💻 computer science

Warp-Geo: Differentiable Geometry Representation for Dynamic-Boundary Simulation and Shape Optimization

The paper introduces Warp-Geo, a differentiable, GPU-accelerated framework that combines uniform grids with implicit Poisson differentiation to enable efficient, frequent geometry updates for gradient-based shape optimization and dynamic boundary simulations.

Original authors: Weijie Zhang, Xiantao Fan, Jian-Xun Wang

Published 2026-11-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Weijie Zhang, Xiantao Fan, Jian-Xun Wang

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

In the world of computer modeling, engineers and scientists often need to simulate how objects move through fluids, like a fish swimming through water or a plane cutting through the air. To do this accurately, the computer must first understand the exact shape of the object and how that shape changes over time. Traditionally, computers have used two main ways to describe these shapes. One method relies on a mesh, a digital net of triangles that wraps around the object like a wireframe. While good for static images, this net becomes a nightmare to update when the object bends or twists, because the computer has to constantly re-weave the connections between the triangles. The other method uses a smooth, invisible field of numbers that fills the space around the object, telling the computer how far every point is from the surface. This approach is much better for moving shapes, but it has historically been slow to update or impossible to use when trying to automatically improve a design.

The challenge has been finding a way to describe a shape that is both fast enough to update every single moment of a simulation and smooth enough to allow a computer to calculate exactly how to change the shape to make it better. If a designer wants to tweak a wing to reduce drag, the computer needs to know not just the new shape, but also how the physics changed in response to that tweak. This requires a mathematical path where the computer can trace its steps backward from the result to the original design, a process known as differentiation. Until now, the tools available could either update the shape quickly without this backward path, or allow the backward path but at a speed so slow it was impractical for real-time use.

A team of researchers has now developed a new system called Warp-Geo that solves this problem by combining speed with mathematical smoothness. The system is designed to take a cloud of scattered points, which might come from a 3D scan or a simulation, and instantly turn them into a smooth, continuous map of distances. This map, known as a signed distance field, tells the computer the exact distance from any point in space to the surface of the object, along with which side of the surface the point is on. What makes this new system unique is that it can rebuild this entire map from scratch in a fraction of a second, every time the object moves, without losing the ability to trace the changes back to the design parameters.

The researchers achieved this by abandoning the complex, tree-like structures that older systems used to organize their data. Instead, they arranged their calculations on a simple, uniform grid, similar to a perfectly regular checkerboard that extends in three dimensions. This regularity allows the system to use powerful graphics processors, the same chips found in video games, to perform millions of calculations simultaneously. When the shape of an object changes, the system does not try to stretch or slide the old map to fit the new shape. Instead, it looks at the new positions of the scattered points and immediately rebuilds the entire distance map from the ground up. Because every step of this rebuilding process is mathematically smooth, the computer can calculate exactly how a tiny change in the shape's design would affect the final result.

To ensure the system works correctly, the researchers first had to teach the computer how to figure out which way is "out" for every point on the surface. When looking at a cloud of points, the computer can guess the direction of the surface, but it often gets confused about whether that direction points inward or outward. The team created a method where the computer shoots a virtual ray from each point to count how many times it crosses the surface. By using the rules of parity—essentially checking if the number of crossings is odd or even—the system can reliably determine the correct orientation for the entire shape, even for very thin or sharp features like a fish's tail or an airplane's wing. This orientation is critical because it allows the system to solve a specific type of equation that fills in the gaps between the points, creating a smooth, continuous surface.

The team tested this new system on a variety of complex shapes, including a swimming fish, an airplane, a car, a jet engine, and even a model of a human heart. In every case, the system successfully reconstructed the smooth surface and the surrounding distance map from the raw point data. They verified that the resulting map was accurate, with the distance values behaving exactly as they should near the surface. The system was also able to handle the rapid, continuous changes required for a simulation of a fish swimming, rebuilding the shape map at every single moment of the motion without needing to pause for slow, manual corrections.

Perhaps most importantly, the researchers demonstrated that the system could be used to automatically improve a design. They set up a task where the computer had to change the shape of a fish model to match a target shape. By using the system's ability to trace gradients backward, the computer adjusted the fish's width and height parameters step by step. The results showed that the computer could find the correct shape much faster than traditional methods that rely on trial and error. In fact, the new system calculated the necessary adjustments about fifteen times faster than older techniques that estimate changes by testing one variable at a time.

The study confirms that this approach is not just a theoretical possibility but a practical tool that can run efficiently on modern computer hardware. The time it takes to rebuild the shape map grows in a predictable way as the grid becomes more detailed, meaning the system can be scaled up for even more complex simulations without breaking down. By removing the bottleneck of slow geometry updates, this new framework opens the door for fully automated design processes where the shape of an object and the physics of its environment are optimized together in a single, seamless workflow. The researchers have made the code for this system available to the public, allowing other scientists and engineers to integrate this capability into their own simulations for everything from aerodynamic design to medical modeling.

Technical Summary: Warp-Geo

Problem Statement

Modern computational workflows, particularly in inverse design, adaptive simulations, and dynamic-boundary problems (e.g., fluid–structure interaction), require geometry representations that are simultaneously accurate, efficient to update, and fully differentiable. Existing methods fail to satisfy all three requirements simultaneously:

Explicit representations (meshes, point clouds) rely on discrete connectivity or sampling patterns. Updating these geometries involves non-smooth operations (e.g., mesh reconnection, point resampling) that break automatic differentiation (AD), preventing gradient propagation for high-dimensional optimization.
Implicit representations (e.g., Level-Set Methods) offer smoothness but face a trade-off between update efficiency and differentiability. Advection-based updates are fast but require periodic reinitialization (solving the Eikonal equation) to maintain the signed distance property. These reinitialization steps involve discrete operations (sign detection, distance correction) that introduce non-smooth discontinuities, rendering them incompatible with end-to-end differentiable optimization.
Poisson surface reconstruction offers a fundamentally differentiable approach by recomputing geometry directly from point clouds without advection. However, classical implementations rely on recursive octree structures. These are computationally irregular, poorly suited for GPU parallelism, and difficult to differentiate efficiently, making them impractical for dynamic scenarios requiring frequent reconstruction at every time step.

The core challenge addressed is developing a geometry representation that enables frequent, high-fidelity updates while maintaining full differentiability for gradient-based optimization.

Methodology

The authors present Warp-Geo, a differentiable, GPU-accelerated framework that reconstructs Signed Distance Fields (SDFs) and surface normals from unstructured point clouds. The pipeline consists of four main modules, all designed for GPU parallelism and automatic differentiation:

1. Normal Vector Estimation (D-PCA)

Instead of standard $k$ -nearest neighbors, the framework uses Distance-Weighted Principal Component Analysis (D-PCA) within a fixed radius $R$ .

A distance-weighted covariance matrix is constructed for each query point, emphasizing nearby samples to suppress noise.
The normal vector is derived from the eigenvector corresponding to the smallest eigenvalue of this matrix.
To ensure numerical stability for AD, Singular Value Decomposition (SVD) is used instead of direct eigendecomposition to avoid ill-conditioning when eigenvalues are nearly equal.
This step is fully parallelized on the GPU using the Warp framework.

2. Parity-Based Normal Orientation Correction

Since D-PCA yields unoriented normals, a global consistency check is required.

Ray Casting: A ray is cast along the estimated normal direction. The parity of surface crossings (even for exterior, odd for interior) determines if the normal points outward or inward.
Point-Cloud Detection: Since no mesh exists, crossings are detected by sampling probe points along the ray and checking for point-cloud occupancy within a detection radius $r_d$ .
Iterative Refinement: To handle sharp features and thin structures where finite-radius detection may fail (false crossings or missed crossings), the framework employs an iterative refinement strategy. This includes checking for ambiguous detection patterns and comparing crossing counts in opposite directions to resolve orientation near complex geometries.

3. GPU-Efficient Poisson Reconstruction on Uniform Grids

The geometry is reconstructed by solving a Poisson equation $\nabla^2 \chi = \nabla \cdot \mathbf{V}$ , where $\mathbf{V}$ is the vector field of oriented normals.

Uniform Grid Discretization: Unlike classical octree-based methods, Warp-Geo uses a uniform Eulerian grid. This eliminates recursive tree construction and irregular memory access, enabling balanced thread workloads and predictable memory patterns on GPUs.
Dense Stencil Interpolation: To overcome the weak coupling issues of standard uniform grids (where a point normal affects only one cell), the framework distributes each point's normal to a $3 \times 3 \times 3$ stencil of surrounding cells using tri-linear interpolation weights. This strengthens local coupling, ensuring smooth and complete surfaces even for sparse point clouds.
Differentiability: The linear system $A\mathbf{a} = \mathbf{b}$ is solved using JAX. While the matrix $A$ depends on the grid topology (fixed) and $\mathbf{b}$ depends on the point cloud, the entire process is differentiable.

4. Surface Extraction and SDF Computation

Isosurface Extraction: The implicit field $\chi$ is converted to a mesh using the Marching Cubes algorithm. This step is piecewise differentiable; non-smoothness only occurs when the discrete topology changes.
SDF Reinitialization: The implicit field is converted to a true Signed Distance Field $\phi$ satisfying the Eikonal property ( $\|\nabla \phi\| = 1$ ). This is achieved by computing the Euclidean distance from each grid point to the extracted triangular mesh and assigning the sign based on the parity-based inside-outside test. This step is fully data-parallel on the GPU.

Differentiability Strategy

The framework combines Warp (for point-wise and grid-based kernels) and JAX (for linear solvers). For the Poisson solve $A(\theta)\mathbf{a}(\theta) = \mathbf{b}(\theta)$ , the gradient with respect to design parameters $\theta$ is computed using the adjoint method. This avoids computing the full Jacobian $\partial \mathbf{a} / \partial \theta$ by solving a single additional linear system, making the gradient computation independent of the number of design parameters.

Key Contributions

Unified Differentiable Pipeline: Warp-Geo integrates normal estimation, orientation correction, Poisson reconstruction, and SDF computation into a single, end-to-end differentiable pipeline optimized for GPUs.
GPU-Accelerated Uniform Grids: By replacing adaptive octrees with uniform grids and dense $3 \times 3 \times 3$ stencils, the authors achieve high geometric fidelity and robustness while enabling efficient, scalable GPU parallelism.
Elimination of Advection/Reinitialization: The framework enables dynamic boundary simulations by directly regenerating the SDF from updated point clouds at every time step, bypassing the non-differentiable and computationally expensive advection and reinitialization steps required by Level-Set Methods.
High-Dimensional Optimization: The framework supports the direct optimization of high-dimensional geometry parameterizations (including full point-cloud coordinates or compact control parameters) without requiring hand-crafted low-dimensional proxies or learned surrogates.

Results

The authors validate Warp-Geo through three primary scenarios:

Forward SDF Reconstruction: The framework successfully reconstructs SDFs for diverse geometries (fish, aircraft, automobiles, jet engines, cardiovascular structures). The reconstructed fields satisfy the Eikonal property globally, with minor deviations only near sharp features where closest-point projections are non-unique.
Dynamic-Boundary Simulation: In a fluid–structure interaction (FSI) simulation of a swimming fish, Warp-Geo was coupled with a differentiable CFD solver. The SDF was regenerated at every time step from the moving point cloud, demonstrating tight integration with physics solvers without advection overhead.
Inverse Shape Optimization: The framework was used to optimize the shape of a fish to match a target SDF. Gradients computed via automatic differentiation were validated against finite differences, showing excellent agreement. AD was approximately 15 times faster than finite differences. The optimization successfully converged from random initialization to the target shape within 500 epochs.
Scalability: Computational cost analysis shows that wall-clock time and GPU memory usage scale favorably (near-linearly) with the number of grid points, confirming the framework's practical feasibility for frequent geometry updates.

Significance and Claims

The paper claims that Warp-Geo addresses a critical gap in differentiable computational modeling by providing a robust, efficient, and flexible geometry representation that seamlessly integrates with downstream differentiable physics solvers.

Enabling End-to-End Optimization: By maintaining differentiability throughout the geometry processing pipeline, Warp-Geo enables true end-to-end shape optimization where gradients flow from physics objectives back to geometric parameters.
Practical Feasibility: The framework demonstrates that frequent geometry updates (e.g., at every CFD time step) are computationally feasible on modern GPUs, removing the geometry update process as a bottleneck in dynamic simulations.
General Applicability: The method does not require training data, guaranteeing high geometric fidelity from the first use, and is applicable to a wide range of complex, non-convex, and thin-structured geometries found in engineering and biological applications.

The authors position Warp-Geo as a foundational tool for next-generation inverse design and adaptive simulation workflows, with future work planned for adaptive resolution and topology-changing geometries.