2D–3D decomposition demand. The 2D model evolves the canonical Zakharov pair (η,ψ)(\eta,\psi) under the exact water-wave Hamiltonian, advancing the Dirichlet–Neumann operator through a homogeneous Craig–Sulem expansion so that every nonlinear correction reduces to FFTs and pointwise products on a regular grid at O(N2log⁡N)O(N^2 \log N) cost. Because the wave side and the 3D bulk solver carry the same physical scalar variable, the surface velocity potential, the coupling becomes natural: both directions of state transfer are realized as Dirichlet–Neumann maps, and a Newtonian relaxation source integrated through its exact exponential decay returns the 3D-side surface state to the canonical wave variables without damping outgoing waves. The Hamiltonian structure of the wave model thus pays off chiefly in the coupling: the canonical pair (η,ψ)(\eta,\psi) serves as a complete interface state that both solvers can read and write consistently. A scalar amplitude parameter ε∈[0,1]\varepsilon\in[0,1] continuously interpolates between the linear Airy limit and the nonlinear Zakharov system, giving the user a single dial to trade nonlinear fidelity against stability. We validate our method on a wide range of experiments, including standalone 2D wave-propagation benchmarks, controlled 2D–3D coupling tests, and a suite of large-scale production-style scenarios such as a surfacing submarine, a seaplane landing, a battleship in heavy seas, two chasing boats, and a shallow pond with irregular boundaries. Across all these cases, our coupled pipeline produces accurate results with minimal coupling artifacts at competitive runtime, suggesting that the method is practical and applicable across a broad range of scenarios within its intended nonlinear regime.

Limitations. Our solver has several limitations. First, the fully nonlinear setting (ε=1\varepsilon= 1) is not unconditionally stable: when strong 3D forcing pushes the surface close to the breaking threshold, the higher-order DNO terms can excite numerical instabilities, and we have to reduce ε\varepsilon to maintain a stable simulation. In our experiments, the submarine uses ε=0.3\varepsilon= 0.3, while the battleship in heavy waves and the pond both use ε=0.2\varepsilon= 0.2. While our ε\varepsilon sweeps (Figure 14, Figure 15 confirm that even a small ε\varepsilon already removes a large fraction of the linear-model error, the user is still left to tune ε\varepsilon per scene rather than always running at full nonlinearity. Second, the truncated DNO introduces a computational overhead that grows with the truncation order: as reported in Table 3, Order 3 is roughly 2× slower than Order 2 per step, and each additional order adds a comparable cost. More broadly, even the Order-2 wave step is roughly 5× more expensive than linear-wave solvers such as SWE and Airy DK (Table 3); in our coupled pipeline this overhead is minor in absolute terms, since the 3D solver dominates the total cost in most of our scenes (Table 5, but for standalone ocean animation where a linear model suffices, our solver is the more expensive choice. In practice, we find that Order 2 is already sufficient for the vast majority of our scenes, with Order 3 providing only marginal accuracy gains in the steepest regimes (Figure 15). Third, our 2D wave model cannot represent breaking waves: once the surface overturns, the single-valued height field η(x,t)\eta(\mathbf{x},t) ceases to be a valid representation. We also note that the wave solver advances the entire 2D grid at every substep, so unlike Tessendorf-style spectral oceans, the simulated wave domain cannot be extended indefinitely. Finally, our coupling does not fully eliminate interface artifacts in the most demanding scenarios, such as the battleship scene, and we currently rely on render-time masking to visually mitigate these residual artifacts (Section 6).

Future Work. Each limitation points to a distinct direction. To address the conditional stability of the fully nonlinear model, we would like to investigate more stable nonlinear time-integration schemes for the Zakharov system, such as Hamiltonian-preserving integrators, spectrally consistent dealiasing tailored to the higher-order DNO terms, or implicit treatments of the leading nonlinear corrections, with the goal of pushing the stable operating point of ε\varepsilon closer to 1 even in strongly forced regimes. To reduce the per-step cost of the nonlinear correction, we will explore cheaper approximations of the higher-order DNO terms, for example mixed-precision FFT pipelines, learned surrogates for Gj(η)G_j(\eta) at j≥2j \ge2, or adaptive-order schemes that locally raise the truncation order only where the surface is steep. Finally, to extend the method to breaking and overturning waves, we plan to swap out the single-valued height-field representation in regions where breaking is detected for a richer surface representation, such as a locally multi-valued surface, a particle-augmented free surface in the spirit of NB-FLIP, or a hybrid Hamiltonian/Lagrangian description that retains the spectral efficiency of our solver in the smooth bulk while gracefully degrading to a particle representation through the breaking event.

Acknowledgments

We sincerely thank the reviewers for their valuable feedback. Georgia Tech authors acknowledge NSF CAREER #2430319, IIS #2433307, OISE #2433313, IIS #2433322, ECCS #2318814, and the NVIDIA Academic Grant for funding support. We credit the Houdini education license for video animations.

References

George Biddell Airy. 1845. Tides and waves. B. Fellowes.

References

  1. [1]George Biddell Airy. 1845. Tides and waves. B. Fellowes.
  2. [2]Ryoichi Ando and Christopher Batty. 2020. A practical octree liquid simulator with adaptive surface resolution. ACM Trans. Graph. 39, 4, Article 32 (Aug. 2020), 17 pages. doi:10.1145/3386569.3392460DOI
  3. [3]Roland Angst, Nils Thuerey, Mario Botsch, and Markus Gross. 2008. Robust and efficient wave simulations on deforming meshes. In Computer Graphics Forum, Vol. 27. Wiley Online Library, 1895–1900.DOI
  4. [4]William JD Ataman, Chris Swan, and Paul H Taylor. 2001. On the efficient numerical simulation of directionally spread surface water waves. J. Comput. Phys. 174, 1 (2001), 277–305.
  5. [5]Jean-Pierre Berenger. 1994. A perfectly matched layer for the absorption of electromagnetic waves. Journal of computational physics 114, 2 (1994), 185–200.
  6. [6]Garrett Birkhoff. 2015. Hydrodynamics. Princeton University Press.
  7. [7]Morten Bojsen-Hansen and Chris Wojtan. 2016. Generalized non-reflecting boundaries for fluid re-simulation. ACM Transactions on Graphics (TOG) 35, 4 (2016), 1–7.
  8. [8]Jeremiah U Brackbill, Douglas B Kothe, and Hans M Ruppel. 1988. FLIP: a low-dissipation, particle-in-cell method for fluid flow. Computer Physics Communications 48, 1 (1988), 25–38.
  9. [9]Robert Bridson. 2015. Fluid simulation for computer graphics. AK Peters/CRC Press.
  10. [10]José A Canabal, David Miraut, Nils Thuerey, Theodore Kim, Javier Portilla, and Miguel A Otaduy. 2016. Dispersion kernels for water wave simulation. ACM Transactions on Graphics (TOG) 35, 6 (2016), 1–10.
  11. [11]Duowen Chen, Zhiqi Li, Taiyuan Zhang, Jinjin He, Junwei Zhou, Bart G van Bloemen Waanders, and Bo Zhu. 2025a. Fluid simulation on compressible flow maps. ACM Transactions on Graphics (TOG) 44, 4 (2025), 1–17.
  12. [12]Duowen Chen, Zhiqi Li, Junwei Zhou, Fan Feng, Tao Du, and Bo Zhu. 2024. Solid-Fluid Interaction on Particle Flow Maps. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–20.arxiv.org/abs/2409.09225
  13. [13]Duowen Chen, Junwei Zhou, and Bo Zhu. 2025b. A neural particle level set method for dynamic interface tracking. ACM Transactions on Graphics 44, 3 (2025), 1–21.
  14. [14]Nuttapong Chentanez and Matthias Müller. 2010. Real-time Simulation of Large Bodies of Water with Small Scale Details. In Symposium on Computer Animation. 197–206.
  15. [15]Nuttapong Chentanez, Matthias Müller, and Tae-Yong Kim. 2015. Coupling 3D Eulerian, heightfield and particle methods for interactive simulation of large scale liquid phenomena. IEEE transactions on visualization and computer graphics 21, 10 (2015), 1116–1128.
  16. [16]Walter Craig and Catherine Sulem. 1993. Numerical simulation of gravity waves. J. Comput. Phys. 108, 1 (1993), 73–83.
  17. [17]Gordon David Crapper. 1984. Introduction to water waves. Cambridge University Press.
  18. [18]Fang Da, David Hahn, Christopher Batty, Chris Wojtan, and Eitan Grinspun. 2016. Surface-only liquids. ACM Transactions on Graphics (TOG) 35, 4 (2016), 1–12.DOI
  19. [19]Yitong Deng, Hong-Xing Yu, Diyan Zhang, Jiajun Wu, and Bo Zhu. 2023. Fluid Simulation on Neural Flow Maps. ACM Transactions on Graphics (TOG) 42, 6 (2023), 1–21.
  20. [20]Douglas G Dommermuth and Dick KP Yue. 1987. A high-order spectral method for the study of nonlinear gravity waves. Journal of Fluid Mechanics 184 (1987), 267–288.
  21. [21]Guillaume Ducrozet, Félicien Bonnefoy, David Le Touzé, and Pierre Ferrant. 2016. HOS-ocean: Open-source solver for nonlinear waves in open ocean based on High-Order Spectral method. Computer Physics Communications 203 (2016), 245–254.
  22. [22]Florian Ferstl, Ryoichi Ando, Chris Wojtan, Rüdiger Westermann, and Nils Thuerey. 2016. Narrow band FLIP for liquid simulations. In Computer Graphics Forum, Vol. 35. Wiley Online Library, 225–232.DOI
  23. [23]Nick Foster and Ronald Fedkiw. 2001. Practical animation of liquids. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques. 23–30.
  24. [24]Nick Foster and Dimitri Metaxas. 1996. Realistic animation of liquids. Graphical models and image processing 58, 5 (1996), 471–483.
  25. [25]Ming Gao, Nathan Mitchell, and Etchyo Fisakis. 2014. Steklov-Poincaré Skinning. In Symposium on Computer Animation. 139–148.DOI
  26. [26]Philippe Guyenne and David P Nicholls. 2008. A high-order spectral method for nonlinear water waves over moving bottom topography. SIAM Journal on Scientific Computing 30, 1 (2008), 81–101.
  27. [27]Francis H Harlow. 1964. The particle-in-cell computing method for fluid dynamics. Methods Comput. Phys. 3 (1964), 319–343.
  28. [28]Jinjin He, Taiyuan Zhang, Zhiqi Li, Junwei Zhou, Duowen Chen, and Bo Zhu. 2026. A Level Set Method on Particle Flow Maps. arXiv preprint arXiv:2601.09939 (2026).arxiv.org/abs/2601.09939
  29. [29]Christopher J Horvath. 2015. Empirical directional wave spectra for computer graphics. In Proceedings of the 2015 Symposium on Digital Production. 29–39.
  30. [30]Libo Huang and Dominik L Michels. 2020. Surface-only ferrofluids. ACM Transactions on Graphics (TOG) 39, 6 (2020), 1–17.DOI
  31. [31]Libo Huang, Ziyin Qu, Yun Tan, Xinxin Zhang, Dominik L Michels, and Chenfanfu Jiang. 2021. Pipes, splashes, and waves on a vast ocean. ACM Transactions on Graphics (TOG) 40, 6 (2021), 1–15.
  32. [32]Markus Ihmsen, Jens Orthmann, Barbara Solenthaler, Andreas Kolb, and Matthias Teschner. 2014. SPH fluids in computer graphics. States of the art reports (2014).
  33. [33]Stefan Jeschke, Tomáš Škřivan, Matthias Müller-Fischer, Nuttapong Chentanez, Miles Macklin, and Chris Wojtan. 2018. Water surface wavelets. ACM Transactions on Graphics (TOG) 37, 4 (2018), 1–13.
  34. [34]Stefan Jeschke and Chris Wojtan. 2015. Water wave animation via wavefront parameter interpolation. ACM Transactions on Graphics (TOG) 34, 3 (2015), 1–14.DOI
  35. [35]Stefan Jeschke and Chris Wojtan. 2017. Water wave packets. ACM Transactions on Graphics (TOG) 36, 4 (2017), 1–12.
  36. [36]Stefan Jeschke and Chris Wojtan. 2023. Generalizing shallow water simulations with dispersive surface waves. ACM Transactions on Graphics (TOG) 42, 4 (2023), 1–12.DOI
  37. [37]Michael Kass and Gavin Miller. 1990. Rapid, stable fluid dynamics for computer graphics. In Proceedings of the 17th annual conference on Computer graphics and interactive techniques. 49–57.
  38. [38]Anita T Layton and Michiel van de Panne. 2002. A numerically efficient and stable algorithm for animating water waves. The Visual Computer 18, 1 (2002), 41–53.DOI
  39. [39]Bernard Le Méhauté. 2013. An introduction to hydrodynamics and water waves. Springer Science & Business Media.
  40. [40]Richard Lee and Carol O’Sullivan. 2007. A Fast and Compact Solver for the Shallow Water Equations. In VRIPHYS. 51–57.
  41. [41]Zhiqi Li, Duowen Chen, Candong Lin, Jinyuan Liu, and Bo Zhu. 2024. Particle-Laden Fluid on Flow Maps. ACM Trans. Graph. 43, 6, Article 266 (Nov. 2024), 12 pages. doi:10.1145/3687916DOI
  42. [42]Zhiqi Li, Duowen Chen, Junwei Zhou, Sinan Wang, Yuchen Sun, and Bo Zhu. 2026. An Impulse-formed Navier-Stokes Solver based on Long-range Particle Flow Maps. arXiv preprint arXiv:2602.00499 (2026).arxiv.org/abs/2602.00499
  43. [43]Zhiqi Li, Candong Lin, Duowen Chen, Xinyi Zhou, Shiying Xiong, and Bo Zhu. 2025a. Clebsch gauge fluid on particle flow maps. ACM Transactions on Graphics (TOG) 44, 4 (2025), 1–12.DOI
  44. [45]Frank Losasso, Ronald Fedkiw, and Stanley Osher. 2006. Spatially adaptive techniques for level set methods and incompressible flow. Computers & Fluids 35, 10 (2006), 995–1010.
  45. [47]Miles Macklin. 2022. Warp: A High-performance Python Framework for GPU Simulation and Graphics. https://github.com/nvidia/warp. NVIDIA GPU Technology Conference (GTC).
  46. [48]Matthias Müller, David Charypar, and Markus Gross. 2003. Particle-based fluid simulation for interactive applications. In Proceedings of the 2003 ACM SIGGRAPH/Eurographics symposium on Computer animation. Citeseer, 154–159.
  47. [49]Mohammad Sina Nabizadeh, Ritoban Roy-Chowdhury, Hang Yin, Ravi Ramamoorthi, and Albert Chern. 2024. Fluid Implicit Particles on Coadjoint Orbits. ACM Transactions on Graphics (TOG) 43, 6 (2024), 270:1–270:38.DOI
  48. [50]Mohammad Sina Nabizadeh, Stefan Heidrich, and Albert Chern. 2022. Covector fluids. ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–16.
  49. [51]Michael B Nielsen and Robert Bridson. 2011. Guide shapes for high resolution naturalistic liquid simulation. In ACM SIGGRAPH 2011 papers. 1–8.
  50. [52]Michael B. Nielsen and Robert Bridson. 2016. Spatially adaptive FLIP fluid simulations in bifrost. In ACM SIGGRAPH 2016 Talks (Anaheim, California) (SIGGRAPH ’16). Association for Computing Machinery, New York, NY, USA, Article 41, 2 pages. doi:10.1145/2987398.2927399DOI
  51. [53]Michael B Nielsen, Andreas Söderström, and Robert Bridson. 2013. Synthesizing waves from animated height fields. ACM Transactions on Graphics (TOG) 32, 1 (2013), 1–9.DOI
  52. [54]Steven A Orszag. 1971. On the elimination of aliasing in finite-difference schemes by filtering high-wavenumber components. Journal of the Atmospheric Sciences 28, 6 (1971), 1074–1074.doi.org/10.1175/1520-0469(1971)028<1074:oteoai>2.0.co;2
  53. [55]Marc Rabaud and Frédéric Moisy. 2013. Ship wakes: Kelvin or Mach angle? Physical review letters 110, 21 (2013), 214503.
  54. [56]Takahiro Sato, Christopher Wojtan, Nils Thuerey, Takeo Igarashi, and Ryoichi Ando. 2018. Extended narrow band FLIP for liquid simulations. In Computer Graphics Forum, Vol. 37. Wiley Online Library, 169–177.DOI
  55. [57]Rohan Sawhney and Keenan Crane. 2017. Boundary first flattening. ACM Transactions on Graphics (ToG) 37, 1 (2017), 1–14.DOI
  56. [58]Camille Schreck, Christian Hafner, and Chris Wojtan. 2019. Fundamental solutions for water wave animation. ACM Transactions on Graphics (TOG) 38, 4 (2019), 1–14.DOI
  57. [59]Camille Schreck and Chris Wojtan. 2022. Coupling 3D liquid simulation with 2D wave propagation for large scale water surface animation using the equivalent sources method. In Computer Graphics Forum, Vol. 41. Wiley Online Library, 343–353.
  58. [60]Han Shao, Libo Huang, and Dominik L Micles. 2022. A fast unsmoothed aggregation algebraic multigrid framework for the large-scale simulation of incompressible flow. ACM Transactions on Graphics (TOG) 41, 4 (2022), 1–18.
  59. [61]Tomas Skrivan, Andreas Soderstrom, John Johansson, Christoph Sprenger, Ken Museth, and Chris Wojtan. 2020. Wave curves: simulating larger water waves with dynamically deforming surfaces. ACM Trans. Graph. 39, 4, Article 65 (Aug. 2020), 12 pages. doi:10.1145/3386569.3392466DOI
  60. [62]Andreas Söderström, Matts Karlsson, and Ken Museth. 2010. A PML-based nonreflective boundary for free surface fluid animation. ACM Transactions on Graphics (TOG) 29, 5 (2010), 1–17.DOI
  61. [63]Jos Stam. 1999. Stable Fluids. In Proceedings of SIGGRAPH 1999. 121–128.
  62. [64]Jos Stam. 2001. A simple fluid solver based on the FFT. Journal of Graphics Tools 6, 2 (2001), 43–52.
  63. [65]James Johnston Stoker. 2019. Water waves: The mathematical theory with applications. Courier Dover Publications.
  64. [66]Alexey Stokman and Andrew Selle. 2017. Fluxed animated boundary method. ACM Transactions on Graphics (TOG) 36, 4 (2017), 1–8.
  65. [67]Yuchen Sun, Linglai Chen, Weiyuan Zeng, Tao Du, Shiying Xiong, and Bo Zhu. 2024. An Impulse GloFuid Method for Simulating Two-Phase Flows. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–12.
  66. [68]Yuchen Sun, Junlin Li, Ruicheng Wang, Sinan Wang, Zhiqi Li, Bart G. van Bloemen Waanders, and Bo Zhu. 2025. Leapfrog Flow Maps for Real-Time Fluid Simulation. ACM Transactions on Graphics (TOG) 44, 4 (2025), 1–12.
  67. [69]Jerry Tessendorf. 2001. Simulating ocean water. Simulating nature: realistic and interactive techniques. SIGGRAPH 1, 2 (2001), 5.
  68. [70]Jerry Tessendorf. 2014. ewave: Using an exponential solver on the iwave problem. Technical Note (2014).
  69. [71]Nils Thuerey, Matthias Muller-Fischer, Simon Schirm, and Markus Gross. 2007. Real-time breaking waves for shallow water simulations. In 15th Pacific Conference on Computer Graphics and Applications (PG’07). IEEE, 39–46.
  70. [72]Nils Thuerey, Ulrich Rüde, and Marc Stamminger. 2006. Animation of open water phenomena with coupled shallow water and free surface simulations. In Proceedings of the 2006 ACM SIGGRAPH/Eurographics Symposium on Computer Animation (Vienna, Austria) (SCA ’06). Eurographics Association, Goslar, DEU, 157–164.
  71. [73]Hui Wang, Jinjin He, and Xubo Yang. 2026. Two-phase Impulse Fluid on Particle Flow Map. IEEE Transactions on Visualization and Computer Graphics (2026).
  72. [74]Huamin Wang, Gavin Miller, and Greg Turk. 2007. Solving general shallow wave equations on surfaces. In Proceedings of the 2007 ACM SIGGRAPH/Eurographics symposium on Computer animation. 229–238.
  73. [76]Sinan Wang, Yitong Deng, Molin Deng, Hong-Xing Yu, Junwei Zhou, Duowen Chen, Taku Komura, Jiajun Wu, and Bo Zhu. 2024. An Eulerian Vortex Method on Flow Maps. ACM Transactions on Graphics (TOG) 43, 6 (2024), 1–14.
  74. [77]Sinan Wang, Junwei Zhou, Fan Feng, Zhiqi Li, Yuchen Sun, Duowen Chen, Greg Turk, and Bo Zhu. 2025b. Fluid Simulation on Vortex Particle Flow Maps. ACM Transactions on Graphics (TOG) 44, 4 (2025), 1–24.
  75. [78]Yu Wang, Mirela Ben-Chen, Iosif Polterovich, and Justin Solomon. 2018. Steklov spectral geometry for extrinsic shape analysis. ACM Transactions on Graphics (TOG) 38, 1 (2018), 1–21.
  76. [79]Bruce J West, Keith A Brueckner, Ralph S Janda, D Michael Milder, and Robert L Milton. 1987. A new numerical method for surface hydrodynamics. Journal of Geophysical Research: Oceans 92, C11 (1987), 11803–11824.
  77. [80]Gerald Beresford Whitham. 2011. Linear and nonlinear waves. John Wiley & Sons.
  78. [81]Cem Yuksel, Donald H House, and John Keyser. 2007. Wave particles. ACM Transactions on Graphics (TOG) 26, 3 (2007), 99–es.
  79. [82]Vladimir E Zakharov. 1968. Stability of periodic waves of finite amplitude on the surface of a deep fluid. Journal of Applied Mechanics and Technical Physics 9, 2 (1968), 190–194.
  80. [83]Junwei Zhou, Duowen Chen, Molin Deng, Yitong Deng, Yuchen Sun, Sinan Wang, Shiying Xiong, and Bo Zhu. 2024. Eulerian-Lagrangian Fluid Simulation on Particle Flow Maps. ACM Transactions on Graphics (TOG) 43, 4 (2024), 1–20.
  81. [84]Bo Zhu, Wenlong Lu, Matthew Cong, Byungmoon Kim, and Ronald Fedkiw. 2013. A new grid structure for domain extension. ACM Transactions on Graphics (TOG) 32, 4 (2013), 1–12.
  82. [85]Yongning Zhu and Robert Bridson. 2005. Animating sand as a fluid. ACM Transactions on Graphics (TOG) 24, 3 (2005), 965–972.

Paper details

vlnc:W_pawgbfz4
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents