Papers
Topics
Authors
Recent
Search
2000 character limit reached

Adaptive Volumetric Mechanical Property Fields Invariant to Resolution

Published 16 Jun 2026 in cs.CV, cs.LG, and cs.RO | (2606.18231v1)

Abstract: Accurate mechanical properties (or materials) Young's modulus ($E$), Poisson's ratio ($ν$) and density ($ρ$) are essential for reliable physics simulation of digital worlds, but most 3D assets lack this information. We propose AdaVoMP, a method for predicting accurate dense spatially-varying ($E$, $ν$, $ρ$) for input 3D objects across representations, improving the resolution, accuracy, and memory efficiency over the state-of-the-art. The foundation of our technique is a sparse and adaptive voxel structure SAV that efficiently represents both the input 3D shape and the material field output. We replace the fixed-voxel model of the most accurate prior method, VoMP, with a novel sparse transformer encoder-decoder model that learns to generate a unique SAV autoregressively for every input shape to represent its materials, achieving a resolution $163\times$ higher than prior art. Experiments show that AdaVoMP estimates more accurate volumetric properties, even with lesser test-time compute than all prior art. This allows us to convert high-resolution complex 3D objects into simulation-ready assets, resulting in realistic deformable simulations.

Summary

  • The paper introduces ADAVOMP, a novel neural architecture that employs adaptive sparse voxel representation and transformer modules to predict high-resolution volumetric mechanical properties.
  • The paper demonstrates significant improvements in resolution scaling and prediction accuracy, achieving up to 1024^3 resolution with substantially reduced errors compared to state-of-the-art methods.
  • The paper advances simulation readiness by automating the generation of physically plausible material fields using compact representations and cross-modal supervision.

Adaptive Volumetric Mechanical Property Fields Invariant to Resolution

Introduction

The paper "Adaptive Volumetric Mechanical Property Fields Invariant to Resolution" (2606.18231) presents ADAVOMP, a neural architecture for predicting high-resolution, spatially-varying mechanical property fields (Young's modulus, Poisson's ratio, density) over 3D shapes. This task is critical for converting visual 3D assets into simulation-ready bodies for robotics and physically-based AI workflows, a process that is currently data- and labor-intensive. State-of-the-art baselines optimize for either geometric fidelity or computational efficiency but are bounded by fixed-resolution representations; as a result, they lack the adaptability required for highly heterogeneous material fields and complex internal structures.

Methodology

ADAVOMP introduces a sparse adaptive voxel (SAV) representation for encoding both geometry and materials, together with a novel transformer-based encoder-decoder architecture specialized for adaptive, hierarchical inference and generation. SAV integrates multi-scale, sparse voxel grids, allowing the model to focus inference on regions of high heterogeneity and represent homogeneous areas compactly. The key advances are:

  • Hierarchical Sparse Adaptive Voxels (SAV): Unlike conventional octree or fixed-resolution sparse voxel representations, SAV learns to allocate resolution dynamically based on local material complexity, a task-specific refinement criterion (not geometry-agnostic). This substantially increases representational compactness while preserving query fidelity for regions with large, homogeneous properties.
  • Adaptive Geometry Transformer (AGT): AGT encodes SAV hierarchies of DINOv3 [Siméoni et al., 2026] multi-view features. Network architecture leverages unified coordinate embeddings, hierarchical positional encodings (RoPE), and sparse shifted-window attention to propagate context across disparate spatial and resolution scales.
  • Adaptive Material Generator (AMG): AMG is a level-conditioned, autoregressive transformer decoding the adaptive tree structure and per-voxel material parameters. It operates coarse-to-fine, at each stage emitting KEEP/SUBDIVIDE/EMPTY signals and latent codes projected through a VAE (MatVAE) mapping to the triplet (E,v,p)(E, v, p). The generator is trained jointly with the encoder via teacher forcing and multi-scale supervision.
  • Cross-modal Supervision: Training utilizes a large dataset with material labels bootstrapped via VLM-based annotation pipelines and incorporates multi-scale loss functions, negative supervision on empty regions, and direct regularization in physical material space.

Experimental Results

The model is prioritized for simulation-centric metrics: absolute and relative errors for (E,v,p)(E, v, p), validity with respect to physically plausible material ranges, and final simulation-mass estimation accuracy. Under rigorous benchmarking, including two challenging datasets (GVT-TEST, GVT-HARD) and the ABO-500 mass estimation benchmark, the following holds:

  • Resolution Scaling: ADAVOMP attains up to 102431024^3 prediction resolution, a 16,384×\times improvement over the 64364^3 grids used by VoMP [Dagli et al., 2025] and Pixie [Le et al., 2025]. This allows the recovery of fine-grained fields necessary for high-fidelity simulation in robotics and synthetic environments.
  • Prediction Accuracy: On both object- and voxel-averaged absolute/relative error metrics for EE, vv, pp, ADAVOMP achieves substantial and consistent reductions over VoMP/Pixie and recent NeRF-/splatting-based methods [Zhai et al., 2024; Shuai et al., 2025; Lin et al., 2025a], even when restricted to fixed computational budgets (Table 1-2 in the main text, Tables 8-9 for full-scale ablations).
  • Compactness and Memory Efficiency: The SAV structure achieves a ∼7\sim7\% node count versus a dense 64364^3 grid on real object data, yielding scalable trade-offs between resolution, inference memory, and compute. Learned generation of structure retains this advantage on held-out test data.
  • Simulation Validity: In FEM/mesh-free simulation scenarios, ADAVOMP’s predicted fields enable physically plausible object interaction and deformation, with superior mass estimation performance.

Ablative and Scaling Studies

Comprehensive ablation isolates the contributions of the adaptive structure, hierarchical transformer design, context and embedding parameterizations, empty-space regularization, and MatVAE material subspace: all yield measurable improvements in convergence rate, supervisory signal propagation, and physical accuracy. Model scale and training data augmentation are shown to further tighten the error bounds, demonstrating a compute-efficient Pareto frontier.

Theoretical and Practical Implications

ADAVOMP advances the representation and prediction of physically relevant material fields for 3D shape-based simulation. Key implications are:

  • Task-Specific Hierarchical Representation: The adaptive, autoregressive voxelization directly targets material field complexity, departing from generic spatial structures in prior neural volume representations.
  • Versatile Input Modalities: The architecture generalizes across diverse 3D data (triangular meshes, splats, neural fields), and is not tied to explicit ground-truth geometry surfaces.
  • Faithful-to-physics Generation: By constraining the predicted latents through a pre-trained material VAE and supervising across material scales, the model robustly maps visual/semantic cues to physically meaningful parameters.
  • Scalability for Large-scale Simulation: The method streamlines asset preparation for high DOF robotic simulation, embodied AI, and digital twin environments—removing the bottleneck of manual or per-object property assignment.

Limitations and Future Directions

Despite the empirical performance, the method is inherently limited to isotropic volumetric properties (scalar, not tensor fields) and linear material parameters. Incorporating anisotropy, directionality, or nonlinear yield moduli would require a generalized latent/material field parameterization, and possibly joint vision-physics modeling. The static setting (no explicit temporal/dynamic cues) means the model may not capture viscoelastic effects or context-specific deformation cues present in video or dynamic interaction data. Future research may address structure-function transfer via joint video/geometry supervision, direct regression of additional properties (fracture, damping, thermal) and adaptation to approximate or differentiable simulation backends.

Conclusion

ADAVOMP establishes a new paradigm for efficient, high-fidelity prediction of volumetric mechanical property fields. The method's architectural innovations, superior accuracy, resolution scalability, and practical efficiency solidify it as a foundational technology for automatic, large-scale simulation asset authoring, supporting the next generation of embodied AI, robotics, and interactive virtual environments. Future work will likely extend this approach toward broader physical property spaces, tighter integration with vision-LLMs, and closed-loop simulation-in-the-training frameworks.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Sign Up to Generate

Explain it Like I'm 14

Easy Explanation of “AdaVoMP: Adaptive Volumetric Mechanical Property Fields Invariant to Resolution”

What this paper is about

This paper is about teaching a computer to guess the physical material properties inside 3D objects—things like how stiff, how squishy, and how heavy different parts are—so those objects can behave realistically in physics simulations (for example, a robot pressing on a cushion or a phone bending a little when squeezed). Most 3D models look good on the outside but don’t include these physical details on the inside. The authors introduce a method called AdaVoMP that fills in those missing properties accurately and at very high detail.

What questions the researchers asked

In simple terms, the paper asks:

  • Can we automatically predict the material properties inside a 3D object, not just on its surface?
  • Can we do this at very high resolution (lots of detail) without using too much memory or time?
  • Can the predictions be accurate enough to plug straight into physics engines for realistic simulations?

The properties they predict across the whole volume are:

  • Young’s modulus (E): how stiff something is (like steel vs. rubber).
  • Poisson’s ratio (ν): how much something squishes sideways when you squeeze it.
  • Density (ρ): how heavy something is for its size.

How the method works (in everyday language)

Think of a 3D object like it’s made of tiny Minecraft-like cubes called “voxels.” But instead of filling the whole object with millions of cubes at the same size, AdaVoMP uses an “adaptive” approach: big cubes where the material is the same, and small cubes only where details change. This saves a lot of memory and makes it faster.

Here are the main ideas, explained with analogies:

  • Adaptive voxel tree (SAV): Imagine a map where big regions are used for empty deserts, but you zoom in with detailed street maps only for busy cities. SAV does this for 3D materials—it uses big cubes where the material is uniform (like a solid block of wood) and tiny cubes where it changes (like the edges between plastic and metal parts).
  • Learning from views: The method looks at multiple images or renderings of the 3D object to gather clues about its materials (color/texture hints), then packs those clues into the voxel structure. This is like studying a statue by walking around it and taking notes from many angles.
  • Transformer model: A transformer is an AI model that pays attention to the most important parts of the data (like a focused reader). Here, it learns which regions need more detail and what materials to put there.
  • Coarse-to-fine generation: The model starts with big cubes and decides for each one: keep it, split it into smaller cubes, or mark it as empty space. It repeats this, zooming in step-by-step only where needed, and filling in the material values as it goes.
  • Staying physically realistic: The system decodes its predictions through a learned “material decoder” (borrowed from earlier work) that helps ensure the values for stiffness, squishiness, and density stay within realistic ranges.

In short, AdaVoMP sees a 3D shape, uses smart attention to focus on important areas, and builds a detailed “material map” inside the object—using only as much detail as necessary in each spot.

What they found and why it matters

The authors tested AdaVoMP against other methods and found:

  • Higher resolution: It reaches an effective detail of up to 1024×1024×1024 cubes, while strong previous methods typically worked at about 64×64×64. That’s a huge jump in detail.
  • Better accuracy: It predicts the three key properties (E, ν, ρ) more accurately than other methods, including strong baselines. Even when run at lower detail (to save time), it often still beats older methods.
  • More efficient: Because it only adds fine detail where needed, it uses memory and compute more wisely than methods that keep everything at the same resolution.

Why this matters:

  • More realistic simulations: Robots pushing, pulling, or gripping objects can be simulated more faithfully. Sofas squish, pillows settle, and metal parts stay rigid where they should.
  • Faster content creation: Game developers, movie artists, and robotics teams can turn ordinary 3D models into physics-ready assets without manually labeling materials.
  • Scales to complex scenes: Works on detailed, multi-part objects and scenes that were too hard or too heavy for previous approaches.

What this could lead to next

  • Improved training for robots: Robots can practice in believable virtual worlds that match real-world behavior more closely.
  • Better design tools: Engineers and artists can quickly experiment with different materials and see how objects would bend, bounce, or break.
  • Future upgrades: The method could expand to even more physical properties (like how materials behave when heated or when they yield under stress) and handle directional materials (those that behave differently in different directions).

In one sentence: AdaVoMP is a smart, high-detail way to fill 3D objects with realistic physical properties so simulations look and feel more like the real world.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

The paper advances high-resolution, adaptive prediction of volumetric mechanical properties, but leaves several concrete issues unresolved:

  • Data provenance and label noise: The training set relies on VLM-based auto-annotation rather than physically measured ground truth. There is no quantification of annotation noise/bias, nor techniques to mitigate it (e.g., noise-robust losses, label cleaning, or uncertainty-aware training).
  • Real-world validation: No systematic comparison against laboratory measurements (e.g., ASTM-tested E, ν, ρ) across standard materials; lack of evaluation on real scanned objects (CT, photogrammetry) to assess domain shift.
  • Interior inference from surface cues: The method infers interior volumetric properties from surface appearance and multi-view features; there is no analysis of failure modes for occluded interiors, hollow vs solid parts, or internal heterogeneity that is not visible externally.
  • Material model scope: The model assumes isotropic linear elasticity with scalar EE, ν\nu, and ρ\rho. It does not handle anisotropy (tensor-valued stiffness), viscoelasticity, plasticity/yield, damping, or thermal properties; pathways to extend and validate these are not provided.
  • Physical constraint enforcement: While MatVAE is used to improve plausibility, the paper lacks explicit guarantees or statistics on constraint satisfaction (e.g., ν∈[0,0.5)\nu \in [0,0.5), E>0E>0, ρ>0\rho>0) and combinations that ensure stable constitutive behavior across simulators.
  • Latent bottleneck in MatVAE: The 2D latent used to decode (E,ν,ρ)(E,\nu,\rho) may limit expressivity for complex or multi-modal material distributions. No ablation of latent dimensionality vs accuracy/stability is provided.
  • Exposure bias in autoregressive decoding: Training uses teacher forcing; the impact of test-time distributional drift and error accumulation across levels is not analyzed. Alternatives (e.g., scheduled sampling, DAgger) are not explored.
  • SAV ground-truth construction sensitivity: The tolerance τ and simple averaging used to build target trees can bias supervision (e.g., homogenizing Poisson’s ratio and modulus). There is no sensitivity analysis of τ nor exploration of more physically grounded aggregation/homogenization rules.
  • Structural refinement decisions: The learned SUBDIVIDE/KEEP policy is not compared against alternative splitting criteria or search strategies; consequences of over-/under-subdivision on boundary fidelity and simulation accuracy are not quantified.
  • Thin structures and small parts: No minimum resolvable thickness analysis, anti-aliasing strategies, or guarantees about capturing slender features without spurious erosion/dilation.
  • Empty space prediction quality: The accuracy of EMPTY vs occupied predictions (false positives/negatives) is not reported, especially for inputs like 3DGS/NeRF where occupancy is uncertain.
  • Viewpoint and feature dependency: There is no ablation of number of views, camera distribution, or feature backbone (DINOv3) on accuracy; robustness to lighting/texture changes is untested.
  • Cross-representation generalization: Performance differences across meshes, SDFs, 3DGS, and NeRF inputs are not isolated or quantified; representation-specific failure modes and domain adaptation strategies are missing.
  • Domain shift and robustness: No tests with realistic sensor noise, imperfect geometry, or varying scale/units; robustness to common asset issues (holes, non-manifold meshes, mis-scaled models) remains unassessed.
  • Scale ambiguity and mass: Many 3D assets lack metric scale. The impact of scale errors on density-derived mass and simulation fidelity is not addressed; no mechanism to infer/correct scale is presented.
  • Composite assemblies and cavities: The approach does not explicitly model internal cavities or multi-material assemblies with sub-voxel structure; how density and stiffness aggregate in such cases is not validated.
  • Simulation-grounded evaluation: Beyond per-voxel errors and a few demos, there is no comprehensive, task-driven evaluation of deformation trajectories, contact responses, and stability across multiple physics engines (FEM, MPM, real-time engines).
  • Portability across simulators: The paper assumes “true” parameters are portable, but does not study mapping to engine-specific stable parameters (e.g., damping, numerical stabilizers) or provide translation procedures.
  • Uncertainty quantification: The model provides point estimates without per-voxel confidence or calibrated uncertainty. No method is offered to propagate uncertainty to simulation risk assessment.
  • Compute and accessibility: Training used large GPU clusters (e.g., 32×A100 for 5 days; data preprocessing at 128×A100). There is no exploration of compute-efficient variants (distillation, pruning, smaller backbones) or guidance for low-resource users.
  • Resolution vs accuracy trade-offs: While higher effective resolution improves metrics, a detailed analysis of diminishing returns, compute–accuracy Pareto curves, and adaptive stopping criteria is missing.
  • Fairness of high-resolution comparisons: Baselines not designed for 1024³ are evaluated at that scale; a resolution-agnostic or resampling-fair comparison protocol is not provided.
  • Dataset breadth and balance: The materials/categories represented in GVT are not characterized for balance or coverage; expansion to more diverse, measured datasets and reporting on long-tail performance is needed.
  • Reproducibility and release: It is unclear whether code, trained models, SAV ground-truth trees, and preprocessing pipelines will be released to enable independent verification and extension.
  • Interpretability and part-awareness: The mapping from visual cues to predicted materials is opaque; no analysis of part-level priors, CAD metadata use, or interpretable attributions is provided.
  • Multi-object context: The method processes individual objects; the effect of scene context (e.g., adjacent objects, assembly context) on material inference is not studied, nor is joint inference in scenes.
View Paper Prompt  View All Prompts 

Practical Applications

Overview

The paper introduces AdaVoMP, a feed-forward method that predicts high‑resolution, spatially varying volumetric mechanical properties (Young’s modulus E, Poisson’s ratio ν, and density ρ) for 3D objects across representations (meshes, Gaussian splats). Its key innovations are:

  • A sparse adaptive voxel (SAV) tree that compactly represents material fields up to 1024³ effective resolution.
  • A sparse transformer encoder–decoder that autoregressively generates both structure (Empty/Keep/Subdivide) and per‑voxel material latents.
  • Integration with a MatVAE decoder to keep predicted triplets physically plausible.

This enables conversion of typical 3D assets—often lacking mechanical parameters—into simulation‑ready objects for physically based, deformable interactions.

Below are practical applications derived from these findings, organized by deployment horizon.

Immediate Applications

These can be deployed now with available tools and standard GPU resources for inference.

  • Simulation‑ready asset conversion for interactive media
    • Sector: Software, Gaming, VFX, AR/VR
    • Use: Auto-assign volumetric E/ν/ρ to marketplace or studio assets to obtain realistic deformation and collision in engines (e.g., Unreal, Unity, NVIDIA Omniverse, Blender).
    • Tools/products/workflows:
    • “Materializer” plug‑in that ingests meshes/splats and exports PhysX- or FEM-ready assets with SAV-backed materials.
    • Batch conversion pipeline for asset libraries.
    • Dependencies/assumptions: Input asset must be renderable and voxelizable; best results with multi-view image features. Assumes isotropic, linear elastic modeling in downstream sim.
  • Robotics training environments with deformable objects
    • Sector: Robotics
    • Use: Populate simulators (Isaac Sim, MuJoCo, PhysX) with soft objects (pillows, cables, plastics) exhibiting realistic compliance/weight, improving manipulation and sim‑to‑real transfer.
    • Tools/products/workflows:
    • “AdaVoMP‑to‑PhysX” exporter.
    • Domain randomization over material fields using SAV refinements to stress‑test policies.
    • Dependencies/assumptions: Linear elastic assumptions; quality of materials tied to training data and visibility of subparts; inference GPU recommended for higher resolutions.
  • E‑commerce/logistics mass estimation
    • Sector: Retail, Logistics, Finance (costing)
    • Use: Estimate object mass by combining predicted density and known/derived volume for shipping cost prediction and inventory planning (paper shows improved ABO mass benchmarks).
    • Tools/products/workflows:
    • “Mass Estimator” API for product 3D models.
    • Integration into PIM/PLM systems to auto-fill missing weight fields.
    • Dependencies/assumptions: Requires reasonably accurate 3D geometry; internal cavities/assembly complexity may bias mass estimates if not visible.
  • QA and plausibility checks for physics assets
    • Sector: Software, Robotics, VFX
    • Use: Validate that material fields fall within physically plausible ranges and detect unrealistic assignments across large asset libraries.
    • Tools/products/workflows:
    • Static checker that flags out-of-range or discontinuous fields at SAV boundaries.
    • Dependencies/assumptions: Uses MatVAE‑bounded ranges; expects adequate multi-view texture information for reliable inference.
  • Educational and demonstrative physics labs
    • Sector: Education
    • Use: Interactive demonstrations where changing geometry or texture leads to updated compliance, showcasing effects of E/ν/ρ on deformation.
    • Tools/products/workflows:
    • Classroom modules built in Blender/Omniverse with AdaVoMP-inferred materials.
    • Dependencies/assumptions: Requires lab workstations with GPUs for real-time or offline inference.
  • Rapid pre‑FEA material field initialization
    • Sector: Engineering software (CAE/CAD), Product Design
    • Use: Pre-populate coarse-to-fine material fields for meshed objects to accelerate FEA setup (e.g., initial guesses for material assignments across assemblies).
    • Tools/products/workflows:
    • Export SAV as per‑element fields for existing FEA solvers.
    • Dependencies/assumptions: Predictions are isotropic and linear elastic; engineering validation still required.
  • Content authoring for AR/VR with haptic approximations
    • Sector: AR/VR, Haptics
    • Use: Produce per‑voxel stiffness fields to drive approximate haptic cues and realistic object interactions in XR scenes.
    • Tools/products/workflows:
    • Middleware that maps AdaVoMP fields to device-specific haptic parameters.
    • Dependencies/assumptions: Approximate mapping from E/ν to device constraints; scalability depends on device bandwidth and XR engine.

Long‑Term Applications

These require additional research, scaling, or broader ecosystem changes.

  • Scene‑scale auto‑materialization for digital twins
    • Sector: Robotics, Smart Infrastructure, AEC
    • Use: One‑shot large‑scene inference to assign volumetric materials across entire scanned environments for simulation-rich digital twins (maintenance, training, emergency response).
    • Tools/products/workflows:
    • Streaming SAV inference over tiled scans; graph-based consistency across objects.
    • Dependencies/assumptions: Efficient scene-level segmentation; consistent material priors; scalable memory and inference throughput.
  • Online perception‑to‑action for robotic manipulation
    • Sector: Robotics
    • Use: On‑the‑fly per‑object property prediction from RGB‑D scans to adjust grasp forces, trajectory speeds, or deformation-aware planning.
    • Tools/products/workflows:
    • Real-time AdaVoMP variants integrated with perception stacks; policy-conditioned inference budgets (coarse-to-fine).
    • Dependencies/assumptions: Low-latency inference; generalization to novel, untextured, or worn objects; robust handling of partial views.
  • Product validation and virtual certification
    • Sector: Consumer Goods, Furniture, Automotive, Policy/Standards
    • Use: Virtual pre‑compliance tests (drop, load, fatigue) using auto‑assigned material fields to reduce reliance on physical prototypes; potential inclusion in standards for digital certification.
    • Tools/products/workflows:
    • Verification pipelines coupling AdaVoMP fields with certified solvers and audit trails.
    • Dependencies/assumptions: Regulatory acceptance; traceability of training data and uncertainty bounds; extensions beyond linear elasticity.
  • Advanced material behavior (anisotropy, plasticity, viscoelasticity, thermal coupling)
    • Sector: Engineering, Healthcare, Energy
    • Use: Predict directional/tensor properties, yield strength, damping, and thermomechanical responses for accurate multiphysics simulations (e.g., composites, batteries, soft tissues).
    • Tools/products/workflows:
    • Expanded MatVAE-style latent spaces and datasets with tensor targets; constitutive-model–aware training.
    • Dependencies/assumptions: New datasets with high-quality ground truth; stable learning of richer constitutive parameters.
  • Surgical and medical training simulators
    • Sector: Healthcare
    • Use: Patient-specific volumetric tissue property inference from imaging for surgical rehearsal (e.g., organ deformation).
    • Tools/products/workflows:
    • Medical imaging adapters (CT/MRI) and domain-specific priors; coupling to real-time nonlinear FEM.
    • Dependencies/assumptions: Clinical data access/ethics; anisotropy/viscoelastic needs; validation against gold standards.
  • Generative shape–material co‑design and optimization
    • Sector: Design, Additive Manufacturing
    • Use: Co‑optimize geometry and spatial material fields to meet target stiffness/mass/comfort using SAV’s coarse‑to‑fine refinement and differentiable surrogates.
    • Tools/products/workflows:
    • Integration with topology optimization and gradient-free search; latent‑guided manufacturability constraints (e.g., multi‑material AM).
    • Dependencies/assumptions: Differentiable pipelines or strong surrogates; expanded material parameterization and manufacturing-aware constraints.
  • Circular economy and recycling triage
    • Sector: Sustainability/Policy
    • Use: Map predicted density/stiffness ranges to likely material classes and volumes for sorting and lifecycle assessments.
    • Tools/products/workflows:
    • Material‑class classifiers fused with AdaVoMP outputs; facility‑level scanners.
    • Dependencies/assumptions: Robust mapping from mechanical triplets to classes; regulatory buy‑in; real-world scanning infrastructure.
  • Real‑time haptics and XR with high‑bandwidth feedback
    • Sector: AR/VR, Haptics
    • Use: Per‑voxel, physics‑consistent updates of haptic devices in collaborative XR environments.
    • Tools/products/workflows:
    • Edge/cloud streaming of SAV updates; device driver support for spatially varying stiffness cues.
    • Dependencies/assumptions: Hardware bandwidth and stability; low-latency SAV querying and downscaling.

Cross‑Cutting Assumptions and Constraints

  • Model scope: Current outputs are isotropic, linear elastic parameters (E, ν, ρ). Anisotropy, plasticity, damping, and temperature effects are not yet modeled.
  • Data dependence: Training relies on part-level annotations created via a VLM pipeline and material reference ranges; hidden internal structures not visible in images may be misrepresented.
  • Input requirements: Requires voxelizable geometry and multi‑view renderability for best performance; texture quality affects reliability.
  • Compute: Training is costly; inference is feed‑forward and scalable with SAV refinement budget, but higher resolutions still benefit from a GPU.
  • Simulator compatibility: Outputs are “true” material properties; stability or fidelity may depend on the downstream solver and its constitutive model implementation.

Taken together, AdaVoMP’s adaptive, high‑resolution volumetric material prediction unlocks immediate, valuable automation for simulation‑ready asset creation and robotics training, and sets a path toward richer digital twins, advanced materials modeling, and broader industrial and policy impact as datasets and models expand.

View Paper Prompt  View All Prompts 

Glossary

  • AdaVoMP: The proposed method for predicting high-resolution volumetric mechanical property fields using sparse adaptive voxels. "AdaVoMP\ generates high-resolution physically accurate volumetric mechanical property fields"
  • Adaptive Geometry Transformer: A transformer-based encoder for adaptive voxel feature trees with positional embeddings and sparse attention. "An Adaptive Geometry Transformer that embeds adaptive DINO feature trees with unified coordinate embeddings and sparse windowed attention (\S\ref{sec:method_encoder})."
  • Adaptive Material Generator: An autoregressive transformer that outputs both material values and adaptive structure across levels. "Adaptive Material Generator (bottom, \S\ref{sec:method_decoder})"
  • ALDE: An evaluation metric reported for errors (used for Young’s modulus) where lower is better. "ALDE (↓\downarrow)"
  • ALRE: An evaluation metric reported for relative errors (used for Young’s modulus) where lower is better. "ALRE (↓\downarrow)"
  • Anisotropic materials: Materials whose mechanical properties vary with direction. "enables the approximation of anisotropic materials via multiscale modeling"
  • Autoregressive: A generation process where outputs are produced sequentially, conditioning on previously generated elements. "learns to generate a unique {SAV} autoregressively for every input shape"
  • BF16 mixed precision: Training with the bfloat16 numeric format to reduce memory while maintaining accuracy. "We train our models end-to-end in BF16 mixed precision"
  • Cross-attention: An attention mechanism that allows candidates to attend to input latents. "We first apply cross-attention from candidates to the input latents"
  • Depth-attenuated averaging: A feature aggregation scheme that downweights projections by depth to preserve details. "we adopt a depth-attenuated averaging of projected features"
  • Density (ρ\rho): Mass per unit volume, a critical mechanical property for simulation. "density (ρ\rho)"
  • DINOv3: A self-supervised vision transformer used to extract patch-token image features. "DINOv3~\cite{simeoni2025dinov3} patch-token features"
  • Distributed Data Parallelism (DDP): A multi-GPU training paradigm that distributes data across devices. "Distributed Data Parallelism (DDP)"
  • Feed-forward methods: Approaches that infer parameters directly from static inputs without iterative simulation. "In contrast, feed-forward methods like~\citet{dagli2025vomppredictingvolumetricmechanical} and ours"
  • FFN (Feed-Forward Network): The MLP component in transformer blocks that processes token embeddings. "following with a feed-forward network (FFN)."
  • FSDP-2: A fully sharded data-parallel training strategy (PyTorch variant) for large models. "FSDP-2~\cite{zhao2023pytorchfsdpexperiencesscaling}"
  • Gaussian splats: Point-based rendering primitives used to represent scenes and surfaces. "Gaussian splat + mesh environment."
  • GVT (Geometry with Volumetric Trees): The dataset used, containing assets with volumetric tree representations. "Our dataset Geometry with Volumetric Trees (GVT) builds on top of the GVM dataset"
  • Hierarchical hash lookup: A multi-level hashing scheme for fast batched queries over sparse voxel structures. "use a hierarchical hash lookup for fast batched queries."
  • Hybrid Sharded Data Parallelism (HSDP): A training setup combining parameter sharding (ZeRO/FSDP) and DDP. "Hybrid Sharded Data Parallelism (HSDP) i.e. ZeRO-3~\cite{rajbhandari2020zeromemoryoptimizationstraining}/FSDP-2~\cite{zhao2023pytorchfsdpexperiencesscaling} + Distributed Data Parallelism (DDP)"
  • Inverse physics: Optimization methods that infer physical parameters by matching simulated behavior to observations. "Inverse physics methods~\cite{physdreamer, huang2024dreamphysicslearningphysicsbased3d, Liu_2025_CVPR, cleach2023differentiablephysicssimulationdynamicsaugmented, liu2024physics3d, lin2025omniphysgs} optimize parameters from video or priors"
  • Latent space: A compact representation space in which the model predicts material codes before decoding. "predicting per-voxel latents zℓ,i∈R2z_{\ell,i}\in\mathbb{R}^2 in its latent space."
  • MatVAE: A variational autoencoder that decodes latent vectors into physically plausible material triplets. "MatVAE decoder from VoMP~\cite{dagli2025vomppredictingvolumetricmechanical}"
  • Megatron-FSDP: A large-scale training framework combining Megatron-LM with PyTorch FSDP for model sharding. "Megatron-FSDP~\cite{shoeybi2020megatronlmtrainingmultibillionparameter}"
  • NeRF: Neural radiance fields for view synthesis and scene representation. "Approaches based on NeRF and Gaussian splats~\cite{mildenhall2020nerf, 10.1145/3592433}"
  • Neo-Hookean: A nonlinear elasticity constitutive model for deformable materials. "constitutive (or material) models (e.g. Neo-Hookean, St. Venant Kirchoff)"
  • Octant id: A discrete identifier of a child voxel’s position within its parent’s eight subcells. "its discrete octant id, o(i)∈ {0,…,7}o(i)\in\ \{0,\dots,7\}:"
  • Octree: A hierarchical spatial data structure subdividing space into eight-child nodes. "operates over an octree"
  • Omniverse: NVIDIA’s platform for simulation and rendering used to create dataset imagery. "rendering in Omniverse~\cite{omniverse} and Blender~\cite{blender}"
  • OpenVDB: A sparse volumetric data structure commonly used in computer graphics. "octrees or OpenVDB~\cite{Museth2013OpenVDB}, which subdivide based on geometric criteria"
  • Poisson's ratio (ν\nu): A material property describing lateral strain response to axial loading. "Poisson's ratio (ν\nu)"
  • Qwen2.5-VL 72B: A large vision-LLM used for auto-labeling materials. "Qwen2.5-VL 72B~\cite{bai2025qwen25vltechnicalreport}"
  • Refinement frontier: The current set of candidate voxels considered for subdivision during decoding. "an explicit sparse candidate set CℓC_\ell (the refinement frontier)"
  • RoPE (Rotary Positional Embeddings): A positional encoding technique applied inside attention to inject coordinates. "applying RoPE~\cite{rope} on its unified coordinates uℓ,iu_{\ell,i}"
  • SAV (Sparse Adaptive Voxels): A learned multi-resolution voxel tree for compactly representing geometry and materials. "A sparse adaptive voxel (SAV) representation for 3D shapes and materials"
  • Shear modulus: A material property relating shear stress to shear strain, used in elasticity. "yield strength, shear modulus, and thermal expansion."
  • Shifted-window self-attention: Local attention over spatial windows shifted across 3D coordinates. "sparse 3D shifted-window self-attention~\cite{swin, swinv2, Xiang_2025_CVPR}"
  • Simulator-dependence: A limitation where inferred parameters depend heavily on the chosen simulator. "suffer from overfitting, simulator-dependence~\cite{mpm, le2025pixie}, and poor scalability."
  • Softmax: A normalization function mapping logits to probabilities in classification. "pℓ,i=softmax(aℓ,i)p_{\ell,i}=\mathrm{softmax}(\mathbf{a}_{\ell,i})"
  • St. Venant Kirchoff: A classical nonlinear elasticity model for deformable materials. "constitutive (or material) models (e.g. Neo-Hookean, St. Venant Kirchoff)"
  • Teacher forcing: A training technique that feeds ground-truth decisions during autoregressive learning. "We train E\mathbf{E} and G\mathbf{G} end-to-end using teacher forcing"
  • Thermal expansion: The tendency of materials to change dimensions with temperature. "yield strength, shear modulus, and thermal expansion."
  • Transformer encoder-decoder: An attention-based architecture that encodes inputs and decodes outputs. "a novel sparse transformer encoder-decoder model"
  • Unified coordinates: A mapping of multi-level voxels to the finest-resolution grid for consistent attention. "we also map each voxel (level ℓ\ell, index ii) to its unified coordinates"
  • VLM (Vision-LLM): Models combining visual and textual understanding used for annotation. "a dataset auto-labeled with a VLM-based pipeline"
  • VoMP: A prior fixed-resolution voxel-based method for volumetric material prediction. "VoMP~\cite{dagli2025vomppredictingvolumetricmechanical}"
  • Voxelization: Discretizing a 3D shape into a grid of voxels for processing. "requiring that the 3D input shape be voxelized"
  • Yield strength: The stress at which a material begins to plastically deform. "yield strength, shear modulus, and thermal expansion."
  • Young's modulus (EE): A measure of material stiffness under tension/compression. "Young's modulus (EE)"
  • ZeRO-3: A memory-optimization and sharding strategy for large-scale distributed training. "ZeRO-3~\cite{rajbhandari2020zeromemoryoptimizationstraining}"
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now

Collections

Sign up for free to add this paper to one or more collections.

Sign Up

Tweets

Sign up for free to view the 4 tweets with 52 likes about this paper.

Sign Up for Free