Graphics x Science: Graphics for Cross-Scale Reliable Scientific Instruments

Lattice metamaterials enable lightweight, multifunctional structures, yet homogenization-based evaluation of their effective properties remains computationally expensive. Neural surrogates offer speed but often lack the accuracy and stability required for engineering-grade simulations. We introduce GMT, a Geometric Multigrid Transformer — a neural solver with high numerical fidelity for fast and reliable lattice homogenization. GMT achieves architectural alignment with Geometric Multigrid (GMG) by restructuring Point Transformer V3 to operate across sparse GMG hierarchies, capturing long-range dependencies and cross-level interactions essential for multigrid convergence. To enforce physical consistency, GMT incorporates physics-aware positional encoding for strict enforcement of periodicity and predicts both the finest-level solution and multi-level residual corrections. These predictions deliver a spectrally-aligned initialization, enabling end-to-end training under physics-informed and solver-aware losses and requiring only a single GMG V-cycle refinement to reach convergence. This fusion of neural prediction and numerical rigor achieves relative residual errors of 10⁻⁵ with a 160× speedup over state-of-the-art GPU-based solvers at equivalent accuracy — particularly at high resolutions (e.g., 512³), where traditional methods become most costly. We validate GMT across mechanical and thermal domains, demonstrate robust generalization to unseen geometries and non-periodic settings, and showcase scalability to high resolutions — enabling real-time design iteration, multi-scale simulations, high-throughput material discovery, and inverse design.

We summarize two graphics pipelines: one for boundary-aware reduced simulation and one for scratch-anisotropy design of specular imagery. The first develops neural model reduction for PDE systems whose boundary configurations vary with geometry or design choices. The second translates target view-dependent shading into scratch orientation, density, and manufacturable curves on specular substrates. Together, the two projects show how graphics methods make physically meaningful quantities explicit within concrete computational pipelines. We describe each project's formulation, representative results, and practical scope.

Simulating large-scale articulated assemblies poses a significant challenge due to the numerical stiffness and geometric complexity of jointed structures. Conventional rigid body solvers struggle with the high nonlinearity induced by rotation parameterization. This difficulty becomes more pronounced for multiple two-way-coupled bodies. This paper introduces a novel framework that leverages the linear kinematic mapping of Affine Body Dynamics (ABD). As ABD targets near-rigid objects, the constitutive variations of different materials become negligible, which justifies a co-rotational approach to isolate geometric nonlinearities of the system. This insight enables the use of constant system matrices that can be pre-factorized throughout the simulation, even with fully implicit integration schemes. To manage the high DOF counts of large-scale systems, we map primal body coordinates onto a compact dual space defined by minimal joint degrees of freedom. By solving the resulting KKT systems, our method ensures exact constraint enforcement and physically accurate motion propagation. We provide a suite of specialized solvers tailored for diverse joint topologies, including chains, trees, closed loops, and irregular networks. Experimental results show that our approach achieves interactive rates for systems with hundreds of thousands of bodies on a single CPU core, while maintaining excellent stability at large time steps.

Holographic displays offer natural 3D visual cues by reconstructing complete light fields through optical interference. Recent advances in computer-generated holography (CGH) have improved display quality, expanded etendue, and reduced computation time. However, most CGH pipelines still rely on image-based supervision, separating rendering with wave propagation, which cannot preserve complete light transport with phase information. As a result, they often lose important 3D cues such as continuous focus, parallax, occlusion, and material-dependent visual effects. Mesh representations, which are widely studied in computer graphics, naturally retain continuous geometry, visibility, texture, and material information. In this extended abstract, we discuss two recent mesh-based holography pipelines that aim to address these limitations. In the first, we develop a camera-calibrated wave propagation model tailored to mesh-derived wave fields, together with complex-valued optimization, enabling optical reconstructions that preserve continuous depth and view-dependent structure. In the second, we further combine path tracing with Rayleigh–Sommerfeld integration to capture complete wave-field transport in photorealistic scenes with reflective, refractive, glossy, and globally illuminated effects. Together, our pipelines bridge CGH and computer graphics, advancing CGH toward more faithful full-3D visual reconstruction.

This paper presents a differentiable pipeline for topology optimization of high-resolution mechanical metamaterials on grid domains, enabling complete geometric freedom within a fixed-resolution design space. Our method begins with a microstructure generation procedure based on random walks, which avoids hand-crafted parameterizations and populates the design space without strong geometric priors, yielding a diverse set of mechanically meaningful microstructures. We train a convolutional neural network to predict homogenized stiffness tensors from these microstructures, enabling a fast and differentiable approximation of mechanical behavior without the need for finite element solves. By plugging this surrogate into a topology optimization loop, we can backpropagate through mechanical objectives and discover high-resolution, fabricable designs across a wide range of densities and target behaviors. We demonstrate our pipeline's inverse design capabilities, producing microstructures with both isotropic and anisotropic stiffness, and validate our predictions through mechanical testing.

Reconstructing object surfaces from point cloud observations is a fundamental problem in both computer graphics and robotics. Classical reconstruction algorithms convert point clouds into mesh representations using geometric priors or analytic formulations; more recently, implicit representations have enabled high-quality reconstructions by learning continuous functions that describe object geometry. Despite their success, many modern reconstruction techniques assume access to large datasets or dense observations. In contrast, robotic systems typically operate with limited sensing capabilities: robots perceive objects through sparse and noisy measurements obtained from sensors such as RGB-D cameras, LiDAR, or tactile arrays, integrated incrementally as the robot interacts with the environment. In such settings, reconstructing the geometry alone is insufficient — robots must also reason about the uncertainty of the reconstructed surfaces in order to plan safe interactions and guide further sensing actions. This work revisits Gaussian process implicit surfaces as a principled framework for uncertainty-aware 3D reconstruction in robotic settings.

We present an algorithm for global knit structure planning that leverages a generalization of power diagrams to triangulated surfaces. This generalization is based on modified geodesic heat kernels and is used to quantize the curl measure of a normalized knitting time function gradient. Knit singularity positions are optimized jointly in a global fashion via an iterative Lloyd-type algorithm, leading to faster and more optimal placement of singularities than prior work, allowing for practical creation of denser knit graphs. In this denser setting, we present singularity ordering constraints that more robustly achieve helix-free knit graphs. The speed and robustness of the method is demonstrated via a diverse array of knits, and a virtual gallery of helix-free knit graphs. We also provide further demonstration of user constraints for knit singularity masking, level set alignment constraints, and apparent seam placement via curl boosting.

Accurate registration of brain MRI scans is fundamental for cross-subject analysis in neuroscientific studies. This involves aligning both the cortical surface of the brain and the interior volume. Traditional methods treat volumetric and surface-based registration separately, which often leads to inconsistencies that limit downstream analyses. We propose a deep learning framework, NeurAlign, that registers 3D brain MRI images by jointly aligning both cortical and subcortical regions through a unified volume-and-surface-based representation. Our approach leverages an intermediate spherical coordinate space to bridge anatomical surface topology with volumetric anatomy, enabling consistent and anatomically accurate alignment. By integrating spherical registration into the learning, our method ensures geometric coherence between volume and surface domains. In a series of experiments on both in-domain and out-of-domain datasets, our method consistently outperforms both classical and machine learning-based registration methods — improving the Dice score by up to 7 points while maintaining regular deformation fields. Additionally, it is orders of magnitude faster than the standard method for this task, and is simpler to use because it requires no additional inputs beyond an MRI scan. With its superior accuracy, fast inference, and ease of use, NeurAlign sets a new standard for joint cortical and subcortical registration.

We present ∂∞-Grid, a differentiable grid-based representation for efficiently solving differential equations (DEs). Grid-based neural fields (e.g., Instant-NGP, K-Planes) train fast via spatial locality but rely on linear interpolation, which lacks higher-order differentiability and cannot solve DEs. On the other hand, coordinate-based MLPs are infinitely differentiable but computationally expensive, requiring hours of training. ∂∞-Grid combines the efficiency of feature grids with radial basis function (RBF) interpolation — smooth and C∞ — enabling accurate computation of gradients, Jacobians, and Laplacians for solving DEs. Our multi-resolution co-located grid captures high-frequency solutions with fast global gradient propagation. We validate ∂∞-Grid on Poisson, Helmholtz, and Kirchhoff-Love thin-shell equations, achieving 5–20× speed-ups over coordinate-based MLPs while maintaining comparable accuracy.

We propose a unified nonlinear optimization framework that achieves a monolithic coupling of incompressibility, viscosity, and surface tension within a single solver. This first particle-based solver resolves their interdependence via position-based nonlocal viscosity formulations, enhancing stability beyond operator-splitting methods.

We propose a GPU-friendly adaptive algebraic coarsening method guided by Green strain increments and combined with affine embedding to avoid topological changes and irregular memory access, yielding the first fully GPU-optimized adaptive IPC solver with up to 3× speedup over state-of-the-art GPU IPC while preserving identical visuals.

This paper presents an ambient-robust inverse rendering method using robot-assisted RGB–NIR imaging. By combining multi-view RGB images with NIR flash images, we reconstruct accurate geometry and reflectance under multiple lighting conditions. We also introduce a robot-assisted imaging system and multi-view RGB–NIR dataset collected across multiple ambient illumination settings.

AtomSlicer is a 3D printing method for fused filament fabrication that generates non-planar layers and continuous toolpaths aligned with user-defined fields. It enables better control of layer orientation, constant thickness, and near-continuous deposition, helping improve print quality, reduce interruptions, and support advanced multi-axis fabrication.

While thin shells have been studied for decades, there is little consensus on how to discretize them. We systematically study bending models for Kirchhoff-Love shells. We develop new models and methods to address discovered simulation gaps, and make practical recommendations of when and how these models should be used.

Neural eigenanalysis enables reduced-order modeling across shape families but is limited to Neumann boundaries. We extend it to Dirichlet, Robin, and mixed conditions by treating boundary configurations as inputs, forming a unified shape–boundary space for consistent spectral analysis and simulation across varying geometries and boundary settings.

The Material Point Method struggles with separating materials due to a shared background grid that couples velocities. We introduce a hybrid approach using separate velocity grids and a unified pressure model, enabling buoyancy-driven phase separation for immiscible fluids.

A field-based computational framework for designing fabric-reinforced composites with optimized topology and manufacturable semi-continuous layers, producing laminates up to 43.8% stiffer than conventional planar-ply designs.

We propose CoherentRaster, an efficient 3D Gaussian Splatting-based light field rendering framework for glasses-free 3D displays. CoherentRaster introduces Cross-view Coherent Attribute Reuse to eliminate redundant computations across neighboring viewpoints and View-coherent Remapping to restore warp-level memory efficiency degraded by interlaced subpixel layouts, enabling real-time, high-quality synthesis.

Complex-valued holographic radiance fields optimize holographic 3D scenes without relying on intensity-based intermediaries. Leveraging multi-view images, our method uses complex-valued Gaussian primitives encoding amplitude and phase aligned with scene geometry. This eliminates expensive single-view hologram recalculation, accelerating speed by 30×–10,000× while matching state-of-the-art quality, bridging wave optics and 3D scene geometry.

We present a computational approach for designing contact-based coordinate-motion assemblies that meet user-specified target appearance and motion. One key enabler of our approach is that we established a theoretical connection between contact geometry within an assembly and unique coordinate motion of the assembly's parts.

The interplay of morphology, control, and anisotropic friction makes optimal design for terrestrial locomotion challenging. We present a computational pipeline that co-designs friction and controllers across diverse robot morphologies, showing anisotropic friction is critical and achieving statistically superior performance over state-of-the-art methods.

Affine Body Dynamics within Incremental Potential Contact enables accurate simulation of near-rigid, extremely stiff solids with strict non-penetration. We present a distributed ABD solver based on consensus ADMM, where nodes solve local subproblems in parallel and synchronize shared bodies through global consensus, achieving robust convergence, non-penetration, and scalable multi-node performance.

We present a quadratic B-spline FEM approach for cloth simulation that captures smoother, more accurate deformations with rich wrinkle detail and few artifacts. Our optimized reduced integration scheme enables better performance compared to traditional linear FEM, while supporting robust, high-quality cloth simulation for realistic garment animation.

To enable seamless mixed-reality telepresence in the wild, we present EgoRelight, a robust human performance capture and HDR illumination recovery approach using a single portable head-mounted device. This dual capability drives the photorealistic and relightable full-body 3D avatars, blending users naturally into surrounding environments.

FastVEM is an efficient boundary-conforming fluid simulation framework. It combines Virtual Element discretization, a simulation-friendly body-fitted grid construction strategy, and a tailored geometric multigrid method to achieve robust, high-fidelity, and efficient fluid–boundary interaction.

We present solutions to the floating-point–induced decision errors arising in CCD for parametric surfaces, including an error-resistant CCD method and a data construction pipeline with known GT for algorithm evaluation under floating-point perturbations. Experimental results demonstrate that the proposed methods are effective and robust.

FreeShell introduces a robust thermal-shrinkage-actuated 4D printing technique for fabricating freeform thin-shell surfaces. By printing triangular tiles connected by shrinkable connectors using a single material, heating triggers the transformation from flat structures into 3D shells. An optimized mesh layout algorithm ensures geometric accuracy and robustness, reducing material and environment dependency.

We present a physics-aware method for reconstructing dynamic fluids from sparse-view videos. By integrating volumetric rendering with physically guided Gaussian optimization, it enforces consistency in density and motion, reducing artifacts and improving realism. The approach achieves accurate, stable 3D smoke reconstruction with high visual and physical fidelity.

Optimal imaging performance requires designing optics with the downstream processing in the loop—especially as AI pipelines grow in importance. Yet the scalar nature of loss functions in the processing-aware setting breaks industry-standard lens design solvers. We generalize classical ray aberrations to task-driven objectives, bringing robust solvers to end-to-end design.

We introduce a highly efficient variational framework for computing optimal vortex cores in 3D unsteady flows. By unifying diverse detection criteria through time-preintegrated Lagrangians, our method accurately reconstructs spacetime vortex manifolds by solving Euler-Lagrange equations in just one time step, enabling rapid and robust flow analysis.

We present a gradient-domain Monte Carlo framework for solving Poisson equations on complex domains. The method directly estimates solution differences between query points and reconstructs the final solution efficiently, reducing variance and improving convergence without introducing additional bias beyond the underlying WoS-based solver.

We propose a conservative algorithm to test the geometrical validity of polynomial finite elements as they deform linearly in time. In elastodynamic simulation, our algorithm guarantees that the system remains physically valid during the entire trajectory, not only at discrete time steps, even when implemented using floating point arithmetic.

To bring physically accurate 3D cues to holography, HoloPathTracer moves beyond prior computer-generated holography pipelines that separate radiance rendering from wave propagation. By tracing complex wave paths directly through the scene, it enables phase-only holograms with natural defocus, reflections, refractions, and more faithful visual realism.

We demonstrate an invisible holographic window: a transparent surface‑relief computer-generated hologram formed on glass. Using a scattering‑suppressed encoding scheme and spatial‑frequency band‑division multiplexing, we achieve high‑transmittance, crosstalk‑free full‑color 3D image reconstruction, allowing clear viewing of a photorealistic holographic image overlaid in front of real‑world scenes through the glass.

We present a computational method to design rigid-deployable Kirigami sheets based on isohedral tilings. By exploring combined topology and geometry spaces, our approach identifies self-deployable structures with controlled expansion. Our approach generates diverse designs, demonstrated through an interactive browser and validated with 3D-printed prototypes.

LagrangianSplats reconstructs physically plausible 3D fluid velocity fields from sparse video by combining divergence-free kernels with Lagrangian Gaussian transport. A sliding-window optimization enables efficient long-range supervision, producing state-of-the-art transport consistency and physical accuracy on synthetic and real fluid datasets.

NEO is a neural framework for fast Laplace-Beltrami spectral analysis on 3D point clouds. Rather than computing eigenvectors with expensive iterative solvers, it predicts low-frequency eigenspaces directly from geometry, enabling near-linear scaling, robustness to irregular sampling, and accurate zero-shot transfer to substantially higher-resolution shapes and downstream spectral tasks.

We explore a new direction in mechanical cloaking: halftoning an image using a porous structure that behaves like a linear, isotropic material and visually matches an image. For an external observer, this creates the surprising effect where the object appears mechanically homogeneous while its porous structure resembles a target image.

Our variant of the Material Point Method discretizes both the velocity and stress fields over mixed finite elements, avoiding the need for repeated costly particle-to-grid transfers within the implicit elastoviscoplastic solve. Our method supports materials ranging from sand and snow to elastic solids, with two-way coupling to rigid-body solvers.

This paper extends Monte Carlo PDE solvers to handle nonlinear radiative boundary conditions using a Picard-style fixed-point iteration framework. It introduces a heteroscedastic regression-based denoising method for boundary estimates and demonstrates the approach on heat radiation simulations with complex geometries, achieving higher accuracy than standard linearization strategies.

MotionBricks is a real-time generative framework that transforms interactive motion control for animation and robotics. By combining a large-scale latent backbone with intuitive "smart primitives," it delivers high-quality, zero-shot motion synthesis at 15,000 FPS, allowing users to effortlessly build complex and fine-grained animations and robotic applications.

This work presents a new class of gear mechanisms, called multi-point conjugation gear mechanisms, as well as computational techniques to modeling these gear mechanisms for exactly generating user-specified 3D motions under external loads.

MPM Lite is a hybrid Lagrangian/Eulerian method that removes particle-based quadrature at solve time. It achieves 15.9 times speedup over implicit MPM, and 1.88 times speedup over explicit MPM.

We present PAColorHolo, a perceptually-aware color management framework that enables accurate color reproduction in holographic displays. Our approach jointly addresses system-level color distortions through color space transformation, adaptive illumination control, and a color-restoration neural network. Simulations, optical experiments, and a user study demonstrate improved perceptual color fidelity.

We present a system for optimizing spacecraft designs under radiation pressure, the force that photons exert on objects. It combines a fast Monte Carlo simulation, a neural proxy of it, and gradient-based optimization to find spacecraft geometry, or operational parameters that minimize travel time or maximize proximity to a target.

We present a reinforcement learning-based retargeting method that transfers human motion to diverse morphologies, including humanoids and a quadruped, without artifacts such as foot-sliding, hovering, or self-penetration. A bilevel optimization framework refines upper-loop retargeting parameters defined via sparse semantic correspondences while training the lower-loop RL policy, eliminating manual correspondence tuning.

We present a computational fabrication pipeline for retrofitting 3D objects with real-time multi-touch sensing without modifying their interiors. From a 3D scan, our method routes and optimizes surface-conforming electrode layouts under geometric, fabrication, and hardware constraints. We guide assembly via projection to enable precise spatial input sensing on curved objects.

We present a simple, robust, and consistent estimator for boundary path integrals in physics-based differentiable rendering. By reformulating boundary integration with kernel-density estimation, our method avoids fragile measure-zero sampling and costly reparameterization. It matches finite-difference derivatives and outperforms state-of-the-art baselines in synthetic inverse-rendering benchmarks with improved efficiency and lower variance.

We propose sample matching, a variance reduction principle for differentiable volume rendering. By jointly estimating scattering and transmittance gradient components at shared sample positions, our method reduces gradient variance by up to 80%, significantly improving convergence speed and reconstruction quality in volumetric inverse rendering.

This work introduces a compact, expressive design space for shellular metamaterials, along with a fast GPU-based homogenization pipeline that evaluates elastic properties in near real time. This enables interactive exploration and inverse design, achieving diverse geometries and a broad range of mechanical behavior at low solid volume.

SMP builds reusable and modular reward models for training motor controllers. Once constructed from a motion dataset, the priors can be reused across diverse tasks while preserving the behaviors in the data, without requiring access to the original dataset or retraining.

We present ST-FLIP, a spatiotemporal extension of the Fluid-Implicit Particle (FLIP) method for free-surface and two-phase simulation. ST-FLIP augments particles with 4D spatiotemporal coordinates and supports time steps up to an order of magnitude larger than in existing solvers, delivering several-fold speedups for multi-billion-particle simulations at very high grid resolutions.

Geomorphological transport is the long-distance transport of quantities which are key elements of terrain erosion. We propose a new stochastic algorithm for the efficient simulation of geomorphological transport with momentum conservation, which enables us, for the first time, to model dynamic emergent deltas, meanders and debris fans in eroded terrains.

SymX is an open-source framework that turns succinct symbolic energy expressions into optimized code, delivering automatic first- and second-order derivatives and parallel assembly for Newton-style optimization time integrators. SymX facilitates building complex simulations, including high-order FEM, rigid-deformable coupling, and frictional contact, accelerating simulation research.

We present a method for exploring the space of uniformly deployable hinged kirigami structures, as a constrained embedding space of an arbitrary planar graph. The design space can be used for desired deployment properties, such as conformal behavior and fully-closed deployed configuration, and for inverse design.

We present a physically-based framework for simulating sand–water–air mixtures by coupling LBM fluids with MPM granular sand under a unified formulation. A water retention model with built-in volume conservation enables stable, realistic simulation of mixtures across diverse, multiscale scenarios.

We introduce a novel variance reduction scheme for Walk on Spheres with Laplace's equation, demonstrate performance, and provide analytic guarantees on performance.

Woodstock is a novel biophysical simulation framework for fungal wood decay. By tightly coupling rot dynamics, wood states, and strand-based mechanics, our method reproduces characteristic decay phenomena such as fungal colonization, trunk hollowing, wood fracturing, and eventual structural collapse.

YASPS is a programming system for automatically computing derivatives of your physical simulation energy from symbolic definitions, specializing to contact-rich scenarios. YASPS captures the structure of the simulation data, parameterization, and sparsity, and produces modular and efficient GPU code with performance on par with manually written code.