Inverse Design of Metainterfaces towards Static Friction Control: Beyond the Hertzian Limit
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "Inverse Design of Metainterfaces towards Static Friction Control".
Mira: Here's the gist of the paper. Programming the static friction of mechanical interfaces is critical for soft robotics, haptics, and precision gripping By utilizing general axisymmetric asperities, we unlock nonlinear macroscopic responses unattainable by standard Hertzian contacts To solve the inverse problem, we embed a fully differentiable contact mechanics engine within a neural network and a quadratic optimizer The resulting optimal unit cells are validated against high-fidelity Boundary Element Method (BEM) simulations This framework bridges data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering functional tribological surfaces.
Kai: Theoretical Foundation of Contact Mechanics The paper establishes the mathematical relationships governing contact between axisymmetric asperities, showing that the contact area scales as A ∼ F 2γ+1.
Mira: The macroscopic contact stiffness is dictated by the derivative of the force with respect to the area, dF/dA = E∗κ(γ)cγ / (πγ+1)2.
Kai: This relationship shows that for a single asperity, the physical limits of the A − F curve are given by A ∼ F for the conical case and A ∼ const for the flat punch.
Paper summary: Kai: What they claim is that by using general axisymmetric asperities, they unlock nonlinear macroscopic responses that you just don't get from standard Hertzian contacts. Mira So, instead of just using simple spheres or cones, they’re looking at a whole range of asperity shapes to engineer these new effects. Kai To solve the inverse problem—figuring out the shape parameters—they embed a fully differentiable contact mechanics engine right inside a neural network and use a quadratic optimizer. Lev That sounds ambitious. For real hardware, training something that handles these complex physics gradients requires some serious setup because you're dealing with non-trivial error landscapes on the learning side, and you need to make sure the backpropagation actually lands where it should >
Mira: What’s interesting is they are not relying on purely data-driven surrogates which just fit a curve. Instead, they use a hybrid Scientific Machine Learning framework that embeds analytical governing equations directly into the learning loop. Kai So, the network learns by trying to minimize the error between what you want and what the physics predicts based on those equations >
Lev: Embedding those analytical equations is smart because it means you can backpropagate physical gradients directly, which helps steer the optimization toward physically valid solutions rather than just mathematically plausible ones that might be physically impossible to build >
Kai: They train this using a composite loss function that balances the shape and intensity of the alpha P star curve, which they define as Ltotal equals Lphysics plus lambda e Lparam >
Mira: And to make sure the optimizer stays in a valid area, they use a supervised curriculum learning approach where that anchor weight lambda e is linearly annealed to zero over early epochs >
Lev: And they also increase the steepness parameter kappa in the softplus contact regularizer way up to one zero five because they want to recover those strict mathematical boundaries of standard contact mechanics >
Kai: They validate all this against full-field Boundary Element Method simulations, and their analytical surrogate predictions track the BEM evaluations perfectly across the entire loading domain with a maximum relative error of less than three percent >
Conclusion: Kai: So, looking at "Inverse Design of Metainterfaces towards Static Friction Control: Beyond the Hertzian Limit," they essentially show how to use this hybrid SciML framework to design surfaces with controllable static friction based on programmable contact behaviors. Mira They move away from the limitations of classical models by using these engineered axisymmetric asperities to get those nonlinear responses they couldn't achieve before >
Kai: The authors are Jacopo Bilotto, Arnav Singhal, Joaquin Garcia-Suarez, Gaëtan Cortes, Lucas Fourel, and Jean-François Molinari from EPFL and Brown University > Lev For me the implication is that this could lead to designing interfaces where you can precisely control how hard something sticks or slips when it’s moving under load >
Mira: It shows that shape variability of the asperities, represented by gamma, is really the key factor here for engineering exotic contact mechanics, and it outperforms spherical baselines even on targets that are outside their normal testing range >
Kai: The paper points toward future work where they want to relax those assumptions about purely elastic non-interacting contacts by integrating localized plasticity or multi-asperity interaction fields > Lev That would be a big step because real hardware isn't just about simple elasticity; it’s about how materials deform locally under stress, and that's where the next level of complexity is needed >
Mira: Right. They also suggest this framework can evolve from just static shape optimization to inverse design for dynamic frictional properties by training the network to capture those more complex nonlinear responses >
Jacopo Bilotto, Arnav Singhal, Joaquin Garcia-Suarez, Gaétan Cortes, Lucas Fourel, Jean-François Molinari
Institute of Civil Engineering, Institute of Materials, Ecole Polytechnique FƓre de Lausanne · Brown University
cond-mat.soft, cs.LG
Submitted: 2026-05-10
Updated: 2026-10-29
Comments: 21 pages, 16 figures
Journal ref: Tribol Lett 74, 128 (2026)
DOI: 10.1007/s11249-026-02210-4
Code: https://github.com/JBil8/frictional_metasurfaces_inverse_design
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: Programming the static friction of mechanical interfaces is critical for soft robotics, haptics, and precision gripping By utilizing general axisymmetric asperities…
Key concepts
- Axisymmetric Asperities
- These are idealized, circular bumps on a surface used to model contact between two curved objects. The paper uses these simple shapes to understand how the total contact area and stiffness change based on the shape of the bumps, which is crucial for controlling friction.
- Hybrid SciML Framework
- This approach combines scientific machine learning with physics-based models. It uses a neural network to predict surface parameters, but it is constrained by analytical equations that describe real contact mechanics. This ensures the AI designs surfaces that are physically possible.
- Inverse Design
- Instead of designing a shape and then calculating its friction (forward problem), this method starts with a target friction curve and works backward to find the exact surface geometry ($\gamma$ and $h$) required to achieve it. This is done by optimizing parameters within a neural network.
- Boundary Element Method (BEM)
- This is a high-fidelity simulation technique used to test the designs generated by the AI. BEM accurately solves complex contact problems across an entire domain, allowing researchers to verify that the simplified neural network predictions match real-world physical behavior with very high accuracy.
Terminology
Summary
Programming the static friction of mechanical interfaces is critical for soft robotics, haptics, and precision gripping By utilizing general axisymmetric asperities, we unlock nonlinear macroscopic responses unattainable by standard Hertzian contacts To solve the inverse problem, we embed a fully differentiable contact mechanics engine within a neural network and a quadratic optimizer The resulting optimal unit cells are validated against high-fidelity Boundary Element Method (BEM) simulations This framework bridges data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering functional tribological surfaces. p. 0 By utilizing general axisymmetric asperities, we unlock nonlinear macroscopic responses unattainable by
standard Hertzian contacts. To solve the inverse problem, we embed a fully differentiable contact
mechanics engine within a neural network and a quadratic optimizer.Where the paper says it
Theoretical Foundation of Contact Mechanics
The paper establishes the mathematical relationships governing contact between axisymmetric asperities, showing that the contact area scales as A ∼ F 2γ+1. The macroscopic contact stiffness is dictated by the derivative of the force with respect to the area, dF/dA = E∗κ(γ)cγ / (πγ+1)2. p. 3 (4)
The macroscopic contact stiffness, directly related to the coefficient of static friction, is dictated by
the derivative of the force with respect to the area. Differentiating F(A) yields:
dF
dA = E
∗κ(γ)cγ
1
π
γ+1
2
A
γ−1
2 . From Equation 2 the contact area A = πa2
is given by
A = π
(γ + 1)F
2γκ(γ)cE∗
2
γ+1
=⇒ A ∼ F
2
γ+1 . (3)
Therefore, for a single asperity, the physical limits of the A − F curve are given by A ∼ F for the
conical case and A ∼ const for the flat punch, see Figure 1b. 4e,f for a depiction. Leveraging the analytical solutions from Sneddon31
,
the total contact load Ftot and total real contact area Atot are given by:
4Where the paper says it
Where the paper says it
Where the paper says it
Hybrid Neural Architecture for Inverse Design
The inverse problem is addressed using a hybrid Scientific Machine Learning (SciML) framework that embeds analytical governing equations directly into the learning loop. The architecture consists of a fully connected Multi-Layer Perceptron (MLP) which takes the target discretized and normalized physical curves and max values as input to predict the parameter vectors γˆ and hˆ. p. 5 We propose the hybrid SciML architecture shown in Fig. The model consists of two coupled
blocks:
• A fully connected Multi-Layer Perceptron (MLP) which takes the target discretized and normalized physical curves (αˆ(Pˆ), and Sˆ = dα/d ˆ Pˆ) and the max values P
∗
max, αmax as input. The Multi-Layer Perceptron (MLP) predicts the parameter vectors γˆ
and hˆ. These predicted parameters are then passed to a differentiable physics layer implementing the
analytical forward model from Eqs. These predicted parameters are then passed to a differentiable physics layer implementing the
analytical forward model from Eqs. 6 and 7, yielding reconstructed physical curves yˆ = F(γˆ, hˆ). heights directly, the network predicts dimensionless inter-asperity gaps. The first gap is forced to zero,
and the absolute depth offsets are recovered via a cumulative sum:
hi =
X
i
j=1
gj where g1 = 0, gj>1 = ˆy
gap
j
·
2∆max
N − 1
(8)
This architectural choice mathematically guarantees that the asperities are sorted by engagement
depth, dramatically reducing the optimization search space.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Training Strategy and Regularization
The training process employs a composite loss function that balances shape and intensity of the α(P∗) curve, defined as Ltotal = Lphysics + λ(e)Lparam. To steer the optimizer into a valid basin of attraction, a supervised curriculum learning approach is used where the anchor weight λ(e) is linearly annealed to zero over early epochs. p. 5 This weight (λ) is linearly brought to zero over the first few epochs of training. This
mechanism safely steers the optimizer into a valid physical basin of attraction, after which the constraint is fully relaxed, allowing the final topography to be driven purely by the macroscopic physics
loss. Simultaneously, we employ a “contact regularization” method1 within the differentiable physics layer:
the steepness of the softplus contact regularizer (κ) is initialized at a low value (κ = 103
), effectively
blurring the contact boundaries and smoothing the non-convex physical loss landscape. As training progresses, κ is exponentially annealed to 105
, recovering the strict discontinuities characteristic
of exact multi-asperity contact mechanics, see App.Where the paper says it
Where the paper says it
Validation and Benchmarking
The framework is rigorously validated against full-field Boundary Element Method (BEM) simulations, where the analytical surrogate predictions track the BEM evaluations perfectly across the entire loading domain with a maximum relative error of less than 3%. The model successfully reproduces complex, nonlinear contact area (and by extension friction) as a function of load trajectories (A(F)). p. 1 Specifically, we deploy a
neural surrogate to provide zero-shot topological initializations, which are subsequently refined using
a multi-stage optimization scheme. This architecture allows us to engineer surfaces that reproduce
complex, nonlinear contact area (and by extension friction) as a function of load trajectories (A(F)). As the dimensionality of the design space increases, the required number of
random starts is expected to grow exponentially. The neural surrogate circumvents this entirely by
mapping the target physics directly into the correct global basin of attraction, requiring only a single
8 We apply this solver across two
distinct design spaces: a generalized formulation that optimizes both asperity heights and shape exponents, and a strictly Hertzian baseline where the shapes are locked to perfect spheres (γi ≡ 2) and
only the height offsets are allowed to vary. To rigorously test the limits of these approaches, we evaluate them against three out-of-distribution
macroscopic targets, as illustrated in Fig.Where the paper says it
Where the paper says it
Where the paper says it
Conclusions and Future Directions
The work introduces a hybrid SciML framework that breaks the classical Hertzian barrier, enabling the inverse design of metainterfaces with programmable static friction. Future work should relax purely elastic, non-interacting assumptions by integrating localized plasticity and multi-asperity interaction fields. p. 4 While exact for sparse distributions, interaction effects become non-negligible at high contact densities 35. However, Boussinesq 9
showed that the elastic deformation field decays as 1/r, so that shielding effects are minimized when
the mean spacing between asperities is sufficiently large compared to the contact radius. 4 demonstrated experimentally that the arrangement of discrete hemispherical Hertzian
asperities can give rise to new frictional responses, not found in natural surfaces. However, the inverse
design of such interfaces is ill-posed: finding the exact topography that yields a target friction curve
is mathematically non-trivial and computationally challenging for brute-force optimization.Where the paper says it
Where the paper says it
Data Availability
The code used to generate the dataset and perform this study is available on Github The project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 945363. J.G.S. and G.C. gratefully acknowledge financial support from the Swiss National Science Foundation, via Ambizione Grant PZ00P2 216341 “Data-Driven Computational Friction”. p. 0 This framework bridges
data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering
functional tribological surfaces. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems.Where the paper says it
References
[1] Allgower, E.L., Georg, K., 1990. Introduction, in: Allgower, E. p. 0 This framework bridges
data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering
functional tribological surfaces. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems. This framework bridges
data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering
functional tribological surfaces. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems.Where the paper says it
Where the paper says it
[2] Archard, J.F., Allibone, T.E., 1997. Elastic deformation and the laws of friction. p. 4 While exact for sparse distributions, interaction effects become non-negligible at high contact densities 35. However, Boussinesq 9
showed that the elastic deformation field decays as 1/r, so that shielding effects are minimized when
the mean spacing between asperities is sufficiently large compared to the contact radius.Where the paper says it
[3] de Avila Belbute-Peres, F., Smith, K., Allen, K., Tenenbaum, J., Kolter, J.Z., 2018. End-to-End Differentiable Physics for Learning and Control, in: Advances in Neural Information Processing Systems, Curran Associates, Inc
[4] Aymard, A., Delplanque, E., Dalmas, D., Scheibert, J., 2024. Designing metainterfaces with specified friction laws. p. 3 2 The Multi-Asperity Differentiable Forward Model
To design complex metainterfaces, we construct a unit cell composed of N = 9 independent axisymmetric asperities distributed on a flat base. While this formulation scales to any arbitrary N, a 9-
asperity basis provides sufficient topological degrees of freedom to approximate nonlinear macroscopic
laws while maintaining good interpretability.Where the paper says it
[5] Bengio, Y., Louradour, J., Collobert, R., Weston, J., 2009. Curriculum learning, in: Proceedings of the 26th Annual International Conference on Machine Learning, Association for Computing Machinery, New York, NY
[6] Bessa, M.A., Bostanabad, R., Liu, Z., Hu, A., Apley, D.W., Brinson, C., Chen, W., Liu, W.K., 2017. A framework for data-driven analysis of materials under uncertainty: Countering the curse of dimensionality. p. 0 The predicted designs are strictly validated against high-fidelity Boundary Element Method (BEM) simulations. This framework bridges
data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering
functional tribological surfaces.Where the paper says it
[7] Bilotto, J., Kolinski, J.M., Lecampion, B., Molinari, J.F., Subhash, G., Garcia-Suarez, J. p. 0 Inverse Design of Metainterfaces for Static Friction
Control: Beyond the Hertzian Limit
Jacopo Bilotto1,*, Arnav Singhal1,2, Joaquin Garcia-Suarez1
, Gaëtan Cortes1
, Lucas Fourel1
,
and Jean-François Molinari1
1
Institute of Civil Engineering, Institute of Materials, École Polytechnique Fédérale de Lausanne (EPFL), CH
1015 Lausanne, Switzerland
2Brown University, 02912 Providence, RI, USA
*Corresponding author: jacopo. bilotto@epfl. bilotto@epfl. ch
May 20, 2026
Abstract
Programming the static friction of mechanical interfaces is critical for soft robotics, haptics,
and precision gripping.Where the paper says it
Where the paper says it
[8] Bonari, J., Paggi, M., Dini, D., 2022. A new finite element paradigm to solve contact problems with roughness. International Journal of Solids and Structures 253, 111643
[9] Boussinesq, J., 1885. Application des potentiels à l’étude de l’équilibre et du mouvement des solides élastiques. Gauthier-Villars
[10] Bowden, F.P., Tabor, D. p. 0 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems. According to the
classical theory of Bowden and Tabor 10, the static friction force is directly proportional to the real area
of contact, if adhesion effects are negligible. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems. According to the
classical theory of Bowden and Tabor 10, the static friction force is directly proportional to the real area
of contact, if adhesion effects are negligible. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems. According to the
classical theory of Bowden and Tabor 10, the static friction force is directly proportional to the real area
of contact, if adhesion effects are negligible. Inverse Design of Metainterfaces for Static Friction
Control: Beyond the Hertzian Limit
Jacopo Bilotto1,*, Arnav Singhal1,2, Joaquin Garcia-Suarez1
, Gaëtan Cortes1
, Lucas Fourel1
,
and Jean-François Molinari1
1
Institute of Civil Engineering, Institute of Materials, École Polytechnique Fédérale de Lausanne (EPFL), CH
1015 Lausanne, Switzerland
2Brown University, 02912 Providence, RI, USA
*Corresponding author: jacopo. bilotto@epfl.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[11] Ciavarella, M., Joe, J., Papangelo, A., Barber, J.R., 2019. The role of adhesion in contact mechanics. p. 0 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems. According to the
classical theory of Bowden and Tabor 10, the static friction force is directly proportional to the real area
of contact, if adhesion effects are negligible.Where the paper says it
[12] Colburn, S., Majumdar, A., 2021. Inverse design and flexible parameterization of meta-optics using algorithmic differentiation. p. 1 Unlike purely data-driven surrogates, differentiable physics approaches embed analytical governing
equations directly into the learning loop, allowing for the backpropagation of physical gradients 3
. This paradigm has been successfully applied to the inverse design of nanophotonic structures 20, topology optimization in optics 12, spinodal metamaterials 19, and general mechanical structures 34.Where the paper says it
[13] Djellouli, A., Albertini, G., Wilt, J., Tournat, V., Weitz, D., Rubinstein, S., Bertoldi, K., 2026. Squeaking at soft–rigid frictional interfaces. p. 0 As demonstrated by the seminal works of Archard2 and Greenwood &
Williamson17, the contact area of a random rough surface scales linearly with normal load, regardless
of the specific asperity shape: the well-known “roughness paradox”. This linearity creates a theoretical
barrier: it restricts standard interfaces to a single, Amontons-Coulomb frictional behavior, effectively
locking the coefficient of friction to a constant value.Where the paper says it
[14] Frérot, L., Anciaux, G., Rey, V., Pham-Ba, S., Molinari, J.F., 2020. Tamaas: a library for elastic-plastic contact of periodic rough surfaces. p. 0 ch
May 20, 2026
Abstract
Programming the static friction of mechanical interfaces is critical for soft robotics, haptics,
and precision gripping. Static friction is governed by the real contact area, and standard rough
surfaces exhibit a linear area-load scaling inherent to classical Archard and Greenwood-Williamson
models, severely restricting their functional range.Where the paper says it
[15] Föppl, L., 1941. Elastische Beanspruchung des Erdbodens unter Fundamenten. Forschung auf dem Gebiet des Ingenieurwesens A 12, 31–39
[16] Garcia-Suarez, J., 2026. A Matter of Shape: Contact Area Optimization in Soft Lubricated Impact. p. 0 ch
May 20, 2026
Abstract
Programming the static friction of mechanical interfaces is critical for soft robotics, haptics,
and precision gripping. Static friction is governed by the real contact area, and standard rough
surfaces exhibit a linear area-load scaling inherent to classical Archard and Greenwood-Williamson
models, severely restricting their functional range.Where the paper says it
[17] Greenwood, J.A., Williamson, J. p. 0 ch
May 20, 2026
Abstract
Programming the static friction of mechanical interfaces is critical for soft robotics, haptics,
and precision gripping. Static friction is governed by the real contact area, and standard rough
surfaces exhibit a linear area-load scaling inherent to classical Archard and Greenwood-Williamson
models, severely restricting their functional range. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems. According to the
classical theory of Bowden and Tabor 10, the static friction force is directly proportional to the real area
of contact, if adhesion effects are negligible. 2 Theoretical Framework: From Asperity Mechanics to Macroscopic
Metasurfaces
2. 1 Axisymmetric asperities contact
The solution for general axisymmetric profiles indenting a flat half-plane was first found by Foeppl 15
and Schubert29 and later on made popular by Sneddon31.Where the paper says it
Where the paper says it
Where the paper says it
[18] Kogut, L., Etsion, I., 2002. Elastic-Plastic Contact Analysis of a Sphere and a Rigid Flat. p. 4 While exact for sparse distributions, interaction effects become non-negligible at high contact densities 35. However, Boussinesq 9
showed that the elastic deformation field decays as 1/r, so that shielding effects are minimized when
the mean spacing between asperities is sufficiently large compared to the contact radius.Where the paper says it
[19] Kumar, S., Tan, S., Zheng, L., Kochmann, D.M., 2020. Inverse-designed spinodoid metamaterials. p. 1 Unlike purely data-driven surrogates, differentiable physics approaches embed analytical governing
equations directly into the learning loop, allowing for the backpropagation of physical gradients 3
. This paradigm has been successfully applied to the inverse design of nanophotonic structures 20, topology optimization in optics 12, spinodal metamaterials 19, and general mechanical structures 34.Where the paper says it
[20] Liu, D., Tan, Y., Khoram, E., Yu, Z., 2018. Training Deep Neural Networks for the Inverse Design of Nanophotonic Structures. p. 1 Unlike purely data-driven surrogates, differentiable physics approaches embed analytical governing
equations directly into the learning loop, allowing for the backpropagation of physical gradients 3
. This paradigm has been successfully applied to the inverse design of nanophotonic structures 20, topology optimization in optics 12, spinodal metamaterials 19, and general mechanical structures 34.Where the paper says it
[21] Loshchilov, I., Hutter, F., 2019. Decoupled Weight Decay Regularization. p. 8 (b) Logarithmic contact regularization schedule increasing the physical steepness boundary (κ) to enforce strict Sneddon mechanics
over the first quarter of epochs. (c) Linear decay of the curriculum weight (λ), shifting the network
from parameter-steering to pure physics reconstruction. The advent of scientific machine learning (SciML) has offered new pathways for materials design6
. Recently,
generative deep learning models, including variational autoencoders (VAEs) 24 have been applied to
1
arXiv:2605.Where the paper says it
Where the paper says it
[22] Ma, Y., Zhang, G., Cao, S., Huo, Z., Han, J., Ma, S., Huang, Z., 2023. A Review of Advances in Fabrication Methods and Assistive Technologies of Micro-Structured Surfaces. p. 0 This framework bridges
data-driven optimization and rigorous physics, offering a scale-invariant pathway for discovering
functional tribological surfaces. 1 Introduction
The precise control of static friction is a fundamental requirement for emerging technologies, from
the skin-like tactile sensors of soft robotics to high-fidelity haptic feedback systems.Where the paper says it
[23] McKay, M.D., Beckman, R.J., Conover, W.J., 1979. A Comparison of Three Methods for Selecting Values of Input Variables in the Analysis of Output from a Computer Code. p. 5 We propose the hybrid SciML architecture shown in Fig. The model consists of two coupled
blocks:
• A fully connected Multi-Layer Perceptron (MLP) which takes the target discretized and normalized physical curves (αˆ(Pˆ), and Sˆ = dα/d ˆ Pˆ) and the max values P
∗
max, αmax as input.Where the paper says it
[24] Mouton, V., Mélot, A., 2026. Friction on demand: A generative framework for the inverse design of metainterfaces. p. 0 Static friction is governed by the real contact area, and standard rough
surfaces exhibit a linear area-load scaling inherent to classical Archard and Greenwood-Williamson
models, severely restricting their functional range. Here, we propose a framework for the inverse
design of tribological metainterfaces engineered for programmable contact behaviors.Where the paper says it
[25] Nordhagen, E., Sveinsson, H.A., Malthe-Sørenssen, A., 2025. Tailoring Frictional Properties of Surfaces Using Diffusion Models. p. 1 Similarly, Denoising Diffusion Probabilistic Models (DDPMs) 25 have
been used to generate the entire surface topography, but only to achieve a specific working point of the
load-contact area curve. While these approaches excel at reconstructing complex surfaces, they often
lack precise control over the functional properties.Where the paper says it
[26] Popov, V.L., Heß, M., Willert, E., 2019. Handbook of Contact Mechanics: Exact Solutions of Axisymmetric Contact Problems. p. 0 By utilizing general axisymmetric asperities, we unlock nonlinear macroscopic responses unattainable by
standard Hertzian contacts. To solve the inverse problem, we embed a fully differentiable contact
mechanics engine within a neural network and a quadratic optimizer.Where the paper says it
[27] Sahli, R., Mikkelsen, J., Boye, M.S., Dias, M.A., Aghababaei, R., 2024. Frictional Contact of Soft Polymeric Shells. p. 0 As demonstrated by the seminal works of Archard2 and Greenwood &
Williamson17, the contact area of a random rough surface scales linearly with normal load, regardless
of the specific asperity shape: the well-known “roughness paradox”. This linearity creates a theoretical
barrier: it restricts standard interfaces to a single, Amontons-Coulomb frictional behavior, effectively
locking the coefficient of friction to a constant value.Where the paper says it
[28] Sanner, A., Kumar, N., Dhinojwala, A., Jacobs, T.D.B., Pastewka, L., 2024. Why soft contacts are stickier when breaking than when making them. p. 0 Here, we propose a framework for the inverse
design of tribological metainterfaces engineered for programmable contact behaviors. By utilizing general axisymmetric asperities, we unlock nonlinear macroscopic responses unattainable by
standard Hertzian contacts.Where the paper says it
[30] Schulze, K.D., Bennett, A.I., Marshall, S., Rowe, K.G., Dunn, A.C., 2016. Real Area of Contact in a Soft Transparent Interface by Particle Exclusion Microscopy. p. 4 The total force and real contact area are normalized by the equivalent modulus E∗ and the nominal unit cell area Anom to yield the dimensionless nominal pressure
P
∗ = P/E∗ = Ftot/(AnomE∗
) and the contact area fraction α = Atot/Anom. 3 Translating Scale-Invariant Models to Physical Interfaces
Having established a scale-invariant forward model based on the dimensionless nominal pressure (P
∗
)
and contact area fraction (α), translating these predictions into a physical, macroscopic interface follows a straightforward, deterministic procedure.Where the paper says it
[31] Sneddon, I.N., 1965. The relation between load and penetration in the axisymmetric boussinesq problem for a punch of arbitrary profile. International Journal of Engineering Science 3, 47–57
[32] Urbakh, M., Klafter, J., Gourdon, D., Israelachvili, J., 2004. The nonlinear nature of friction. p. 1 Specifically, we deploy a
neural surrogate to provide zero-shot topological initializations, which are subsequently refined using
a multi-stage optimization scheme. This architecture allows us to engineer surfaces that reproduce
complex, nonlinear contact area (and by extension friction) as a function of load trajectories (A(F)).Where the paper says it
[33] Xu, Y., Li, X., Chen, Q., Zhou, Y., 2024. Persson’s theory of purely normal elastic rough surface contact: A tutorial based on stochastic process theory. p. 0 However, for standard rough surfaces, this design space is
fundamentally constrained. As demonstrated by the seminal works of Archard2 and Greenwood &
Williamson17, the contact area of a random rough surface scales linearly with normal load, regardless
of the specific asperity shape: the well-known “roughness paradox”.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI system can perform inverse design of metainterface unit cells to reproduce complex target friction laws as a function of load trajectories (A(F)). This is achieved by embedding a fully differentiable contact mechanics engine within a neural network and a quadratic optimizer, allowing the discovery of non-standard topographies that
reproduce complex, nonlinear contact area (and by extension friction) as a function of load trajectories (A(F))
-
Improved AI system can identify optimal geometric parameters such as shape exponents gamma and height offsets h for programmable surfaces. This is accomplished through a hybrid SciML architecture where a Multi-Layer Perceptron predicts these parameters, which are then passed to a differentiable physics layer implementing the analytical forward model from Equations 6 and 7.
-
Improved AI system can navigate complex, non-convex design spaces efficiently using a multi-stage optimization scheme with L-BFGS. This refinement module utilizes
a multi-stage, memory-efficient Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm with Wolfe line search
and dynamically modulates the steepness parameter kappa to enforcestrict mathematical boundaries of standard contact mechanics.
-
Improved AI system can ensure topological consistency by enforcing a strict monotonic depth ordering on asperity heights. This is managed by anchoring the height array with h0 = 0 and dynamically sorting the active heights during each forward pass, which is
mathematically guarantees that the asperities are sorted by engagement depth.
-
Improved AI system can accurately predict macroscopic contact stiffness S star even when new asperities make initial contact. The framework accounts for this singularity by formulating it as
S star = dPstar/dα = PN i=1 dPstar i / P d∆ N i=1 dαi/d∆
and employing an L1 gradient loss to capture the resultingsharp, discontinuous cliffs in the macroscopic S star response.
-
Improved AI system can perform rigorous validation of its designs against high-fidelity simulations. The predicted unit cells are validated against
high-fidelity Boundary Element Method (BEM) simulations,
confirming that the framework's assumptions hold with a maximum relative error of less than 3% across all tested regimes.
Abstract
Programming the static friction of mechanical interfaces is critical for soft robotics, haptics, and precision gripping. Static friction is governed by the real contact area, and standard rough surfaces exhibit a linear area-load scaling inherent to classical models, severely restricting their functional range. Here, we propose a framework for the inverse design of tribological metainterfaces engineered for programmable contact behaviors. By utilizing general axisymmetric asperities, we unlock nonlinear macroscopic responses unattainable by standard Hertzian contacts. To solve the inverse problem, we embed a fully differentiable contact mechanics engine within a neural network and a quadratic optimizer. We leverage regularized physical gradients to automatically discover non-standard topographies that reproduce complex target contact laws, with only a few asperities in unit cells. The predicted designs are validated against high-fidelity Boundary Element Method (BEM) simulations. By front-loading computational costs during training, the neural surrogate yields near-instantaneous design inference, bypassing the bottlenecks of traditional multi-start optimization. This framework bridges data-driven optimization and rigorous physics, offering a scale-invariant pathway for the high-throughput discovery and real-time control of functional tribological surfaces.
Sources
Related papers
- Coarse-graining dense, deformable active particles
- Three-body corrections to the Brownian viscosity of hard spheres without hydrodynamic interactions
- Thin active nematohydrodynamic layers: asymptotic theories and instabilities
- Origins of suppressed self-diffusion of nanoscale constituents of a complex liquid
- Soft material mechanical property degradation at high strain rates
- Avalanche criticality emerges by thermal fluctuation in a quiescent glass