DEM-NeRF: A Neuro-Symbolic Method for Scientific Discovery through Physics-Informed Simulation

Alvaro Velasquez; Houbing Song; Wenkai Tan

REVIEW 3 major objections 6 minor 33 references

DEM-NeRF: A Neuro-Symbolic Method for Scientific Discovery through Physics-Informed Simulation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A neural radiance field combined with a physics-driven network reconstructs a 3D object and predicts its elastic deformation directly from images.

desk verdict A plausible but unevaluated swap of MPM for DEM inside a NeRF+physics pipeline; the central accuracy claim is unsupported and the dynamics language overreaches. read the letter →

arxiv 2507.21350 v1 pith:E4I6R3L2 submitted 2025-07-28 cs.LG

Wenkai Tan , Alvaro Velasquez , Houbing Song This is my paper · ORCID

classification cs.LG

keywords neuro-symbolicNeuralRadianceFieldPhysics-InformedNetworkDeepEnergyMethodhyperelasticdeformationsolidreconstructionreal-timesimulationNeo-Hookean

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

The paper proposes DEM-NeRF, a neuro-symbolic pipeline that reconstructs a deformable solid from sparse multi-view images and then predicts its deformation under load, without explicit geometry, labeled data, or specialized hardware. The reconstruction stage uses a neural radiance field to turn images into a 3D particle cloud, and the simulation stage uses a physics-informed network that minimizes the total potential energy of a compressible Neo-Hookean (a standard hyperelastic) solid. The authors claim this runs in about 45 seconds of NeRF training, 20 seconds of DEM training, and roughly one second per prediction, which would make interactive real-time elastic simulation practical. The attraction is that it sidesteps the usual bottleneck of numerical solvers, which require complete geometric models and known boundary conditions.

What carries the argument

The load-bearing mechanism is the Deep Energy Method, in which a neural network acts as a global shape function mapping each particle's rest position $\mathbf{X}$ to a displacement $\mathbf{u}$, so the deformation gradient and the strain energy are computed from first-order derivatives of the network output. Instead of enforcing the strong-form equilibrium equation as standard physics-informed neural networks do, DEM minimizes the total potential energy $\Pi$ of Eq. (9) — strain energy minus the work of body forces and tractions — together with a boundary-condition mean-squared-error term, which the authors argue needs only first-order differentiation and therefore converges faster. Around this sit the NGP-NeRF reconstruction, which supplies the initial geometry, and the particle-sampling step (random, Poisson disc, or uniform mesh), which converts that geometry into the DEM training domain.

What would settle it

Recreate the T-bar experiment with the same multi-view images and independently measure the actual 3D displacement field (for example with digital image correlation or tracked markers), then compare it with DEM-NeRF's prediction using the paper's unreported material and load parameters; a mismatch beyond rendering noise, or any observable time-dependent motion that the quasi-static energy minimization cannot represent, would undercut the 'reliable deformation prediction' claim.

Watch

Extended reading notes

Core claim

DEM-NeRF's central claim is that coupling NGP-NeRF for image-based geometry with a Deep Energy Method (DEM) network for quasi-static Neo-Hookean elasticity produces high-quality 3D reconstruction and reliable deformation prediction directly from image sequences, in hyperelastodynamic and loading scenarios. The pipeline reconstructs the undeformed T-bar as a mesh, samples it into a particle cloud, applies a fixed base and a torque formed by two opposing forces, and trains the DEM network by minimizing the total potential energy of the solid plus a boundary-condition penalty. The predicted displacement field is fed back into the NeRF renderer to generate new views of the deformed object. The paper reports DEM training of about 20 seconds and prediction of about 1 second, against roughly 600 seconds for the MPM-based PAC-NeRF baseline and 1,140 seconds for full-order FEM, and calls DEM-NeRF the first neuro-symbolic method for real-time solid deformation reconstruction and prediction.

Load-bearing premise

The whole pipeline assumes the T-bar's deformation is exactly a compressible Neo-Hookean quasi-static response with manually chosen Young's modulus, Poisson's ratio, and torque loads, and none of those values are reported in the paper.

Editorial extensions

If this is right

Real-time interactive simulation of elastic solids becomes feasible from ordinary camera images, since the reported per-prediction cost is about one second after short training.
The method removes the need for explicit geometric models, manually specified meshes, or known hard-boundary conditions, which are the main obstacles to applying FEM and BEM to real scenes.
Because it uses images plus physics rather than labeled displacement data, it can be applied to new objects and loading setups without collecting ground-truth deformation fields.
If the quasi-static Neo-Hookean assumption holds, the same pipeline extends naturally to any object whose material can be described by a strain energy function, including soft robotics and biomedical tissue models.

Reading between the lines

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

The authors stop short of reporting any quantitative accuracy metric (error against measured displacement or rendered image); a head-to-head error comparison with PAC-NeRF and FEM would be the natural next test of the 'reliable prediction' claim.
The same architecture can be turned into an inverse method: by treating Young's modulus and Poisson's ratio as trainable parameters, DEM-NeRF could identify material properties from video alone, something the paper mentions as motivation but does not implement.
The abstract's 'hyperelastodynamics' suggests time-dependent behavior, but Eq. (9) is a quasi-static energy minimization with no inertial terms; adding kinetic energy and time stepping would be needed to justify that label.
A testable extension is to demonstrate the pipeline on a second object with a genuinely complex boundary (e.g., a teapot) using Poisson disc sampling, which the paper argues should outperform uniform sampling but does not evaluate.

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.

Desk Editor's Note

read the letter

The paper takes the PAC-NeRF/PIE-NeRF recipe—reconstruct a solid with NeRF, then simulate deformation with a physics-based neural solver—and swaps MPM for the deep energy method (DEM). That specific combination is new, and the DEM formulation is described cleanly: only first-order derivatives are needed, which is a real computational advantage over strong-form PINNs. If the speed numbers hold, the pipeline is also much faster to train than PAC-NeRF (~20s DEM training vs ~600s MPM), which is a meaningful selling point.

That's where the credit ends. The paper never demonstrates that the predicted deformation is accurate. The only quantitative results are a runtime table. There is no displacement error, no strain error, no comparison against the ground-truth deformed T-bar in Figure 6, and no accuracy comparison against PAC-NeRF or PIE-NeRF. The central claim of "reliable deformation predictions" is simply asserted. This is a load-bearing absence: if the simulation is wrong, the whole pipeline is just a NeRF with a fancy solver.

The modeling frame also overreaches. The abstract promises hyperelastodynamics, but the FDEM network in Eq. (1) has no time input, and Eq. (9) is a quasi-static energy minimization without inertial terms. The experiment is a static torque load. That's fine as a quasi-static showcase, but the paper should not claim dynamic capability. Relatedly, the material parameters (E, ν) and the torque magnitudes are never reported, so the experiment is not reproducible, and the loss weight W_u is unspecified.

The novelty is incremental, as the reader notes, but that alone wouldn't sink it. The sinker is the missing evaluation.

Given the internal contradiction between the stated scope and the actual model, and the lack of any accuracy metric, I don't think this paper is ready for publication. But the core idea—DEM as the physics constraint inside a NeRF pipeline—is worth exploring, and with a proper evaluation (error against ground truth, material parameter reporting, and a time-dependent extension if they want to claim dynamics) it could become a solid contribution. I'd send it to review rather than desk-reject, because the combination is new and the community could benefit from a careful critique. My own verdict on the current version: reject and resubmit after major revision.

Referee Report

3 major / 6 minor

Summary. DEM-NeRF proposes a neuro-symbolic pipeline that reconstructs a 3D solid from multi-view images with NGP-NeRF, samples it into particles, and predicts deformation with a Deep Energy Method network minimizing a Neo-Hookean strain-energy functional. The paper claims high-quality reconstructions and reliable deformation predictions in hyperelastodynamic and loading scenarios, with a runtime advantage over PAC-NeRF and FEM. The experimental section applies the pipeline to a T-bar under torque, but reports only computational and rendering times together with qualitative renderings; no quantitative accuracy metrics, material parameters, or load values are provided.

Significance. If substantiated, the direction is significant: combining NeRF-based geometry acquisition with an energy-constrained network could enable real-time, label-free deformation simulation for digital twins, soft robotics, and biomechanics. The paper has two genuine strengths: the DEM loss is an explicit physical energy minimization rather than a fitted surrogate, and the runtime figures (about 20 s DEM training plus 1 s prediction) are potentially attractive. However, the central claims of reliable prediction and hyperelastodynamic capability are not backed by any quantitative evaluation in the current manuscript. The evidence is limited to images and a timing table, so the contribution, as presented, is not yet established.

major comments (3)

[Section IV-A, Table IV-A] The paper reports only computational and rendering times; there is no displacement error, strain error, or any quantitative comparison of the predicted T-bar deformation against the ground-truth deformed shape in Figure 6 or against FEM, PAC-NeRF, or PIE-NeRF. Since the abstract and Section I claim "reliable deformation predictions" and "high-quality reconstructions," the absence of any accuracy metric leaves the central contribution unsubstantiated. The lack of ablations for the boundary weight Wu, particle count, and sampling method compounds this problem.
[Section III-D, Eqs. (1), (9)-(14)] The total potential energy Π in Eq. (9) is minimized as a quasi-static equilibrium, with no kinetic-energy or inertial term, and FDEM in Eq. (1) maps (x,y,z) directly to u with no time coordinate. Thus the architecture cannot represent the "hyperelastodynamics" or the "spatiotemporal representation" claimed in the abstract and Impact Statement. A model that assumes quasi-static Neo-Hookean equilibrium is not a dynamical model, and the claim of dynamic prediction is not supported by the stated equations.
[Section IV-A, Eqs. (7)-(9)] The Young's modulus E, Poisson's ratio ν, torque magnitudes, body force fb, and boundary weight Wu are never reported. These quantities determine the predicted displacement field through the energy in Eq. (9); without them the experiment is not reproducible, and if they were hand-tuned to match the observed deformation, that tuning should be disclosed. The manuscript provides no evidence that the assumed compressible Neo-Hookean model with the chosen loads corresponds to the physical T-bar experiment.

minor comments (6)

[References] References [4] and [26] are the same Karniadakis et al. paper, and references [5] and [25] are the same Raissi et al. paper; the duplicate entries should be consolidated.
[Section III-A] The text contains "particle-could representation," which should be "particle-cloud representation."
[Eq. (14)] The traction boundary set is written with "NP i" and the traction term uses t, while Eq. (9) writes \bar{t}; these notations should be made consistent and the typo corrected.
[Section IV-A] The manuscript mentions "I.C./B.C." but initial conditions are never defined or used in the quasi-static formulation; the abbreviation should be removed or the initial conditions specified.
[Section IV-A, Table] The performance table has no caption and no table number; it should be given a proper number and a caption defining the hardware, the meaning of training time, and which stages are included in each reported number.
[Section IV-B] The comparison of random, Poisson disc, and uniform mesh sampling in Figure 7 does not state the particle spacing or radius used for each method; specifying these parameters would make the qualitative comparison meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the deformation output is produced by a forward energy-minimization solve under stated constitutive and loading assumptions, not by fitting the target deformation or by a self-referential uniqueness claim.

full rationale

The paper's central derivation chain is a forward physics-constrained simulation. Equation (1) defines FDEM as a map from particle coordinates to displacement; Section III-D defines the total potential energy Π in Eq. (9) from a compressible Neo-Hookean strain energy Eq. (7) and external force/traction terms; Eq. (13) obtains the network parameters by minimizing Π plus a boundary-condition MSE. The predicted displacement field is therefore the minimizer of an energy functional whose material parameters (E, ν) and loads are inputs, not quantities fitted to match the deformed ground truth of Figure 6. Nothing in the paper's equations defines the predicted output in terms of the measured deformation or in terms of its own conclusion; the rendered deformed object in Figure 5 comes from applying the DEM solution to the NeRF-reconstructed geometry. The paper does not invoke a uniqueness theorem from the authors' prior work, and all cited physics methods (DEM, PINN, PAC-NeRF, PIE-NeRF) are external prior work, not self-citations carrying the argument. The lack of any quantitative displacement-error comparison against the ground-truth deformed T-bar, and the mismatch between the claimed 'hyperelastodynamics' and the quasi-static formulation in Eq. (9), are serious validation and correctness concerns, but they are not circularity: the model is self-contained against external benchmarks in principle. Accordingly, the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The claims rest entirely on manually set material and load parameters, which are never specified, and on standard continuum-mechanics assumptions that are not checked against data. No new entities are introduced.

free parameters (5)

Young's modulus E = not reported
Material parameter in the Neo-Hookean strain energy (Eq. 8); chosen by hand for the T-bar simulation, directly controls the predicted deformation and is never stated.
Poisson's ratio ν = not reported
Second Lamé parameter input in Eq. (8); chosen by hand, not reported.
applied torque magnitude = not reported
Two opposing forces on the T-bar (Section IV-A) define the traction boundary condition; the magnitude is not given and is necessary to reproduce the deformation.
loss weight Wu for boundary condition = not reported
Weight in the DEM total loss Eq. (13); no value or tuning procedure is reported.
particle count / sampling density = not reported
Uniform mesh sampling in Section IV-A; the number of particles affects both geometry fidelity and DEM accuracy and is not stated.

assumptions (5)

domain assumption The solid is a homogeneous, isotropic, compressible Neo-Hookean hyperelastic material (Eq. 7).
Used throughout Section III-D as the constitutive model; no evidence is given that the T-bar behaves this way.
domain assumption Quasi-static equilibrium is sufficient: inertial effects are neglected and the deformation minimizes total potential energy (Eq. 9).
Section III-D minimizes Π to enforce equilibrium, yet the abstract claims hyperelastodynamics; this tension is never resolved.
domain assumption The NeRF-reconstructed mesh and the sampled particle cloud faithfully represent the true object geometry and its boundary.
Section III-A uses NGP-NeRF mesh conversion and sampling without any quantitative check of reconstruction error.
standard math The network can represent the exact displacement field and minimize the energy to a global minimum.
Relies on the universal approximation theorem and the DEM framework from Nguyen-Thanh et al. [31]; no convergence or accuracy analysis is given.
standard math The Dirichlet boundary condition can be imposed through the MSE loss term in Eq. (15).
A standard PINN boundary soft-constraint; used to fix the bottom of the T-bar.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DEM-NeRF: A Neuro-Symbolic Method for Scientific Discovery through Physics-Informed Simulation." pith.science (2026). https://pith.science/paper/E4I6R3L2

@misc{pith2026250721350,
  author       = {Pith},
  title        = {Pith review of: DEM-NeRF: A Neuro-Symbolic Method for Scientific Discovery through Physics-Informed Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E4I6R3L2}},
  note         = {Machine review of arXiv:2507.21350}
}

read the original abstract

Neural networks have emerged as a powerful tool for modeling physical systems, offering the ability to learn complex representations from limited data while integrating foundational scientific knowledge. In particular, neuro-symbolic approaches that combine data-driven learning, the neuro, with symbolic equations and rules, the symbolic, address the tension between methods that are purely empirical, which risk straying from established physical principles, and traditional numerical solvers that demand complete geometric knowledge and can be prohibitively expensive for high-fidelity simulations. In this work, we present a novel neuro-symbolic framework for reconstructing and simulating elastic objects directly from sparse multi-view image sequences, without requiring explicit geometric information. Specifically, we integrate a neural radiance field (NeRF) for object reconstruction with physics-informed neural networks (PINN) that incorporate the governing partial differential equations of elasticity. In doing so, our method learns a spatiotemporal representation of deforming objects that leverages both image supervision and symbolic physical constraints. To handle complex boundary and initial conditions, which are traditionally confronted using finite element methods, boundary element methods, or sensor-based measurements, we employ an energy-constrained Physics-Informed Neural Network architecture. This design enhances both simulation accuracy and the explainability of results.

Figures

Figures reproduced from arXiv: 2507.21350 by the authors.

**Figure 1.** The pipeline of DEM-NeRF. B. Neural Scene Representation for Solid Building upon the universal approximation theorem, our approach utilizes neural networks to characterize the spatial properties of a deformable object. Concretely, we introduce two neural networks, FNeRF and FDEM in Eq. 1, which approximate the continuous radiance field and the solid’s displacement field, respectively. FNeRF : (x, y, z) → (c, σ), FDE… view at source ↗

**Figure 2.** Deep Energy Method overview. MSEu({Xi u} Nu i=1, Θ) = 1 Nu X Nu i=1 |uˆ(Xi u , Θ) − u(Xi u )| 2 . (15) IV. EVALUATION AND DISCUSSION A. Experiment In this experiment, we mainly utilized the T bar dataset as a showcase. At the default setup, the NeRF can capture not only the texture and location of the targeted object but also the background. Therefore, there are two ways we can extract clear structure from the NeRF.… view at source ↗

**Figure 4.** T bar with training domain (blue), Dirichlet Boundary Condition [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Figures from the paper (4 more)

**Figure 3.** Figure 3: Object reconstruction from undeformed multi-view dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

**Figure 5.** Figure 5: NeRF rendering from DEM output [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 7.** Figure 7: Particle sampling methods on Mesh. The equilibrium equation is defined as DivP + fb = 0, in B (17) with displacement and traction boundary condition u = ¯u on Γu, P · N = t¯ on Γt. (18) [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Physics-Informed Neural Network with Quasi-Static Condition. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 32 canonical work pages

[1]

Computational discovery of scientific knowledge,

S. D ˇzeroski, P. Langley, and L. Todorovski, “Computational discovery of scientific knowledge,” in Computational Discovery of Scientific Knowl- edge: Introduction, Techniques, and Applications in Environmental and Life Sciences, Springer, 2007, pp. 1–14

work page 2007
[2]

Simulations of subatomic many-body physics on a quantum frequency processor,

H.-H. Lu et al. , “Simulations of subatomic many-body physics on a quantum frequency processor,” Phys. Rev. A , vol. 100, no. 1, Art. no. 012320, Jul. 2019

work page 2019
[3]

Time evolution of galaxy scaling relations in cosmological simulations,

P. Taylor and C. Kobayashi, “Time evolution of galaxy scaling relations in cosmological simulations,” Mon. Not. R. Astron. Soc. , vol. 463, no. 3, pp. 2465–2479, Nov. 2016

work page 2016
[6]

A neuro-symbolic approach to the logic of scientific discov- ery,

I. V otsis, “A neuro-symbolic approach to the logic of scientific discov- ery,” in Proc. Int. Conf. Model-Based Reasoning , 2023, pp. 306–330

work page 2023
[7]

Quarteroni and A

A. Quarteroni and A. Valli, Numerical Approximation of Partial Differ- ential Equations. Berlin, Germany: Springer, 2008

work page 2008
[8]

A finite mass based method for Vlasov– Poisson simulations,

D. J. Larson and C. V . Young, “A finite mass based method for Vlasov– Poisson simulations,” J. Comput. Phys. , vol. 284, pp. 171–185, Feb. 2015. 8 IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, MONTH 2020

work page 2015
[9]

XSEDE: accelerating scientific discovery,

J. Towns et al. , “XSEDE: accelerating scientific discovery,” Comput. Sci. Eng., vol. 16, no. 5, pp. 62–74, Sep./Oct. 2014

work page 2014
[10]

The opportunities, challenges, and risks of high performance computing in computational science and engineering,

D. E. Post, R. P. Kendall, and R. F. Lucas, “The opportunities, challenges, and risks of high performance computing in computational science and engineering,” Adv. Comput., vol. 66, pp. 239–301, 2006

work page 2006

Show all 33 references

[11]

Surveying neuro-symbolic approaches for reliable artificial intelligence of things,

Z. Lu, I. Afridi, H. J. Kang, I. Ruchkin, and X. Zheng, “Surveying neuro-symbolic approaches for reliable artificial intelligence of things,” J. Rel. Intell. Environ. , vol. 10, no. 3, pp. 257–279, Sep. 2024

2024
[12]

Neuro- symbolic approaches in artificial intelligence,

P. Hitzler, A. Eberhart, M. Ebrahimi, M. K. Sarker, and L. Zhou, “Neuro- symbolic approaches in artificial intelligence,” Natl. Sci. Rev., vol. 9, no. 6, Jun. 2022, Art. no. nwac035

2022
[13]

Taichi: A language for high-performance computation on spatially sparse data structures,

Y . Hu, T.-M. Li, L. Anderson, J. Ragan-Kelley, and F. Durand, “Taichi: A language for high-performance computation on spatially sparse data structures,” ACM Trans. Graph., vol. 38, no. 6, pp. 1–16, Dec. 2019

2019
[14]

R. L. Taylor and O. C. Zienkiewicz, The Finite Element Method. Oxford, UK: Butterworth-Heinemann, 2013

2013
[15]

C. A. Brebbia and J. Dominguez, Boundary Elements: An Introductory Course. Southampton, UK: WIT Press, 1994

1994
[16]

NeRF: Representing scenes as neural radiance fields for view synthesis,

B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” Commun. ACM, vol. 65, no. 1, pp. 99–106, Jan. 2021

2021
[17]

D- NeRF: Neural radiance fields for dynamic scenes,

A. Pumarola, E. Corona, G. Pons-Moll, and F. Moreno-Noguer, “D- NeRF: Neural radiance fields for dynamic scenes,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 10318–10327

2021
[18]

Physics-informed neural fields for smoke reconstruction with sparse data,

M. Chu, L. Liu, Q. Zheng, E. Franz, H.-P. Seidel, C. Theobalt, and R. Zayer, “Physics-informed neural fields for smoke reconstruction with sparse data,” ACM Trans. Graph., vol. 41, no. 4, pp. 1–14, Jul. 2022

2022
[19]

PAC-NeRF: Physics augmented continuum neural radi- ance fields for geometry-agnostic system identification,

X. Li, Y .-L. Qiao, P. Y . Chen, K. M. Jatavallabhula, M. Lin, C. Jiang, and C. Gan, “PAC-NeRF: Physics augmented continuum neural radi- ance fields for geometry-agnostic system identification,” arXiv preprint arXiv:2303.05512, 2023

2023 arXiv
[20]

M. A. Sutton, J. J. Orteu, and H. Schreier, Image Correlation for Shape, Motion and Deformation Measurements: Basic Concepts, Theory and Applications. Berlin, Germany: Springer, 2009

2009
[21]

Schumann and M

W. Schumann and M. Dubas, Holographic Interferometry: From the Scope of Deformation Analysis of Opaque Bodies . Berlin, Germany: Springer, 2013

2013
[22]

RGB-D-DIC technique for low-cost 3D displacement fields measure- ments,

J. M. Franco, J. M. Caicedo, J. Marulanda, M. Sutton, and P. Thomson, “RGB-D-DIC technique for low-cost 3D displacement fields measure- ments,” Eng. Struct., vol. 197, Oct. 2019, Art. no. 109457

2019
[23]

3D displacement mea- surement using a single-camera and mesh deformation neural network,

Y . Shao, L. Li, J. Li, Q. Li, S. An, and H. Hao, “3D displacement mea- surement using a single-camera and mesh deformation neural network,” Eng. Struct., vol. 318, Apr. 2024, Art. no. 118767

2024
[24]

Artificial neural networks for solving ordinary and partial differential equations,

I. E. Lagaris, A. Likas, and D. I. Fotiadis, “Artificial neural networks for solving ordinary and partial differential equations,” IEEE Trans. Neural Netw., vol. 9, no. 5, pp. 987–1000, Sep. 1998

1998
[25]

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

M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” J. Comput. Phys., vol. 378, pp. 686–707, Feb. 2019

2019
[26]

Physics-informed machine learning,

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,” Nat. Rev. Phys., vol. 3, no. 6, pp. 422–440, Jun. 2021

2021
[27]

Deep learning of vortex-induced vibrations,

M. Raissi, Z. Wang, M. S. Triantafyllou, and G. E. Karniadakis, “Deep learning of vortex-induced vibrations,” J. Fluid Mech. , vol. 861, pp. 119–137, Jan. 2019

2019
[28]

NSFnets (Navier–Stokes flow nets): Physics-informed neural networks for the incompressible Navier–Stokes equations,

X. Jin, S. Cai, H. Li, and G. E. Karniadakis, “NSFnets (Navier–Stokes flow nets): Physics-informed neural networks for the incompressible Navier–Stokes equations,” J. Comput. Phys. , vol. 426, May 2021, Art. no. 109951

2021
[29]

PIE- NeRF: Physics-based interactive elastodynamics with NeRF,

Y . Feng, Y . Shang, X. Li, T. Shao, C. Jiang, and Y . Yang, “PIE- NeRF: Physics-based interactive elastodynamics with NeRF,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2024, pp. 4450–4461

2024
[30]

Instant neural graphics primitives with a multiresolution hash encoding,

T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,” ACM Trans. Graph. , vol. 41, no. 4, pp. 1–15, Jul. 2022

2022
[31]

A deep energy method for finite deformation hyperelasticity,

V . M. Nguyen-Thanh, X. Zhuang, and T. Rabczuk, “A deep energy method for finite deformation hyperelasticity,” Eur. J. Mech. A Solids , vol. 80, Jul. 2020, Art. no. 103874

2020
[32]

The mixed deep energy method for resolv- ing concentration features in finite strain hyperelasticity,

J. N. Fuhg and N. Bouklas, “The mixed deep energy method for resolv- ing concentration features in finite strain hyperelasticity,” J. Comput. Phys., vol. 451, Jun. 2022, Art. no. 110839

2022
[33]

Physics-informed deep learning for computational elastodynamics without labeled data,

C. Rao, H. Sun, and Y . Liu, “Physics-informed deep learning for computational elastodynamics without labeled data,” J. Eng. Mech., vol. 147, no. 8, Aug. 2021, Art. no. 04021043

2021
[34]

Reduced order modeling for nonlinear structural analysis using Gaussian process regression,

M. Guo and J. S. Hesthaven, “Reduced order modeling for nonlinear structural analysis using Gaussian process regression,” Comput. Methods Appl. Mech. Eng. , vol. 341, pp. 807–826, Feb. 2018

2018
[35]

Ceccato, J

F. Ceccato, J. Fern, M. Martinelli, A. Rohe, and A. Yerro, Anura 3D MPM Software. Anura3D MPM Research Community, 2022

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Computational discovery of scientific knowledge,

S. D ˇzeroski, P. Langley, and L. Todorovski, “Computational discovery of scientific knowledge,” in Computational Discovery of Scientific Knowl- edge: Introduction, Techniques, and Applications in Environmental and Life Sciences, Springer, 2007, pp. 1–14

work page 2007

[2] [2]

Simulations of subatomic many-body physics on a quantum frequency processor,

H.-H. Lu et al. , “Simulations of subatomic many-body physics on a quantum frequency processor,” Phys. Rev. A , vol. 100, no. 1, Art. no. 012320, Jul. 2019

work page 2019

[3] [3]

Time evolution of galaxy scaling relations in cosmological simulations,

P. Taylor and C. Kobayashi, “Time evolution of galaxy scaling relations in cosmological simulations,” Mon. Not. R. Astron. Soc. , vol. 463, no. 3, pp. 2465–2479, Nov. 2016

work page 2016

[4] [6]

A neuro-symbolic approach to the logic of scientific discov- ery,

I. V otsis, “A neuro-symbolic approach to the logic of scientific discov- ery,” in Proc. Int. Conf. Model-Based Reasoning , 2023, pp. 306–330

work page 2023

[5] [7]

Quarteroni and A

A. Quarteroni and A. Valli, Numerical Approximation of Partial Differ- ential Equations. Berlin, Germany: Springer, 2008

work page 2008

[6] [8]

A finite mass based method for Vlasov– Poisson simulations,

D. J. Larson and C. V . Young, “A finite mass based method for Vlasov– Poisson simulations,” J. Comput. Phys. , vol. 284, pp. 171–185, Feb. 2015. 8 IEEE TRANSACTIONS ON ARTIFICIAL INTELLIGENCE, VOL. 00, NO. 0, MONTH 2020

work page 2015

[7] [9]

XSEDE: accelerating scientific discovery,

J. Towns et al. , “XSEDE: accelerating scientific discovery,” Comput. Sci. Eng., vol. 16, no. 5, pp. 62–74, Sep./Oct. 2014

work page 2014

[8] [10]

The opportunities, challenges, and risks of high performance computing in computational science and engineering,

D. E. Post, R. P. Kendall, and R. F. Lucas, “The opportunities, challenges, and risks of high performance computing in computational science and engineering,” Adv. Comput., vol. 66, pp. 239–301, 2006

work page 2006