Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

UniPhy: Learning a Unified Constitutive Model for Inverse Physics Simulation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read UniPhy claims that a single latent-conditioned neural network can act as the constitutive model for elastic, plasticine, sand, and fluids, and that latent optimization through differentiable simulation recovers an object's material from…

desk verdict A solid unified-constitutive-model idea with a genuine no-teacher-forcing comparison, but the inference evidence is undercut by a K-means initialization contradiction and a lack of parameter-recovery validation. read the letter →

arxiv 2505.16971 v1 pith:ANKRT5R3 submitted 2025-05-22 cs.CV

Himangi Mittal , Peiye Zhuang , Hsin-Ying Lee , Shubham Tulsiani This is my paper · ORCID
classification cs.CV
keywords inversephysicssimulationneuralconstitutivemodellatent-conditioneddynamicsmaterialpointmethoddifferentiablepropertyinferenceunified
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

UniPhy's central claim is that one shared neural network can play the role of the constitutive model — the law that turns deformation into internal forces — for elastic solids, plasticine, sand, and Newtonian and non-Newtonian fluids, with a per-object latent vector standing in for the material type. The claim matters because current inverse-simulation methods either require a user to specify the material category in advance and then fit a few continuous parameters, or they train a separate neural network for each scene. UniPhy instead trains once on many simulated trajectories and, at inference, freezes the network and optimizes only the latent so that a differentiable simulator reproduces the observed motion. If the claim holds, material identification becomes a search in a learned latent space, and the recovered latent can drive re-simulation of the same material under new forces, velocities, and geometries without further training. The paper reports lower reconstruction error than the instance-specific NCLaw baseline and than spline-, neural-, and graph-network-based simulators on both replay and generalization tasks across all five material families.

What carries the argument

The load-bearing object is the latent-conditioned pair of networks $(g_\phi, f_\theta)$ sharing a per-trajectory latent $z$, embedded in the differentiable Material Point Method (MPM) simulator. The projection network $g_\phi$ predicts the residual that maps a deformation gradient onto the material's elastic region (the return mapping), and the constitutive-law network $f_\theta$ predicts stress from the projected gradient; both take $z$ as a conditioning input, so material identity is carried entirely by the latent. Training minimizes the L2 losses of Eq. 5 between predicted and ground-truth projected gradients and stresses, and inference minimizes the position loss of Eq. 7 through the differentiable MPM rollout, whose affine particle-in-cell (APIC) transfer carries gradients from particle positions back to the latent. The additive-residual and symmetric-stress constructions keep the learned functions aligned with the physical targets. The latent, jointly forced by both losses, must encode the return mapping and the stress law at once, and the differentiable simulator is what turns position error into a latent gradient.

What would settle it

Train on a parameter grid, then build a test object whose parameters fall strictly between two training samples, for instance a Newtonian viscosity halfway between two trained values. Infer its latent from position observations alone using a random initialization rather than the cluster-center warm start, and re-simulate the object under a novel scenario. If the re-simulation matches a neighboring training material instead of the intermediate ground truth, or if random-init inference lands far from the ground truth while cluster-center-init inference succeeds, then the latent is retrieving a training cluster and inference fails for materials outside the training distribution.

Watch

Extended reading notes

Core claim

UniPhy's central discovery is that the constitutive behavior of diverse materials lives on a common learned manifold: a single latent-conditioned model, embedded inside a differentiable Material Point Method (MPM) simulator, can both represent and identify them. The model replaces the two material-dependent functions of a simulator — the deformation-gradient projection (the plastic return mapping that pulls deformation back onto the elastic region) and the stress-generating constitutive law — with neural networks $g_\phi$ and $f_\theta$ that share one latent code $z$. During training, simulated trajectories across geometries and material parameters supply ground-truth deformation gradients, projected gradients, and stresses; jointly fitting the network weights and per-trajectory latents to those targets (Eq. 5) pushes the latent space to encode material identity. At inference the weights are frozen, and the latent is optimized (Eq. 7) so that the rollout of the differentiable simulator matches observed particle positions; the optimized latent then re-simulates the material under novel conditions such as doubled time horizons, unseen velocities, and different geometry. The claim is therefore that position observations alone, run through this latent optimization, recover a material representation that generalizes beyond the observed trajectory.

Load-bearing premise

The load-bearing premise is that matching the observed particle positions is enough to recover the true constitutive behavior: if the latent optimization lands in poor local minima, or if the K-means-based warm start reveals which training material the latent belongs to, the 'inference' could amount to retrieving a training material rather than identifying the object's actual properties.

Editorial extensions

If this is right

  • A user can infer an object's material from its observed motion — 3D or 2D particle positions — without stating whether it is elastic, plasticine, sand-like, or fluid, and can then replay the same trajectory with the trained network.
  • The inferred latent supports re-simulation under conditions not present in the observed trajectory: extended time horizons, unseen linear and angular velocities, and different initial geometries, which the paper's Table 2 reports across all five materials.
  • Shared training across materials improves accuracy and reliability relative to per-scene networks, suggesting that material behavior is a continuous variation around a common physics rather than five unrelated laws.
  • Because inference does not rely on teacher forcing (periodically restarting the simulation from ground-truth internal states), the method is compatible with observers that see only positions, not internal simulator state.
  • The same trained model covers Newtonian and non-Newtonian fluids, which the paper notes prior neural constitutive modeling found difficult.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension beyond the paper: if the latent space is continuous in physical parameters, a linear readout from the optimized latent should predict stiffness, viscosity, or yield stress quantitatively; the paper demonstrates behavior-level replay but does not show such a parameter-to-latent mapping.
  • The paper's Section 4 states the inference latent is initialized randomly from $\mathcal{N}(0, I)$, while its implementation details warm-start from the best K-means cluster centers of the training latent space; a fair test would run inference from random initializations and measure how much of the reported accuracy survives, since cluster centers may leak material-class information.
  • The position-only objective of Eq. 7 has no identifiability analysis: a stiff elastic solid and a highly viscous fluid could produce similar bulk motion under slow forcing, so a natural extension is an uncertainty-aware or multi-hypothesis treatment of the inferred latent.
  • The authors list non-homogeneous scenes and the need for known initial geometry and exact 3D positions as limitations; a concrete next step is per-region latents combined with a reconstruction front end, turning this unified material model into a component of a vision-to-physics pipeline.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes UniPhy, a unified latent-conditioned neural constitutive model trained on simulated trajectories of elastic, plasticine, sand, Newtonian, and non-Newtonian materials. The model consists of a latent-conditioned deformation-gradient projection network and a latent-conditioned constitutive law network, trained jointly with per-trajectory latents using L2 losses on projected deformation gradients and stresses. At inference, given observed particle positions, the latent is optimized through a differentiable MPM simulator to match the trajectory, enabling material inference and re-simulation under novel conditions without user-specified material type information. The authors compare against NCLaw, spline, neural, and GNN baselines on reconstruction and generalization tasks, and report ablations on latent specificity, component substitution, and latent size.

Significance. If the central claim holds, UniPhy is a meaningful step beyond per-scene neural constitutive models (NCLaw) and system-identification approaches, offering a common latent space for diverse materials and inference from motion observations. The paper's strengths include a clear formulation, an internally consistent set of ablations (optimized versus random latents, ground-truth component substitution), and evaluation across several material families including non-Newtonian fluids. However, the load-bearing claim that latent optimization recovers true constitutive behavior is not yet supported: the inference initialization is described contradictorily, the reported metrics are exclusively trajectory reconstruction errors, and no parameter-level or stress-level validation is provided. The quantitative evaluation also lacks error bars or significance tests. The significance of the contribution is therefore conditional on resolving these issues with additional experiments.

major comments (5)
  1. [Sec. 4 vs Sec. 5 (Implementation Details)] Sec. 4 states that the inference latent is 'randomly initialize[d] ... from N(0,I)', while Sec. 5 states 'We initialize the latent during inference from the best performing cluster centers found using K-means clustering on the learned latent space.' Because the learned latent space is built from per-trajectory training latents, K-means centers can correspond to material classes, and selecting the best-performing center is effectively a material-type search. This directly affects the abstract and Sec. 1 claim that UniPhy 'does not rely on user-specified material type information.' Please resolve the contradiction, report results with a truly random N(0,I) initialization, and compare with the K-means initialization to quantify the dependence of the reported gains on this choice.
  2. [Table 3, Random Latent row] The 'random latent' baseline uses another trajectory's learned latent (z'), not a randomly initialized latent optimized through Eq. 7. This does not test whether Eq. 7 converges from a neutral start, nor whether the optimized latent is determined by the observed motion. Please add a baseline with the same optimization procedure but random N(0,I) initialization, ideally with multiple restarts, to support the specificity claim.
  3. [Tables 1-4 and Sec. 5.1-5.2] All reported metrics are L2 trajectory reconstruction errors. None measures whether the optimized latent recovers the correct constitutive behavior on held-out deformations, e.g., stress prediction error on unseen deformation gradients, projected deformation gradient error, or decoding of the latent to constitutive parameters such as Young's modulus, yield stress, or friction angle. Position-only matching can in principle overfit the observed trajectory, so trajectory error alone does not establish that z corresponds to the true material. Please add parameter-level or state-level validation of the inferred latent.
  4. [Tables 1-4] No error bars, standard deviations, or significance tests are reported; Table 3 says 'over 10 trajectories' but without variance. The claimed improvements over baselines (e.g., Table 1 'ours' vs 'nclaw') cannot be assessed for statistical reliability. Please report mean and standard deviation over multiple seeds or test splits, and state the number of runs for each table.
  5. [Sec. 9 (Supplementary) and dataset description in Sec. 5] The networks are trained on trajectories generated by the same analytical constitutive laws and parameter ranges described in Sec. 9, and the test distribution appears to be drawn from the same generative process. It is unclear whether the test trajectories in Tables 1-2 involve material parameters held out from training; the 'unseen' settings in Table 2 refer to time horizon, velocity, and geometry, not necessarily unseen material parameters. To support generalization, please clarify the train/test split over trajectories and material parameters, and evaluate on material parameters not used in training (e.g., interpolation or extrapolation within the ranges).
minor comments (6)
  1. [Eq. (5)] The loss function L in Eq. (5) is not defined; please specify the norm (e.g., squared L2) and the summation ranges over particles and timesteps.
  2. [Algorithm 1 (Supplementary)] Algorithm 1 lists the input dataset as D = {F, Fproj, C, S} but line 4 samples Cp,t_i, and line 6 passes U, V^T to gϕ, whereas Eq. (3) of the main paper defines gϕ(F, z). Please align the notation between the main paper and the algorithm.
  3. [Table 1] Table 1 omits non-Newtonian materials even though the abstract and Section 5 mention them; please include non-Newtonian in the main reconstruction table or state explicitly why it is deferred to Table 3.
  4. [Figure 5 caption] The phrase 'Random latent' in Figure 5 is ambiguous; please specify whether it is a latent from another trajectory or a randomly initialized and optimized latent.
  5. [Table 4] The latent size ablation covers 4, 32, and 256 but not intermediate sizes; please justify the choice or extend the sweep to give a clearer trend.
  6. [General notation] The terms 'projected deformation gradient' and 'deformation gradient projection' are used interchangeably; please define once and use a single term consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the training objective (Eq. 5) fits latent-conditioned networks to ground-truth deformation-gradient projections and stresses, while the inference objective (Eq. 7) optimizes a latent against particle positions; the two are not the same quantity by construction.

full rationale

The paper's derivation chain is not circular. Training (Eq. 5) jointly optimizes g_phi, f_theta, and per-trajectory latents z_n to reproduce ground-truth projected deformation gradients and stresses produced by standard analytical constitutive laws (Supplementary Section 9). Inference (Eq. 6-7) freezes those networks and optimizes a scene-specific latent against observed particle positions through the differentiable MPM simulator. The inference loss is a position-matching objective, not a stress or deformation-projection objective, so the recovered latent is not the training loss minimizer by construction. The random-latent control in Table 3 provides an independent check that the optimized latent carries trajectory-specific material content rather than being an arbitrary point in latent space. One in-scope concern is a protocol inconsistency: Section 4 says the inference latent is 'randomly initialize[d] ... from N(0,I)', while Section 5 Implementation Details says it is initialized 'from the best performing cluster centers found using K-means clustering on the learned latent space'. This could mean the reported material inference is warm-started by training-derived material clusters, which is a potential leakage or validity issue, but it is not circular because the final latent is still obtained by minimizing Eq. 7. No load-bearing self-citation occurs: the key baseline [39] has disjoint authors, and the paper compares against external methods (NCLaw, spline, neural, gnn). The Discussion candidly states limitations (non-homogeneous scenes, need for known initial geometry and 3D motion), which are scope restrictions rather than hidden circular assumptions. Overall, the central inference result is not equivalent to its inputs by construction, so no circularity is identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the simulator's analytical constitutive laws serving as ground truth, on the capacity of a single latent-conditioned MLP pair to approximate all materials, and on the assumption that train and test trajectories come from the same distribution. The learned latent is the main fitted entity; no new physical laws or particles are introduced.

free parameters (3)
  • per-trajectory latent z_n = 32-dimensional vector per trajectory
    Each training trajectory is assigned a latent optimized jointly with network weights to reproduce its deformation projection and stress (Eq. 5). At inference the latent is optimized to match observed positions, so it is the main fitted quantity in both stages.
  • latent regularization coefficient 1/sigma^2 = not reported
    Controls the norm of latent vectors in Eq. 5 and is chosen by hand; the exact value is not given in the paper.
  • K-means cluster centers for inference initialization = not specified (number of clusters not stated)
    The implementation initiates inference latent from the best-performing cluster centers of the learned latent space. The number of clusters and the selection procedure are not detailed, and this choice is inconsistent with the random initialization described in Sec. 4.
assumptions (4)
  • domain assumption MPM with the listed analytical constitutive laws (neo-Hookean, StVK with von Mises return, Drucker-Prager, viscoplastic) generates ground-truth trajectories and internal states.
    All training and evaluation data come from this simulator; the network is never tested against real-world data.
  • ad hoc to paper A single latent-conditioned MLP pair has enough capacity to approximate the deformation projection and constitutive functions of all materials across their parameter ranges.
    Architectural assumption; no capacity or universality argument is provided beyond empirical results.
  • domain assumption The test trajectories (geometries, velocities, extended horizons) lie within the distribution spanned by the training simulator and parameter ranges.
    Generalization is assessed against the same simulator; out-of-distribution behavior is untested.
  • ad hoc to paper The learned latent space is sufficiently structured that K-means initialization from training latents transfers to test trajectories.
    Inference quality depends on this initialization; the main text says random initialization while the implementation uses cluster centers.
invented entities (1)
  • material latent vector z (32-dimensional)
    purpose: Encodes the deformation projection behavior and constitutive law of a trajectory's material in one code shared by both networks.
    The latent is fit to training data and test trajectories; no external physical quantity is predicted, so it has no falsifiable handle outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniPhy: Learning a Unified Constitutive Model for Inverse Physics Simulation." pith.science (2026). https://pith.science/paper/ANKRT5R3

@misc{pith2026250516971,
  author       = {Pith},
  title        = {Pith review of: UniPhy: Learning a Unified Constitutive Model for Inverse Physics Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ANKRT5R3}},
  note         = {Machine review of arXiv:2505.16971}
}
read the original abstract

We propose UniPhy, a common latent-conditioned neural constitutive model that can encode the physical properties of diverse materials. At inference UniPhy allows `inverse simulation' i.e. inferring material properties by optimizing the scene-specific latent to match the available observations via differentiable simulation. In contrast to existing methods that treat such inference as system identification, UniPhy does not rely on user-specified material type information. Compared to prior neural constitutive modeling approaches which learn instance specific networks, the shared training across materials improves both, robustness and accuracy of the estimates. We train UniPhy using simulated trajectories across diverse geometries and materials -- elastic, plasticine, sand, and fluids (Newtonian & non-Newtonian). At inference, given an object with unknown material properties, UniPhy can infer the material properties via latent optimization to match the motion observations, and can then allow re-simulating the object under diverse scenarios. We compare UniPhy against prior inverse simulation methods, and show that the inference from UniPhy enables more accurate replay and re-simulation under novel conditions.

Figures

Figures reproduced from arXiv: 2505.16971 by the authors.

Figure 1
Figure 1. We present UniPhy, a unified latent-conditioned neural model which learns a common latent space to encode the properties of diverse materials. At inference, given motion observations for a system with unknown material parameters, UniPhy allows material inference via differentiable simulation-based latent optimization. These inferred material latents can be used to simulate new trajectories that reflect the behavior … view at source ↗
Figure 2
Figure 2. Overview (a) We use the Material Point Method simulator to generate a dataset with various geometries, motions, and material parameters (e.g. Young’s modulus) as represented in the parameter grid. The dataset comprises of the particle to node deformation gradient matrix, projected deformation gradient matrix, and stress matrix. The visualizations show some example trajectories of different materials. (b) Next, we us… view at source ↗
Figure 3
Figure 3. We show the inference setup in the figure where given an object trajectory along with a random latent, we do not have the material [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: We show our results (w/ and w/o teacher forcing (TF)) in columns four and three respectively compared with NCLaw (w/ and w/o [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative Results on non-Newtonian and Elastic ma [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Pixie: Fast and Generalizable Supervised Learning of 3D Physics from Pixels

    cs.CV 2025-08 reject novelty 5.0 of 10

    A supervised 3D U-Net predicts per-voxel material fields from CLIP feature grids, enabling fast MPM-based animation, but the reported evidence depends on pseudo-labels and a VLM judge from the same model family as the...

Reference graph

Works this paper leans on

64 extracted references · 61 canonical work pages · cited by 1 Pith paper

  1. [1]

    A three-dimensional constitutive model for the large stretch behavior of rubber elastic materials

    Ellen M Arruda and Mary C Boyce. A three-dimensional constitutive model for the large stretch behavior of rubber elastic materials. Journal of the Mechanics and Physics of Solids, 41(2), 1993. 2

    work page 1993

  2. [2]

    The material-point method for granular materials

    SG Bardenhagen, JU Brackbill, and Deborah Sulsky. The material-point method for granular materials. Computer methods in applied mechanics and engineering , 187(3-4),

    work page

  3. [3]

    Combining differentiable pde solvers and graph neu- ral networks for fluid flow prediction

    Filipe De Avila Belbute-Peres, Thomas Economon, and Zico Kolter. Combining differentiable pde solvers and graph neu- ral networks for fluid flow prediction. In ICML, 2020. 3

    work page 2020

  4. [4]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In ECCV, 2020. 2

    2020

  5. [5]

    Vir- tual elastic objects

    Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne V ouga, and Christoph Lassner. Vir- tual elastic objects. In CVPR, 2022. 2

    work page 2022

  6. [6]

    Bubbles, drops, and particles in non- Newtonian fluids

    Raj P Chhabra. Bubbles, drops, and particles in non- Newtonian fluids. 2006. 2, 14

    work page 2006

  7. [7]

    End-to-end differen- tiable physics for learning and control

    Filipe de Avila Belbute-Peres, Kevin Smith, Kelsey Allen, Josh Tenenbaum, and J Zico Kolter. End-to-end differen- tiable physics for learning and control. NeurIPS, 31, 2018. 2

    work page 2018

  8. [8]

    A differentiable physics engine for deep learning in robotics

    Jonas Degrave, Michiel Hermans, Joni Dambre, and Francis Wyffels. A differentiable physics engine for deep learning in robotics. Frontiers in neurorobotics, 13, 2019

    work page 2019

Show all 64 references
  1. [9]

    Functional optimization of flu- idic devices with differentiable stokes flow

    Tao Du, Kui Wu, Andrew Spielberg, Wojciech Matusik, Bo Zhu, and Eftychios Sifakis. Functional optimization of flu- idic devices with differentiable stokes flow. ACM Transac- tions on Graphics (TOG), 39(6), 2020

    2020

  2. [10]

    Diffpd: Differentiable projective dynamics

    Tao Du, Kui Wu, Pingchuan Ma, Sebastien Wah, Andrew Spielberg, Daniela Rus, and Wojciech Matusik. Diffpd: Differentiable projective dynamics. ACM Transactions on Graphics (TOG), 41(2), 2021. 2

    2021

  3. [11]

    Silly rubber: an implicit material point method for simulating non-equilibrated viscoelastic and elastoplastic solids

    Yu Fang, Minchen Li, Ming Gao, and Chenfanfu Jiang. Silly rubber: an implicit material point method for simulating non-equilibrated viscoelastic and elastoplastic solids. ACM Transactions on Graphics (TOG), 38(4), 2019. 2, 14

    2019

  4. [12]

    Elasticity of soft tissues in simple elongation

    YC Fung. Elasticity of soft tissues in simple elongation. American Journal of Physiology-Legacy Content , 213(6),

  5. [13]

    Add: Analytically differentiable dynamics for multi-body systems with frictional contact

    Moritz Geilinger, David Hahn, Jonas Zehnder, Moritz B¨acher, Bernhard Thomaszewski, and Stelian Coros. Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG), 39(6), 2020. 2

    2020

  6. [14]

    Deformable part models are convolutional neural net- works

    Ross Girshick, Forrest Iandola, Trevor Darrell, and Jitendra Malik. Deformable part models are convolutional neural net- works. In CVPR, 2015. 2

    2015

  7. [15]

    Forward flow for novel view synthesis of dynamic scenes

    Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiao- qing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In ICCV, 2023. 2

    2023

  8. [16]

    Real2sim: Visco-elastic parameter estimation from dynamic motion

    David Hahn, Pol Banzet, James M Bern, and Stelian Coros. Real2sim: Visco-elastic parameter estimation from dynamic motion. ACM Transactions on Graphics (TOG), 38(6), 2019. 2

    2019

  9. [17]

    Learning physics-consistent material behavior without prior knowledge

    Zhichao Han, Mohit Pundir, Olga Fink, and David S Kam- mer. Learning physics-consistent material behavior without prior knowledge. arXiv preprint arXiv:2407.20273, 2024. 2

    2024 arXiv

  10. [18]

    Chainqueen: A real-time differen- tiable physical simulator for soft robotics

    Yuanming Hu, Jiancheng Liu, Andrew Spielberg, Joshua B Tenenbaum, William T Freeman, Jiajun Wu, Daniela Rus, and Wojciech Matusik. Chainqueen: A real-time differen- tiable physical simulator for soft robotics. In ICRA. IEEE,

  11. [19]

    Difftaichi: Differentiable programming for physical simulation

    Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fr´edo Durand. Difftaichi: Differentiable programming for physical simulation. ICLR,

  12. [20]

    Learning constitutive relations from indirect observations us- ing deep neural networks.Journal of Computational Physics, 416, 2020

    Daniel Z Huang, Kailai Xu, Charbel Farhat, and Eric Darve. Learning constitutive relations from indirect observations us- ing deep neural networks.Journal of Computational Physics, 416, 2020. 2

    2020

  13. [21]

    Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes

    Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In CVPR,

  14. [22]

    Plasticinelab: A soft-body manipulation benchmark with differentiable physics

    Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics. ICLR, 2021. 2

    2021

  15. [23]

    The affine particle-in-cell method

    Chenfanfu Jiang, Craig Schroeder, Andrew Selle, Joseph Teran, and Alexey Stomakhin. The affine particle-in-cell method. ACM Transactions on Graphics (TOG) , 34(4),

  16. [24]

    Learning category-specific mesh reconstruc- tion from image collections

    Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruc- tion from image collections. In ECCV, 2018. 2

    2018

  17. [25]

    Physics-informed ma- chine learning

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed ma- chine learning. Nature Reviews Physics, 3(6), 2021. 3

    2021

  18. [26]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. 2

    2023

  19. [27]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In ICCV, 2023. 2

    2023

  20. [28]

    Drucker-prager elastoplasticity for sand animation

    Gergely Kl ´ar, Theodore Gast, Andre Pradhana, Chuyuan Fu, Craig Schroeder, Chenfanfu Jiang, and Joseph Teran. Drucker-prager elastoplasticity for sand animation. ACM Transactions on Graphics (TOG), 35(4), 2016. 2, 12

    2016

  21. [29]

    Polyconvex anisotropic hy- perelasticity with neural networks

    Dominik K Klein, Mauricio Fern ´andez, Robert J Martin, Pa- trizio Neff, and Oliver Weeger. Polyconvex anisotropic hy- perelasticity with neural networks. Journal of the Mechanics and Physics of Solids, 159, 2022. 2

    2022

  22. [30]

    Plasticitynet: Learning to simulate metal, sand, and snow for optimization time integration

    Xuan Li, Yadi Cao, Minchen Li, Yin Yang, Craig Schroeder, and Chenfanfu Jiang. Plasticitynet: Learning to simulate metal, sand, and snow for optimization time integration. NeurIPS, 35, 2022. 2

    2022

  23. [31]

    Pac-nerf: Physics augmented continuum neural ra- diance fields for geometry-agnostic system identification

    Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural ra- diance fields for geometry-agnostic system identification. ICLR, 2023. 2, 3

    2023

  24. [32]

    Dynibar: Neural dynamic image-based rendering

    Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2

    2023

  25. [33]

    Differen- tiable cloth simulation for inverse problems

    Junbang Liang, Ming Lin, and Vladlen Koltun. Differen- tiable cloth simulation for inverse problems. NeurIPS, 32,

  26. [34]

    A learning-based multiscale method and its application to inelastic impact problems

    Burigede Liu, Nikola Kovachki, Zongyi Li, Kamyar Az- izzadenesheli, Anima Anandkumar, Andrew M Stuart, and Kaushik Bhattacharya. A learning-based multiscale method and its application to inelastic impact problems. Journal of the Mechanics and Physics of Solids, 158, 2022. 2

    2022

  27. [35]

    Soft ras- terizer: A differentiable renderer for image-based 3d reason- ing

    Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft ras- terizer: A differentiable renderer for image-based 3d reason- ing. In ICCV, 2019. 2

    2019

  28. [36]

    Dynamic 3d gaussians: Tracking by per- sistent dynamic view synthesis

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by per- sistent dynamic view synthesis. In 3DV. IEEE, 2024. 2

    2024

  29. [37]

    Diffaqua: A differentiable computational design pipeline for soft underwater swimmers with shape interpolation

    Pingchuan Ma, Tao Du, John Z Zhang, Kui Wu, Andrew Spielberg, Robert K Katzschmann, and Wojciech Matusik. Diffaqua: A differentiable computational design pipeline for soft underwater swimmers with shape interpolation. ACM Transactions on Graphics (TOG), 40(4), 2021. 2

    2021

  30. [38]

    Risp: Rendering-invariant state pre- dictor with differentiable simulation and rendering for cross- domain parameter estimation

    Pingchuan Ma, Tao Du, Joshua B Tenenbaum, Wojciech Ma- tusik, and Chuang Gan. Risp: Rendering-invariant state pre- dictor with differentiable simulation and rendering for cross- domain parameter estimation. ICLR, 2022. 2

    2022

  31. [39]

    Learning neural constitutive laws from motion observations for generalizable pde dynamics

    Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik. Learning neural constitutive laws from motion observations for generalizable pde dynamics. In ICML. PMLR, 2023. 2, 3, 6, 7, 8, 12

    2023

  32. [40]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 65(1), 2021. 2

    2021

  33. [41]

    Mechanik der festen k ¨orper im plastisch- deformablen zustand

    R v Mises. Mechanik der festen k ¨orper im plastisch- deformablen zustand. Nachrichten von der Gesellschaft der Wissenschaften zu G ¨ottingen, Mathematisch-Physikalische Klasse, 1913, 1913. 2

    1913

  34. [42]

    Hypernerf: A higher- dimensional representation for topologically varying neural radiance fields

    Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin- Brualla, and Steven M Seitz. Hypernerf: A higher- dimensional representation for topologically varying neural radiance fields. SIGGRAPH Asia, 2021. 2

    2021

  35. [43]

    Learning mesh-based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In ICLR, 2020. 3

    2020

  36. [44]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2

    2021

  37. [45]

    Differentiable simulation of soft multi-body systems

    Yiling Qiao, Junbang Liang, Vladlen Koltun, and Ming Lin. Differentiable simulation of soft multi-body systems. NeurIPS, 34, 2021. 2

    2021

  38. [46]

    Physics-informed neural networks: A deep learning frame- work for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning frame- work for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computa- tional physics, 378, 2019. 3

    2019

  39. [47]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In ICML. PMLR, 2020. 3, 6, 7

    2020

  40. [48]

    Learning parameters and constitutive relationships with physics informed deep neural networks

    Alexandre M Tartakovsky, Carlos Ortiz Marrero, Paris Perdikaris, Guzel D Tartakovsky, and David Barajas-Solano. Learning parameters and constitutive relationships with physics informed deep neural networks. arXiv preprint arXiv:1808.03398, 2018. 2

    2018 arXiv

  41. [49]

    The elasticity of a network of long-chain molecules—ii

    LRG Treloar. The elasticity of a network of long-chain molecules—ii. Transactions of the Faraday Society , 39,

  42. [50]

    Sobolev training of thermodynamic-informed neural networks for interpretable elasto-plasticity models with level set hardening

    Nikolaos N Vlassis and WaiChing Sun. Sobolev training of thermodynamic-informed neural networks for interpretable elasto-plasticity models with level set hardening. Computer Methods in Applied Mechanics and Engineering, 377, 2021. 2

    2021

  43. [51]

    Component-based machine learning paradigm for discovering rate-dependent and pressure-sensitive level-set plasticity models

    Nikolaos N Vlassis and WaiChing Sun. Component-based machine learning paradigm for discovering rate-dependent and pressure-sensitive level-set plasticity models. Journal of Applied Mechanics, 89(2), 2022

    2022

  44. [52]

    Geometric learning for computational mechanics part ii: Graph embedding for interpretable multiscale plasticity

    Nikolaos N Vlassis and WaiChing Sun. Geometric learning for computational mechanics part ii: Graph embedding for interpretable multiscale plasticity. Computer Methods in Ap- plied Mechanics and Engineering, 404, 2023

    2023

  45. [53]

    Geo- metric deep learning for computational mechanics part i: Anisotropic hyperelasticity

    Nikolaos N Vlassis, Ran Ma, and WaiChing Sun. Geo- metric deep learning for computational mechanics part i: Anisotropic hyperelasticity. Computer Methods in Applied Mechanics and Engineering, 371, 2020

    2020

  46. [54]

    Molecular dynamics inferred trans- fer learning models for finite-strain hyperelasticity of mon- oclinic crystals: Sobolev training and validations against physical constraints

    Nikolaos N Vlassis, Puhan Zhao, Ran Ma, Tommy Sewell, and WaiChing Sun. Molecular dynamics inferred trans- fer learning models for finite-strain hyperelasticity of mon- oclinic crystals: Sobolev training and validations against physical constraints. International Journal for N...

    2022

  47. [55]

    Learning elastic constitutive material and damping models

    Bin Wang, Yuanmin Deng, Paul Kry, Uri Ascher, Hui Huang, and Baoquan Chen. Learning elastic constitutive material and damping models. In Computer Graphics Forum. Wiley Online Library, 2020. 2, 6, 7

    2020

  48. [56]

    A multiscale multi- permeability poroplasticity model linked by recursive ho- mogenizations and deep learning

    Kun Wang and WaiChing Sun. A multiscale multi- permeability poroplasticity model linked by recursive ho- mogenizations and deep learning. Computer Methods in Ap- plied Mechanics and Engineering, 334, 2018. 2

    2018

  49. [57]

    Fluid- lab: A differentiable environment for benchmarking com- plex fluid manipulation

    Zhou Xian, Bo Zhu, Zhenjia Xu, Hsiao-Yu Tung, Antonio Torralba, Katerina Fragkiadaki, and Chuang Gan. Fluid- lab: A differentiable environment for benchmarking com- plex fluid manipulation. ICLR, 2023. 2

    2023

  50. [58]

    Physgaussian: Physics- integrated 3d gaussians for generative dynamics

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics- integrated 3d gaussians for generative dynamics. In CVPR,

  51. [59]

    Nonlinear material design using principal stretches

    Hongyi Xu, Funshing Sin, Yufeng Zhu, and Jernej Barbi ˇc. Nonlinear material design using principal stretches. ACM Transactions on Graphics (TOG), 34(4), 2015. 6, 7

    2015

  52. [60]

    Continuum foam: A material point method for shear-dependent flows

    Yonghao Yue, Breannan Smith, Christopher Batty, Changxi Zheng, and Eitan Grinspun. Continuum foam: A material point method for shear-dependent flows. ACM Transactions on Graphics (TOG), 34(5), 2015. 2, 8, 14

    2015

  53. [61]

    Physdreamer: Physics-based interac- tion with 3d objects via video generation

    Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interac- tion with 3d objects via video generation. In ECCV, 2025. 2 UniPhy: Learning a Unified Constitutive Model for Inverse Phys...

    2025

  54. [62]

    The ground truth state includes the position x, velocity v, affine velocity C, and deformation gradient F

    Implementation Details Teacher Forcing : The baseline, NCLaw [39] uses a teacher-forcing scheme that restarts the predicted simula- tion from ground truth state periodically. The ground truth state includes the position x, velocity v, affine velocity C, and deformation gradien...

  55. [63]

    Algorithm We detail our training and inference algorithm in Algo- rithm 1 and Algorithm 2 respectively where i is the trajec- tory index, p represents the particle index and t represents the time

  56. [64]

    In Material Point Method (MPM), each particle has a deformation gradient F which is projected on to the yield surface using a return mapping G

    Analytical Constitutive Laws In this section, we discuss the constitutive model and the deformation gradient projection/mapping function for the materials that are used to simulate the trajectories used in training. In Material Point Method (MPM), each particle has a deformati...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.