💻 computer science

Residual-Guided Adaptive Pressure Bases for Dimension-Reduced FLIP Projection

This paper introduces a residual-guided adaptive pressure basis for dimension-reduced FLIP simulations that dynamically augments the coarse pressure basis in regions with high error or complex geometry, achieving lower approximation error and divergence than uniform reduction with only a modest computational overhead.

Original authors: Jong-Hyun Kim

Published 2026-10-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jong-Hyun Kim

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Creating realistic water in computer graphics has long been a battle between two opposing forces: the desire for visual detail and the limits of computer power. To simulate a fluid, researchers often use a grid, a virtual mesh that divides space into tiny boxes. Within this grid, the computer must solve a complex mathematical puzzle at every single moment of the simulation to ensure the liquid behaves like a real, incompressible substance. This puzzle, known as a pressure projection, acts as a global constraint, forcing the water to flow smoothly without tearing or compressing unnaturally. The problem is that the more detailed the grid, the larger and more expensive this puzzle becomes. If a computer tries to solve the puzzle for every single tiny box in a high-resolution scene, the calculation can take so long that the simulation grinds to a halt. To speed things up, scientists often simplify the problem by solving it on a much coarser, lower-resolution grid. However, this shortcut has a flaw: while it works well for calm, smooth water, it often fails when the water crashes, splashes, or forms thin sheets. The coarse grid simply cannot see the fine details of these chaotic moments, leading to a simulation that looks blurry or physically incorrect in the very places where the action is most intense.

A researcher at Inha University has developed a new approach that bridges this gap, allowing for fast simulations that do not sacrifice the sharpness of complex water movements. Instead of trying to solve the entire pressure puzzle at full resolution or simplifying the whole scene uniformly, their method acts like a smart spotlight. It keeps the fast, simplified calculation running across the entire scene but automatically detects when and where the water is behaving in a complicated way. When the computer notices a splash forming, a thin sheet of water stretching, or a violent impact against an obstacle, it temporarily adds extra calculation power just to that specific area. This happens without the computer needing to re-solve the entire puzzle from scratch or slow down the overall process. The system is guided by a "residual," which is essentially a measure of how much the simplified solution is struggling to match the rules of physics in a specific spot. If the simplified math leaves a gap, the system knows to fill it with more detail right there.

The researcher tested this technique in both two-dimensional and three-dimensional simulations, pitting it against standard high-resolution methods and the usual simplified approaches. They found that by adding only a small number of extra calculation points—roughly 3.5 percent more than the simplified version—they could achieve a level of accuracy that was nearly indistinguishable from the full, slow, high-resolution simulation. In scenes involving dam breaks, colliding waves, or water jets, the standard simplified method produced noticeable errors, smoothing out the splashes and missing the fine details of the water's surface. The new method, however, preserved these intricate details. It successfully captured the thin, fragile sheets of water and the sharp pressure changes near obstacles that the coarse grid usually missed. The key to this success was not just adding more detail everywhere, but adding it only where it was needed. The system used a combination of clues to decide where to focus: it looked at the errors left over from the previous moment in time, the sharpness of the water's surface curves, and how quickly the water's speed was changing in a local area.

One of the most significant aspects of this work is how efficiently it manages time. In many adaptive systems, the computer might have to pause, check the results, decide to add more detail, and then solve the problem again within the same frame of animation. This "re-solve" cycle can be very expensive. The new method avoids this by using a clever prediction. It looks at the errors from the previous moment to guess where the trouble spots will be in the next moment, allowing them to set up the extra detail before the calculation even begins. This means the computer only needs to solve the pressure puzzle once per frame, keeping the simulation fast while still capturing the complex physics. The results showed that the new approach reduced the error in the pressure calculation by about 61 percent compared to the standard simplified method, while only increasing the total computation time by a small margin. In three-dimensional tests, the method maintained this advantage, keeping the simulation roughly 56 percent faster than a full-resolution run while delivering much sharper visual results than the standard low-resolution approach.

The study confirms that the most effective way to simulate complex fluids is not to treat every part of the water the same, but to adapt the level of detail to the local behavior of the flow. By focusing computational resources only on the regions where the water is doing something difficult—like crashing into a wall or breaking into a spray—the system achieves a balance that was previously difficult to reach. The researcher noted that while the method is highly effective, it is not a perfect replacement for full-resolution simulation in every possible scenario, particularly if the entire scene becomes chaotic at once. However, for the vast majority of fluid simulations where complexity is localized, this adaptive strategy offers a powerful tool. It allows computer graphics artists and scientists to render water that looks physically accurate and visually stunning without requiring the massive computing power that was once thought necessary. The work demonstrates that with the right kind of guidance, a computer can learn to see the small, important details in a fluid world without having to look at every single drop.

Technical Summary: Residual-Guided Adaptive Pressure Bases for Dimension-Reduced FLIP Projection

Problem Statement

In Eulerian and FLIP (Fluid-Implicit-Particle) liquid simulations, enforcing incompressibility requires solving a large sparse Poisson system for pressure at every time step. As grid resolution increases, the computational cost of this pressure projection becomes a dominant factor. To mitigate this, dimension-reduced pressure projection represents the high-resolution pressure field using low-resolution variables and an interpolation basis. While surface-aware formulations (e.g., Ando et al.) successfully reduce the number of unknowns while satisfying free-surface Dirichlet conditions, they typically employ a spatially uniform reduction. This uniformity assigns the same representation capacity to smooth interior regions and to areas with complex dynamics, such as impacts, thin sheets, or rapidly varying free surfaces. Under aggressive coarsening, these localized features are often under-resolved, leading to increased pressure approximation errors and post-projection velocity divergence.

Methodology

The proposed method introduces a residual-guided adaptive pressure basis that augments a global coarse surface-aware basis only in regions predicted to require additional representation capacity. The core algorithm operates within a single reduced pressure solve per time step, avoiding iterative refinement cycles.

1. Adaptive Basis Construction

The pressure solution space is constructed by combining:

A global coarse surface-aware basis ( $U$ ) that handles the bulk of the domain.
Local fine-grid identity bases ( $U_a$ ) inserted selectively into the reduced system.

The total basis is defined as $B = \hat{\Phi}[U \ U_a]$ , where $\hat{\Phi}$ incorporates the free-surface Dirichlet boundary conditions. The pressure field is represented as $p = Bq$ , where $q$ contains coefficients for both coarse and adaptive fine variables.

2. Selection Criteria (Importance Field)

The selection of cells for adaptive refinement is driven by a composite importance field $I_i$ , which combines three indicators:

Temporal Residual Predictor ( $R_i$ ): The primary signal is the fine-grid residual of the reduced pressure solution from the previous time step. To ensure temporal coherence and avoid frame-to-frame flickering, this residual is updated using an exponential temporal filter: $\bar{r}^n = \alpha_r |r^n| + (1-\alpha_r)\bar{r}^{n-1}$ . This allows the basis to be constructed before the current solve, using historical error data.
Surface Curvature ( $K_i$ ): The magnitude of the free-surface curvature, normalized within a narrow band. This detects newly formed high-curvature structures (e.g., splashes) that may not yet have a residual history.
Velocity Variation ( $G_i$ ): The Frobenius norm of the local velocity gradient, identifying regions of strong kinematic variation.

The importance score is calculated as $I_i = w_r R_i + w_\kappa K_i + w_g G_i + S_i$ , where $S_i$ is a surface priority term ensuring a minimum refinement level at the free surface.

3. Budget Management

To prevent the reduced system from growing excessively, a strict DOF budget is enforced. Candidates exceeding a threshold $\tau_I$ are sorted by importance. The algorithm selects high-importance "seeds" and expands their support to neighboring cells (4-connected in 2D, 6-connected in 3D) until the adaptive DOF count reaches a predefined limit (e.g., 3.5% of fluid cells or a hard cap of 120 additional DOFs).

4. Execution Pipeline

State Reconstruction: Transfer particle velocities to the grid and reconstruct the level set.
Indicator Evaluation: Compute the residual predictor (from the previous step), curvature, and velocity gradients to determine the importance field.
Basis Selection: Select adaptive cells under the DOF budget and construct the augmented basis $B$ .
Reduced Solve: Solve the reduced system $B^T A B q = B^T b$ using a preconditioned conjugate gradient method.
Prolongation & Correction: Prolongate the solution to the fine grid and apply a fixed number (3) of Jacobi iterations to reduce fine-scale divergence.
Residual Update: Compute the new fine-grid residual (pre-Jacobi) to update the temporal predictor for the next step.

Key Contributions

Residual-Guided Adaptive Basis: A method that adapts the pressure solution space rather than the grid or particle representation, retaining a coarse global basis while locally augmenting it based on error signals.
Hybrid Refinement Criterion: A selection strategy that combines the fine-grid residual (as a direct error measure) with geometric (curvature) and kinematic (velocity gradient) indicators to detect both persistent and newly forming complex structures.
Temporal Predictor for Single-Solve Efficiency: The use of a filtered previous-frame residual as a predictor allows the adaptive basis to be constructed prior to the current solve, ensuring only one reduced pressure solve is required per time step.
Quantitative Evaluation: Comprehensive evaluation in both 2D and 3D FLIP simulations, demonstrating reduced pressure approximation error and post-projection divergence compared to uniform dimension reduction, with only a modest increase in computational cost.

Experimental Results

The method was evaluated across 11 diverse 2D scenes (e.g., dam breaks, colliding waves, fountains) and extended to 3D.

Accuracy: Compared to uniform dimension reduction (stride 4), the proposed method reduced the average relative pressure approximation error from 0.127 to 0.049 (a ~61% reduction) in 2D. In 3D, the error decreased from 0.134 to 0.057.
Divergence: Post-projection divergence was significantly lowered. In 2D, the average RMS divergence dropped from $1.82 \times 10^{-3}$ (uniform reduced) to $6.7 \times 10^{-4}$ (proposed), narrowing the gap toward the full-resolution solution.
Efficiency: The method adds only a small number of adaptive DOFs (average ~84 in 2D, ~108 in 3D) to the coarse system. Consequently, the pressure solve time increased by only ~25% over the uniform reduced method (1.58 ms vs. 1.26 ms in 2D) but remained ~73% faster than full-resolution projection (5.84 ms).
Ablation: Removing the residual term increased error to 0.081, confirming the residual as the dominant signal. Removing curvature/velocity terms increased error to 0.061, showing their value in detecting new structures. The full indicator combination yielded the best results.
Temporal Predictor: Using a filtered residual predictor reduced the adaptive cell turnover rate to 17.9% (compared to 29.7% for unfiltered previous residuals) while maintaining accuracy comparable to a more expensive "same-frame refine + re-solve" approach.

Significance and Claims

The paper claims that the proposed method offers a favorable accuracy–cost trade-off. It addresses the limitation of uniform dimension reduction, which fails to capture localized pressure variations under aggressive coarsening, without incurring the high computational cost of full-resolution projection or iterative refinement cycles.

The author emphasizes that the method does not aim to exactly reproduce full-resolution pressure but rather to significantly improve the accuracy of reduced representations by selectively allocating degrees of freedom where the Poisson equation is insufficiently satisfied. The approach is presented as a practical enhancement to existing FLIP pipelines, requiring no changes to the underlying particle-grid transfer or grid topology, and is effective in both 2D and 3D contexts. The significance lies in achieving near-full-resolution accuracy in complex dynamic regions while maintaining the computational efficiency of a reduced-order solver.