acceptodds
Under review as a conference paper at ICLR 2027

Second-Order Neural Contact Potentials

Abstract

While conventional contact models are often defined over separate discrete contact cases, continuous surface-integral potentials provide a unified alternative. However, approximating these potentials with neural networks places new demands on smoothness and derivative quality. We introduce Second-Order Neural Contact Potentials (SNCP), a smooth neural representation for sampling-free contact handling in physics-based simulation. Building on Neural Collision Fields (NCF), which replace explicit contact sampling with a learned integral over pairs of mesh primitives, we address two key limitations that prevent their use in second-order solvers: nonsmooth derivatives and difficulties representing energies near zero. Our formulation replaces the Gaussian kernel of NCF with a -continuous Wendland polynomial and introduces a neural parameterization that enforces a smooth outer cutoff and supports consistent differentiation, yielding stable and well-behaved first- and second-order derivatives. The resulting contact potential generalizes to different meshes without retraining and supports both position-based dynamics and Newton-based finite-element methods. Across challenging simulations involving rigid bodies, volumetric deformable bodies, and thin shells, our method produces accurate and stable contact responses while enabling robust second-order optimization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
collision handling, neural fields, physical simulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JxebOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper extends Neural Collision Fields (NCF) to second-order optimization by introducing a twice-differentiable Wendland kernel and a factorized neural representation that preserves smoothness and compact support. The authors demonstrate their method on several standard contact simulation examples using both XPBD and Newton-based solvers.

Critical strengths
  • The paper addresses the limitations of NCF in second-order optimization.
  • The authors demonstrate the proposed approach on a variety of contact simulation scenarios.
Critical weaknesses
  1. Quadrature comparison. The comparison with numerical quadrature is strange. The authors use adaptive order-15 quadrature to generate their training data but compare their learned approach against uniform quadrature. Why? Uniform quadrature is generally inefficient for localized integrands, and high-order Gaussian or adaptive quadrature would be more appropriate baselines. In particular, the authors already have an adaptive quadrature implementation. Why not compare against it? Furthermore, Table 1 shows that computing the energy, gradient, and Hessian with SNCP is more than twice as slow as their uniform quadrature baseline.
  2. Accuracy of derivatives. The reported relative errors in the gradients (19%) and Hessians (29%) are substantial. How do these errors affect the convergence and accuracy of the simulations? In particular, does using the learned Hessian increase the number of Newton iterations compared with evaluating the derivatives of the original integral numerically? The paper does not sufficiently investigate the practical consequences of these approximation errors.
  3. CPU/GPU comparison. The adaptive quadrature timings are reported using a 32-thread CPU implementation, whereas the learned approach runs on the GPU. This makes it difficult to assess the actual computational advantage of the neural representation. Why not implement and benchmark adaptive quadrature on the GPU as well?
  4. Missing comparisons with existing methods. The paper does not provide direct runtime or convergence comparisons against established non-learning second-order contact methods, such as IPC. Since the main motivation is enabling second-order contact simulation, these comparisons are important to establish whether the proposed approach offers any practical advantage in terms of speed, convergence, accuracy, or robustness.
Questions for authors

Please address the issues raised in the weaknesses, particularly the choice of quadrature baseline, the effect of derivative approximation errors on Newton convergence, and the lack of comparisons against existing second-order contact methods.

2 comments

We thank reviewer for the detailed assessment. We agree that the original comparisons did not sufficiently establish the accuracy–cost trade-off or the practical consequences of derivative approximation. We have expanded the quadrature benchmarks, evaluated downstream Newton convergence, and added an IPC comparison.

  1. Quadrature baselines, Hessian cost, and GPU implementation. (regarding Weakness 1 and 3) The uniform barycentric rule was included for continuity with the NCF-style numerical baseline. We agree that this comparison alone was insufficient, particularly because adaptive quadrature is already used to generate our training labels and is better suited to localized integrands.
  • We have therefore added Uniform80 and a GPU implementation of the same adaptive order-15 quadrature algorithm used by the CPU implementation described in the paper, evaluated at tolerances 10−510^{-5} and 10−610^{-6}. | Method | Energy MRE (%) | Gradient MRE (%) | Hessian MRE (%) | E time (s) | E+G time (s) | E+G+H time (s) | |---|---:|---:|---:|---:|---:|---:| | SNCP | 16.781 | 18.670 | 29.485 | 0.253 | 0.523 | 10.570 | | Uniform40 | 59.367 | 52.112 | 86.688 | 0.602 | 0.640 | 4.703 | | Uniform80 | 23.818 | 18.975 | 25.114 | 10.773 | 11.549 | 82.034 | | GPU adaptive, 10−510^{-5} | 2.309 | 4.385 | 13.307 | 2.790 | 3.551 | 17.766 | | GPU adaptive, 10−610^{-6} | 1.32×10−131.32\times10^{-13} | 2.06×10−132.06\times10^{-13} | 1.54×10−111.54\times10^{-11} | 13.040 | 19.212 | 150.561 |

  • CPU versus GPU. At tolerance 10−610^{-6}, the 32-thread CPU implementation takes 38.14 s for energy and 58.40 s for energy plus gradient. The corresponding GPU times are 13.040 s and 19.212 s, giving speedups of approximately 2.92× and 3.04×, respectively. Thus, the GPU port provides roughly threefold acceleration in these tests. Efficient parallelization remains challenging for this adaptive algorithm because refinement introduces nonuniform computational work.

  • Accuracy versus evaluation cost. SNCP’s E+G+H evaluation is approximately 2.25× slower than Uniform40, although its errors are substantially lower. The expanded benchmarks clarify how this trade-off changes with quadrature accuracy:

    • Uniform80 achieves gradient error close to SNCP’s and lower Hessian error, although its energy error remains higher. Its E+G+H evaluation costs approximately 7.76× as much as SNCP’s.

    • GPU adaptive quadrature at 10−510^{-5} achieves lower errors on all three metrics, at approximately 1.68× SNCP’s E+G+H cost.

    • GPU adaptive quadrature at 10−610^{-6} closely matches the numerical reference used for evaluation, with the near-zero MREs reported above, at approximately 14.24× SNCP’s E+G+H cost.

  • These results assess the computational benefit using a GPU adaptive baseline and make the accuracy–cost trade-off explicit. We will revise the efficiency discussion to distinguish energy, gradient, and full Hessian evaluation and qualify the computational benefit by the accuracy required.

  1. Effect of approximation errors on Newton convergence and simulation behavior. (regarding weakness 2)
  • To examine the practical consequences of the approximation, we replaced the learned contact potential with numerical integration in the same Newton runtime. We also evaluated a Gauss–Newton control that preserves SNCP’s energy and gradient while changing the contact-curvature approximation. It uses κ2E∇E∇E⊤\frac{\kappa}{2E}\nabla E\nabla E^\top.
Mean Newton updates per stepStackMass ratioCloth
Numerical integration/Newton1.0755.2403.575
SNCP/full Hessian1.0509.6104.503
SNCP/Gauss–Newton1.055112.346*9.355
* The starred mean covers 81 recorded frames, including the frame that reaches the iteration cap of 1024; frames 82–200 are unobserved.
  • Effect on Newton convergence. SNCP requires approximately 83% more Newton updates on Mass ratio and 26% more on Cloth than numerical integration, while the Stack counts are close. Thus, the approximation has a measurable downstream cost. This comparison reflects the combined effect of approximating the energy and its derivatives.

  • Role of full curvature. The separate Gauss–Newton control shows that the full Hessian is useful when optimizing the learned potential. Full curvature reduces the Cloth iteration count from 9.355 to 4.503 updates per step. Gauss–Newton completes Stack and Cloth, but completes only 80 of 200 Mass ratio advances before reaching the iteration cap at frame 81. SNCP with the full Hessian completes the entire Mass ratio horizon. These results support the value of full curvature for convergence in the tested scenes, alongside its evaluation cost reported above.

  • Simulation behavior. Numerical integration and SNCP with the full Hessian both complete all steps under the native convergence criterion. We observe visually similar contact behavior using the same contact stiffness parameter, without retuning, and will add visual comparisons to the revised paper. This is qualitative evidence of similar behavior in these scenes; it does not establish a quantitative bound on trajectory or contact-force error.

  1. Comparison with IPC. (regarding Weakness 4)
  • We agree that IPC is an important practical reference and have added a comparison between SNCP and IPC contact within the same libuipc framework: | Scene | SNCP time (ms/step) | IPC time (ms/step) | SNCP Newton updates/step | IPC Newton updates/step | |---|---:|---:|---:|---:| | Stack | 35.44 | 9.55 | 1.050 | 1.035 | | Mass ratio | 445.41 | 88.85 | 9.610 | 10.665 | | Cloth | 173.62 | 46.40 | 4.503 | 3.485 |

  • Although both methods requires neerly identical Newton updates. IPC is 3.71–5.01× faster across these scenes. This is a strong baseline: IPC has benefited from several years of algorithmic development since its introduction in 2020, and libuipc incorporates state-of-the-art GPU IPC techniques. We acknowledge the current runtime gap and will report it explicitly.

  • An important motivation for SNCP is its potential to extend learned contact evaluation to more complex geometric representations, including higher-order surfaces. Direct contact handling on curved geometry remains challenging; for example, High-Order IPC [Ferguson et al., 2023] employs linear collision proxies to avoid costly direct curved-surface collision detection. By learning contact integrals and their derivatives, SNCP offers a promising alternative for handling richer geometry without relying on explicit contact sampling. While higher-order geometry is beyond the scope of this work, we will highlight this important motivation and future direction in the revision.

Reviewer_j2kMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

SNCP extends Neural Collision Fields to Newton-based contact simulation. It replaces the Gaussian contact kernel with a compactly supported Wendland kernel, factors the predicted energy into a smooth geometric envelope and a positive learned coefficient, and trains a SIREN with value and first-derivative supervision. The complete expression is differentiated to obtain world-space gradients and Hessians. I think the main contribution is the smooth support construction and its use in second-order solvers.

Critical strengths
  1. The idea of separating a smooth geometric cutoff from a positive learned coefficient is well motivated because it will give us a support guarantee and allow using the potential for second-order optimization.

  2. The Newton comparisons show a robustness improvement over the NCF adaptation.

  3. The same trained model is used across different geometries, deformable bodies, and thin shells without retraining. The demonstrations with many simultaneous contacts and self-contact are useful too.

Critical weaknesses
  1. The unfactored Sobolev model in Table 2 already has similar pointwise prediction errors to the full method. How does it compare in Newton convergence and runtime, particularly when contacts enter or leave the active region? I think this would clarify the benefit of the geometric factor.

  2. The comparison with numerical integration uses only one quadrature resolution, with substantially different gradient and Hessian errors. A comparison at similar accuracy would help assess the computational benefit of learning the integral, particularly since the full Hessian evaluation is slower at the resolution.

  3. No comparisons on contact behaviour. How much does the learned approximation change the contact stiffness?

  4. It is unclear whether the coordinate-gradient metrics in Table 2 evaluate the log coefficient or the reconstructed energy. Eq. (57) subtracts the geometric factor’s contribution from the training target, so a negative cosine for the coefficient does not necessarily imply an incorrect force direction, right?

I would say there are some issues in the writing too, the main text should explain how qBq_B is constructed, rather than only stating its properties. I think Eqs. 21, 22 should appear in Section 4.2.

Questions for authors
  1. I was trying to understand how the method is implemented. Eq. (50) defines CS=2αE^ABC_S=\sqrt{2\alpha\widehat E_{AB}}, with α=1/κ\alpha=1/\kappa. This would mean that compliant energy CS2/(2α)=E^ABC_S^2/(2\alpha)=\widehat E_{AB}, whereas the Newton formulation uses a stiffness-scaled energy. Substituting Eq. (50) into Eq. (51) also cancels α\alpha from the position correction, making the stated contact update independent of κ\kappa?

  2. Eqs. (49) and (51) omit the accumulated-multiplier term −(α/h2)λ-(\alpha/h^2)\lambda from the numerator. Isn’t that term required after the first iteration when compliance is nonzero?

  3. Since (g_B) can be positive when the exact energy is zero, the log target in Eq. (56) is undefined for some configurations. How are these handled?

  4. How much of the Newton improvement requires the full neural Hessian? Maybe you can do a Gauss–Newton comparison using the same energy and gradient

  5. What convergence criteria are used for each method in comaprison?

2 comments

We thank reviewer for the detailed questions. They identify several places where the manuscript should more clearly distinguish approximation accuracy, the role of the geometric factor, and downstream solver behavior. We address the experimental concerns and implementation questions below.

  1. Factorization versus the unfactored Sobolev model. (regarding Weakness 1)
  • We agree that similar pointwise prediction errors do not establish similar downstream convergence or runtime. We compared the unfactored Sobolev (US) and factored Sobolev (FS) models using three paired training seeds (42, 43, and 44), with identical scene and solver settings:
SceneFS completed trajectoriesUS completed trajectories
Spike/wedge3/30/3
Wedge in crack3/30/3
Internal edges3/32/3
  • FS completes all nine tested trajectories, whereas US completes two. For the two internal-edge trajectories completed by both models, we additionally compare total solve time and Newton updates:
SeedFS solve time (s)US solve time (s)FS Newton updatesUS Newton updates
43179.04368.59210514
44177.89435.71212634
  • On these matched completed runs, FS is 2.06–2.45× faster and requires 59.1–66.6% fewer Newton updates than US. Thus, the geometric factor improves trajectory completion across the tested scenes and reduces solver work on the internal-edge runs that both models complete, despite their similar pointwise prediction errors.

  • Sobolev supervision constrains energy and first derivatives at sampled configurations, but does not itself impose finite support or directly constrain curvature throughout a simulated trajectory. The geometric factor supplies a prescribed cutoff at the envelope boundary. The unfactored decoder does not impose this cutoff and can retain residual forces and curvature beyond it. This structural distinction is relevant when contacts enter or leave the active region, although the trajectory-level measurements do not isolate the behavior at individual activation or deactivation events.

  1. Quadrature at comparable accuracy. (regarding Weakness 2) We agree that a single quadrature resolution is insufficient to assess the computational benefit. We have added Uniform80 and GPU adaptive quadrature with tolerance 10−510^{-5}: | Method | Energy MRE (%) | Gradient MRE (%) | Hessian MRE (%) | E time (s) | E+G time (s) | E+G+H time (s) | |---|---:|---:|---:|---:|---:|---:| | SNCP | 16.781 | 18.670 | 29.485 | 0.253 | 0.523 | 10.570 | | Uniform40 | 59.367 | 52.112 | 86.688 | 0.602 | 0.640 | 4.703 | | Uniform80 | 23.818 | 18.975 | 25.114 | 10.773 | 11.549 | 82.034 | | GPU adaptive, 10−510^{-5} | 2.309 | 4.385 | 13.307 | 2.790 | 3.551 | 17.766 | | GPU adaptive, 10−610^{-6} | 1.32×10−131.32\times10^{-13} | 2.06×10−132.06\times10^{-13} | 1.54×10−111.54\times10^{-11} | 13.040 | 19.212 | 150.561 |
  • Uniform80 provides a comparison at similar gradient accuracy: its gradient MRE is 18.975%, compared with 18.670% for SNCP. It also has lower Hessian error, although its energy error is higher. Its E+G+H evaluation costs approximately 7.76× as much as SNCP’s. These results clarify the accuracy–cost trade-off.
  1. Contact behavior and the role of the full Hessian. (regarding Weakness 3 and Question 4).
  • We compare numerical integration/Newton, SNCP with the full contact Hessian, and SNCP with Gauss–Newton contact curvature in the same Newton runtime. The Gauss–Newton control preserves SNCP’s energy and gradient while changing the contact-curvature approximation. It uses κ2E∇E∇E⊤\frac{\kappa}{2E}\nabla E\nabla E^\top. | Mean Newton updates per step | Stack | Mass ratio | Cloth | |---|---:|---:|---:| | Numerical integration/Newton | 1.075 | 5.240 | 3.575 | | SNCP/full Hessian | 1.050 | 9.610 | 4.503 | | SNCP/Gauss–Newton | 1.055 | 112.346* | 9.355 | * The starred mean covers 81 recorded frames, including the frame that reaches the iteration cap of 1024; frames 82–200 are unobserved.

  • Contact behavior. Numerical integration and SNCP with the full Hessian both complete all steps under the native convergence criterion. We observe visually similar contact behavior using the same stiffness parameter. We will add visual comparisons to the revised paper. These observations provide qualitative evidence in the tested scenes; visual agreement at the same parameter value does not, by itself, quantify the change in effective contact stiffness.

  • Role of the full Hessian. The separate Gauss–Newton control shows that full curvature improves convergence for the learned potential: it reduces the Cloth iteration count and completes the Mass ratio horizon, whereas Gauss–Newton completes only 80 of 200 Mass ratio advances. This supports the practical value of full curvature in these tests, while the additional Hessian evaluation cost remains as reported above.

  1. What the coordinate-gradient metrics measure. (regarding Weakness 4)
  • Your interpretation is correct: a negative cosine similarity for the coefficient gradient alone would not necessarily imply an incorrect force direction.
  • The coordinate-gradient metrics in Table 2 compare derivatives of the reconstructed normalized energy, including the geometric-envelope contribution, against numerical energy-derivative labels. The subtraction in Eq. (57) is used to construct the coefficient’s training target. During evaluation, we reconstruct the energy gradient, including the envelope contribution, before computing the errors and cosine similarities. We will make this distinction explicit in the manuscript.
  1. XPBD equations. (regarding Question 1 and 2)
  • Stiffness scaling. Thank you for identifying the inconsistency in Eq. (50). The factor α\alpha under the square root is erroneous. The corrected expression is
\qquad \alpha=\frac{1}{\kappa}.$$ With this definition, the position update retains its dependence on $\kappa$ through the compliance. - **Accumulated multiplier.** We also missed the accumulated-multiplier $-(\alpha/h^2)\lambda$ term in Eq. (49) and (51). We will correct these in the paper. 6. Zero-energy targets. (regarding **Question 3**) - The unshifted log target shown in the manuscript is incomplete. The implemented target on $g_B>0$ is $z^*=\log(E_{num}/g_B^4 + \epsilon_c)$, with $\epsilon_c=10^{-12}$, and inactive rows use the finite target $\log(\epsilon_c)$. The derivative target uses $E_{num}+\epsilon_c g_B^4$ in the denominator. Thus zero labels do not require evaluating $\log(0)$. 7. Convergence criteria. (regarding **Question 5**) - The relative Newton tolerance is $10^{-3}$, measured using the maximum absolute component of the proposed vertex displacement relative to the first Newton direction in the same timestep. Thus, the relative convergence criterion is $||\Delta x_{k}||_{\infty}/||\Delta x_{0}||_{\infty}\le10^{-3}$: the proposed positional update has decreased to at most 0.1% of its initial magnitude. - We use a CG relative error tolerance of $10^{-3}$ for Mass ratio and Cloth scene and $10^{-10}$ for Stack scene to control the accuracy of the linear solve within each Newton iteration. - Position-based runs use fixed iteration budgets and do not have extra convergence criteria. We will state these settings explicitly in the revised manuscript. 8. Construction of the support coordinate. - We also agree that the construction of $q_B$ belongs in the main method. We will move the barycentric barrier problem and the definition $q_B=||A \alpha^* − B \beta^*||^2−6 \kappa_B$ beside the definition of $g_B$.
Reviewer_MKWHOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes second-order neural contact potentials (SNCP) that aims t omake neual collision fields (NCFs) suitable for Newton solvers. It does that by approximating a finite-range surface contact energy with a neural coefficient multiplied by a prescribed geometric factor. Comparisons focus on CFS and demonsrate improved convergence.

Critical strengths
  • The motivation and the approach seem reasonable
  • The paper does a good job explaining the high- and mid-level details. I am far from an expert and do not understand the details, but can understand the coarse ideas well
  • The qualitative results are good, and the convergence analysis is good. The method leads to better numerical behaviour.
Critical weaknesses
  • No comparisons to IPC. Comparing only to NCF seems not very reasonable. IPC should be a cleare reference
  • The method seems to be slower than fixed quadrature with full Hessian evaluation. It has better fidelity so there is an inherent trade-off
Questions for authors

See weaknesses

1 comment

We thank reviewer for highlighting the missing IPC comparison and the cost of full Hessian evaluation. We agree that both are needed to establish the practical scope of the method.

  1. Comparison with IPC. (regarding Weakness 1)
  • We agree that IPC is an important practical reference and have added a comparison between SNCP and IPC contact within the same libuipc framework: | Scene | SNCP time (ms/step) | IPC time (ms/step) | SNCP Newton updates/step | IPC Newton updates/step | |---|---:|---:|---:|---:| | Stack | 35.44 | 9.55 | 1.050 | 1.035 | | Mass ratio | 445.41 | 88.85 | 9.610 | 10.665 | | Cloth | 173.62 | 46.40 | 4.503 | 3.485 |
  • Although both methods requires neerly identical Newton updates. IPC is 3.71–5.01× faster across these scenes. This is a strong baseline: IPC has benefited from several years of algorithmic development since its introduction in 2020, and libuipc incorporates state-of-the-art GPU IPC techniques. We acknowledge the current runtime gap and will report it explicitly. An important motivation for SNCP is its potential to extend learned contact evaluation to more complex geometric representations, including higher-order surfaces. Direct contact handling on curved geometry remains challenging; for example, High-Order IPC [Ferguson et al., 2023] employs linear collision proxies to avoid costly direct curved-surface collision detection. By learning contact integrals and their derivatives, SNCP offers a promising alternative for handling richer geometry without relying on explicit contact sampling. While higher-order geometry is beyond the scope of this work, we will highlight this important motivation and future direction in the revision.
  1. Full Hessian cost and the accuracy–cost trade-off. (regarding Weakness 2)
  • When evaluating energy, gradient, and Hessian together, SNCP is approximately 2.25× slower than Uniform40 (10.570 s versus 4.703 s), while achieving substantially lower errors. The original comparison thWerefore demonstrates an accuracy–cost trade-off. To characterize this trade-off more fully, we have added Uniform80 and GPU adaptive quadrature with tolerance 10−510^{-5}: | Method | Energy MRE (%) | Gradient MRE (%) | Hessian MRE (%) | E time (s) | E+G time (s) | E+G+H time (s) | |---|---:|---:|---:|---:|---:|---:| | SNCP | 16.781 | 18.670 | 29.485 | 0.253 | 0.523 | 10.570 | | Uniform40 | 59.367 | 52.112 | 86.688 | 0.602 | 0.640 | 4.703 | | Uniform80 | 23.818 | 18.975 | 25.114 | 10.773 | 11.549 | 82.034 | | GPU adaptive, 10−510^{-5} | 2.309 | 4.385 | 13.307 | 2.790 | 3.551 | 17.766 | | GPU adaptive, 10−610^{-6} | 1.32×10−131.32\times10^{-13} | 2.06×10−132.06\times10^{-13} | 1.54×10−111.54\times10^{-11} | 13.040 | 19.212 | 150.561 |
  • Uniform80 achieves gradient error close to SNCP’s and lower Hessian error, although its energy error remains higher. Its E+G+H evaluation is approximately 7.76× more expensive than SNCP’s. GPU adaptive quadrature achieves lower errors on all three metrics, at approximately 1.68× SNCP’s E+G+H cost.
  • These results clarify how the trade-off changes with quadrature resolution and adaptivity. SNCP has the lowest E and E+G evaluation times among the tested methods, while full Hessian evaluation remains a substantial cost and is slower than Uniform40. We will revise the efficiency discussion to distinguish these evaluation settings and qualify the performance claims by approximation accuracy.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.