Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models

arXiv:2609.00359 · physics.comp-ph, cond-mat.mtrl-sci · Submitted 2026-08-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity".

Title and authors: Mira: Let’s talk about the authors and what this title really implies for us as a community. The full title is "Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models."

Kai: It suggests they are trying to build something foundational, not just a niche model for one material type. They’re aiming to create a structure that works across many different material families.

Lev: When I read the title, I think it points toward solving the problem of scaling up constitutive models from specific cases to general families. It’s about finding a way to learn the rules of deformation that apply broadly.

Mira: Yeah, and they are proposing this mixture architecture specifically because it allows them to use only polyconvex strain invariant measures as inputs for each shared component, which keeps the mathematical structure sound.

Kai: So they aren't just throwing any neural network at it; they’re using specialized input convex neural networks that enforce convexity and monotonicity constraints on those basic building blocks.

Lev: That constraint enforcement is key because without it, you risk losing the polyconvexity property, which could cause problems when simulating things like material failure or complex stress states.

Mira: And the conditioning network then just needs to map the material descriptors to non-negative weights for those components. It’s a clean separation between what defines the mechanics and what defines the material itself.

Kai: That separation is what makes it modular, which is a big deal when you're dealing with complex physical systems where different inputs drive different behaviors.

Lev: For real hardware, that modularity is helpful because you can potentially train or fine-tune just the components related to specific physics if you know where the failure mode lies.

Mira: It’s a shift from fitting a single complex function to learning a small set of fundamental physical building blocks and then mixing them appropriately.

The paper's summary: Kai: To summarize what they did, they moved away from using one monolithic partially input-convex neural network, or pICNN, and instead learned several shared strain energy components whose weighted sum defines the final potential.

Mira: So the key is that each of these components only uses polyconvex strain invariant measures as inputs and satisfies convexity and monotonicity constraints on its own.

Lev: And then they have this separate conditioning network that takes your material descriptors, like composition, and outputs non-negative mixture weights for those components. That’s how you get the final strain energy density function.

Kai: The whole point is that because you form the final potential as a non-negative linear combination of these convex-strain invariant components, this entire architecture preserves polyconvexity.

Mira: This construction is motivated by the idea that varying material behaviors can be represented through a relatively small set of shared strain energy components, which might reduce optimization interference and provide a more interpretable structure.

Lev: They also mention that this modular decomposition helps in reducing optimization interference, which is something I worry about when trying to train very large networks on limited data.

Kai: It suggests that instead of learning one huge function that tries to capture every single material quirk, you learn a library of fundamental mechanical responses and then mix them.

Mira: This approach is designed specifically for modeling families of hyperelastic materials where the behavior varies with composition or microstructure, which is a huge practical win over fitting separate forms.

The paper's improvements: Lev: The paper points out several improvements over previous constructions, mainly focusing on how this mixture architecture performs in practice when applied to specific benchmarks.

Kai: In the PolyJet experimental benchmark, they found that the mixture ICNN reliably predicted the stress responses of held-out materials across independent training runs. That’s a strong indicator of good generalization.

Mira: It also showed that on the synthetic Gent-type benchmark, this mixture model generalized more reliably even when sampling was sparse in the material descriptor space. That’s a significant result for data efficiency.

Lev: The paper notes that compared to the monolithic pICNN baseline, the mixture ICNN showed strong held-out performance across tested configurations. It seems less sensitive to choices like network size or the sharpness of the softplus activation functions used in those convex layers.

Kai: That sensitivity reduction is interesting because it means you don't have to spend so much time tuning those hyperparameters just to get a reasonable fit on a new material.

Mira: And they found that this shared mixture structure provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data, which is exactly what we want in applied science.

Conclusion: Kai: So to wrap up the paper "Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models," the main point is that this shared mixture structure can improve generalization across material families, especially when you only have limited data available.

Mira: The primary value they highlight is inspecting the mixture decomposition itself, which helps with the interpretability of trained models and not necessarily identifying specific physical mechanisms right away.

Lev: From an engineering standpoint, it means that related materials can be modeled effectively using a small set of shared latent strain energy components, which simplifies the complexity immensely.

Kai: And they show that this method is less sensitive to model hyperparameters like network size or activation smoothness compared to monolithic architectures.

Mira: They also showed that the mixture architecture generalizes more reliably under sparse sampling of the material descriptor space on synthetic benchmarks, which is a huge step for data efficiency in material modeling.

Lev: For real hardware, this suggests that we might be able to use fewer training materials to get a robust constitutive model ready for deployment.

Kai: So yeah, they’re giving us a more structured way to learn the behavior of many materials at once by decomposing the problem into shared convex pieces.

Steven J. Yang, Govinda Anantha Padmanabha, D. Thomas Seidl, Nikolaos Bouklas

Cornell University · École Polytechnique Fédérale de Lausanne (EPFL) · Sandia National Laboratories

physics.comp-ph, cond-mat.mtrl-sci

Submitted: 2026-08-31

Updated: 2026-11-16

Comments: 19 Pages, 12 Images, 4 Tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 98/100

The gist: Hyperelastic constitutive models enable modeling large deformations in elastic solids by prescribing a strain energy density function and calibrating model-specific parameters from experiments In…

Key concepts

Hyperelastic Constitutive Models
These models describe how elastic solids deform under large strains by defining a strain energy density function, $\psi(F)$. For isotropic materials, this function depends only on the invariants of the deformation gradient tensor $C$. The authors focus on incompressible and nearly incompressible materials, splitting the energy into isochoric and volumetric parts.
Polyconvexity
This mathematical property ensures that a strain energy function is well-behaved for numerical simulations. It means the function can be expressed as a non-negative linear combination of convex functions. The proposed mixture architecture guarantees polyconvexity because it forms the final potential as a non-negative linear combination of convex, strain invariant components.
Mixture Architecture
This is the core proposal where several shared strain energy components, $\hat{y}_k(u)$, are learned using input convex neural networks. These components depend only on invariant measures $u$. A separate network learns non-negative mixture weights, $\hat{w}_k(v)$, based on material descriptors $v$, resulting in the total strain energy: $\psi_{iso} = igoplus_k w_k(v) [ ext{component}_k - ext{baseline}]$. This allows the model to learn material-specific behavior within a shared framework.
Material Descriptors
These are variables $v$ used by the neural network to condition or guide the learning process. Instead of training a separate model for every material, these descriptors capture key characteristics of different material families. The mixture architecture learns how the shared strain energy components should be weighted based on these descriptors, enabling generalization across unseen materials.

Terminology

Summary

Hyperelastic constitutive models enable modeling large deformations in elastic solids by prescribing a strain energy density function and calibrating model-specific parameters from experiments In this work, we propose using mixtures of convex neural potentials based on input convex neural networks as a modular and data efficient approach to modeling material families This construction preserves polyconvexity because the final potential is formed as a non-negative linear combination of convex-strain invariant components The mixture architecture generalized better to material descriptors not seen during training across both the PolyJet experimental benchmark and the synthetic Gent-type benchmark These results suggest that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data. p. 0

Where the paper says it

In the PolyJet experimental benchmark, we showed that the mixture architecture is less sensitive to model hyperparameters, while in the Gent-type benchmark it generalized more reliably with sparse data in the materialdescriptor space. These results suggest that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data.

Constitutive Theory and Invariant Representation

Hyperelastic materials are defined by their strain energy density function ψ(F), where F = ∂x/∂X is the deformation gradient mapping a material point X in the reference configuration to its deformed position x(X). For isotropic hyperelasticity, objectivity implies that the strain energy depends only on the invariants of the right Cauchy-Green tensor C = F T F. p. 2

Where the paper says it

2 Constitutive Theory for Hyperelasticity A hyperelastic material is defined by its strain energy density function ψ(F), where F = ∂x/∂X is the deformation gradient mapping a material point X in the reference configuration to its deformed position x(X), with detF > 0. For isotropic hyperelasticity, objectivity implies that the strain energy depends only on the invariants of the right Cauchy-Green tensor C = F TF.

The authors focus on incompressible and nearly incompressible materials and adopt a standard isochoric–volumetric split of the strain energy, which is written as ψ = ψiso(¯I1, ¯I2) + ψvol(J). p. 2
Where the paper says it

Accordingly, the strain energy may be written in the form ψ(I1, I2, I3), or equivalently ψ(I1, I2, J). We focus on incompressible and nearly incompressible materials and adopt the standard isochoric–volumetric split of the strain energy, ψ = ψiso( ¯I1, ¯I2) + ψvol(J), (2) where C¯ = J −2/3C, ¯I1 = trC¯ = J −2/3 I1, ¯I2 = 1 2 h (trC¯) 2 − tr(C¯2 ) i = J −4/3 I2.

The neural network learns only the isochoric contribution ψiso. p. 2
Where the paper says it

For nearly incompressible materials, volumetric changes are penalized through the volumetric energy ψvol(J). In the present work, the neural network learns only the isochoric contribution ψiso.

To ensure polyconvexity, the isochoric strain energy is constructed in the form ψiso(ξ1, ξ2) where ψiso is convex and non-decreasing in each argument. p. 2
Where the paper says it

Polyconvexity is preserved under the addition of constants. We therefore construct the isochoric strain energy in the form ψiso(ξ1, ξ2), (4) where ψiso is convex and non-decreasing in each argument.

Network Architectures and Baseline Comparison

The paper reviews several constrained neural network architectures, including Input Convex Neural Networks (ICNNs), Monotonic Neural Networks (MNNs), and the monolithic partially input-convex neural network (pICNN) as a baseline. ICNNs enforce convexity by requiring convex and non-decreasing activation functions and non-negative weight matrices for hidden states. p. 3

Where the paper says it

For the polyconvex neural potentials considered in this work, since we use shifted invariants, the network must satisfy both input convexity and input monotonicity. Accordingly, the activation functions σℓ(·) are required to be convex and non-decreasing, the weight matrices Wℓ and Wfℓ are constrained to be element-wise non-negative, and no constraints are imposed on the biases bℓ.

The monolithic pICNN architecture partitions the input vector as x0 = [u, v], where u contains polyconvex strain invariant measures and v contains material descriptor (conditioning) variables. p. 4
Where the paper says it

We adopt this approach as the baseline architecture. Let the input vector be partitioned as x0 = [u, v], where u contains polyconvex strain invariant measures and v contains material descriptor (conditioning) variables.

This results in a single constrained architecture that must capture a range of material behaviors within a shared parameterization. p. 1
Where the paper says it

In typical pICNN implementations, latent features derived from material descriptors are injected throughout the convex branch. The result is a single constrained architecture that must capture a range of material behaviors within a shared parameterization.

Mixture Architecture Formulation

The proposed mixture architecture learns several shared strain energy components yˆk(u), each depending only on the isochoric invariant-based measures u. These components are produced by a neural network satisfying the convexity and monotonicity constraints described in Section 3. p. 4

Where the paper says it

In the mixture architecture, the model produces K convex components yˆk(u), each depending only on the isochoric invariant-based measures u. These components are produced by a neural network satisfying the convexity and monotonicity constraints described in Section 3.

1. The material descriptor variables v are processed by a separate neural network that produces non-negative mixture weights wˆk(v). p. 4
Where the paper says it

These components are produced by a neural network satisfying the convexity and monotonicity constraints described in Section 3. The material descriptor variables v are processed by a separate neural network that produces nonnegative mixture weights wˆk(v).

The resulting strain energy is expressed as ψiso = X K k=1 wˆk(v) [ˆyk(u) − yˆk(u0)]. p. 4
Where the paper says it

The material descriptor variables v are processed by a separate neural network that produces nonnegative mixture weights wˆk(v). The resulting strain energy is expressed as: ψiso = X K k=1 wˆk(v) [ˆyk(u) − yˆk(u0)] .

Evaluation and Generalization Performance

The models were evaluated on an experimental multi-material PolyJet dataset and a synthetic Gent-type benchmark. In the PolyJet study, the mixture ICNN reliably predicted the stress responses of held-out materials across independent training runs. p. 6

Where the paper says it

The goal is to predict the full uniaxial tensile stress response of the held-out material. This process is repeated for all six compositions across three independent training runs, giving 18 evaluations for each model configuration.

The synthetic benchmark demonstrated that the mixture model generalized more reliably under sparse sampling of the material descriptor space. p. 0
Where the paper says it

In the PolyJet experimental benchmark, we showed that the mixture architecture is less sensitive to model hyperparameters, while in the Gent-type benchmark it generalized more reliably with sparse data in the materialdescriptor space. These results suggest that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data.

The results show that the mixture ICNN shows strong held-out performance across the tested configurations compared to the monolithic pICNN baseline. p. 7
Where the paper says it

5 activation, while increasing network size from small to large does not improve overall performance. In contrast, the mixture ICNN shows strong held-out performance across the tested configurations.

Conclusion and Future Directions

The work demonstrates that the shared mixture structure can improve generalization across material families, particularly when only limited data are available. The primary value of inspecting the mixture decomposition is to support the interpretability of the trained models and not necessarily identify physical mechanisms. p. 8

Where the paper says it

Together with the improved held-out generalization shown above, these results support the idea that related materials can be modeled effectively using a small set of shared latent strain energy components. 0 1 2 3 4 »1 0 1 2 ^yi(u) ¡ ^yi(u0) (a) Component 1 2 3 A DM40 DM50 DM60 DM70 DM85 0 1 2 3 4 wi (b) 0 1 2 3 4 »1 0 5 10 Ãiso (c) A DM 60 DM 70 DM 85 Figure 4: Learned mixture decomposition for a representative PolyJet training run: (a) latent strain energy components, (b) composition-dependent mixture weights, and (c) resulting material-dependent strain energy functions.

Future work could explore including volumetric deformation through J = detF in the learned strain energy function. p. 2
Where the paper says it

Finally, Section 6 summarizes the main findings and discusses limitations and directions for future work. 2 Constitutive Theory for Hyperelasticity A hyperelastic material is defined by its strain energy density function ψ(F), where F = ∂x/∂X is the deformation gradient mapping a material point X in the reference configuration to its deformed position x(X), with detF > 0.

The approach could also be applied to other parametric modeling problems where convex potentials provide useful mathematical structure. p. 0
Where the paper says it

Recent work has developed data-driven constitutive models that learn flexible strain energy functions while incorporating key physical constraints. In this work, we propose using mixtures of convex neural potentials based on input convex neural networks as a modular and data efficient approach to modeling material families.

Pruning redundant reasoning tokens

Hyperelastic constitutive models enable modeling large deformations in elastic solids by prescribing a strain energy density function and calibrating model-specific parameters from experiments In this work, we propose using mixtures of convex neural potentials based on input convex neural networks as a modular and data efficient approach to modeling material families This construction preserves polyconvexity because the final potential is formed as a non-negative linear combination of convex-strain invariant components The mixture architecture generalized better to material descriptors not seen during training across both the PolyJet experimental benchmark and the synthetic Gent-type benchmark.

How the training signal is built

Training is performed by minimizing the mean-squared error (MSE) loss between predicted and ground-truth stress components. Model parameters are optimized using AdamW with regularization applied through its decoupled weight decay term. p. 5

Where the paper says it

Training is performed by minimizing the mean-squared error (MSE) loss between predicted and ground-truth stress components. Model parameters are optimized using AdamW (Loshchilov and Hutter (2017)), with regularization applied through its decoupled weight decay term.

The training protocol for the PolyJet benchmark involves training on five compositions and using a leave-one-out procedure to evaluate generalization across material compositions. p. 6
Where the paper says it

0 We use a leave-one-out procedure to evaluate generalization across material compositions. In each train–test split, five compositions are used for training and the remaining one is held out for testing.

Where the gains come from

The key benefit of the mixture architecture is that the learned mixture weights can reveal how the contributions of these latent components vary across the material-descriptor space. This structure enables reliable generalization with far fewer training materials compared to monolithic architectures. p. 9

Where the paper says it

The synthetic benchmark evaluates generalization to unseen deformation modes, materials, and their combinations simultaneously. Model and Training Details: For this benchmark, we use a monolithic pICNN with convex and conditioning hiddenlayer widths of [15, 15], giving 872 trainable parameters.

The mixture model was less sensitive than the monolithic pICNNs to network size and sharpness of the Softplus activation functions used in the convex layers. p. 7
Where the paper says it

In contrast, the mixture ICNN shows strong held-out performance across the tested configurations. This suggests that representing the material family through a small set of shared strain energy components imposes a useful constraint, making the architecture less sensitive to particular choices of network size or activation smoothness.

The synthetic Gent-type benchmark showed that the mixture model generalized more reliably under sparse sampling of the material descriptor space. p. 0
Where the paper says it

In the PolyJet experimental benchmark, we showed that the mixture architecture is less sensitive to model hyperparameters, while in the Gent-type benchmark it generalized more reliably with sparse data in the materialdescriptor space. These results suggest that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data.

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

  1. Mixture of convex neural potentials can be used to model families of hyperelastic materials by learning shared latent strain energy components weighted by material descriptors, which preserves polyconvexity and simplifies simulations. This allows for a more data efficient approach to modeling material families compared to fitting separate forms for each composition.

  2. The mixture architecture can generalize better to material descriptors not seen during training, as shown in the PolyJet experimental benchmark where it was less sensitive to model hyperparameters. This suggests that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data.

  3. The system can reliably predict stress responses for held-out materials across independent training runs, as evidenced by the high test R2 scores in the PolyJet dataset. This indicates that the shared mixture structure imposes a useful constraint, making the architecture less sensitive to particular choices of network size or activation smoothness.

  4. The model can generalize more reliably with sparse data in material descriptor space on synthetic benchmarks, as seen when comparing it to monolithic pICNNs on the Gent-type benchmark. This suggests that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data.

  5. The system can effectively capture distinct mechanical behaviors within the learned latent components, as seen in Figure 4 where components specialize into distinct energy shapes corresponding to different loading paths and material compositions. This capability supports the idea that related materials can be modeled effectively using a small set of shared latent strain energy components.

Abstract

Hyperelastic constitutive models enable modeling large deformations in elastic solids. In common practice, a strain energy density function is prescribed in advance and model-specific parameters are calibrated from experiments. However, many applications require constitutive models for a family of related materials whose mechanical behavior varies with composition. A fixed constitutive model-form may not capture the full range of behavior across the family, while fitting separate forms does not provide a direct way to predict the response of new compositions. Recent work has developed data-driven constitutive models that learn flexible strain energy functions while incorporating key physical constraints. In this work, we propose using mixtures of convex neural potentials based on input convex neural networks as a modular and data efficient approach to modeling material families. Each potential is convex and monotonic with respect to polyconvex strain invariants, while a conditioning network maps material descriptors to mixture weights. We compare the approach with a monolithic partially input-convex neural network using experimental data from PolyJet 3D-printed materials and a synthetic Gent-type benchmark. Across both benchmarks, the mixture architecture generalized better to material descriptors not seen during training. In the PolyJet experimental benchmark, we showed that the mixture architecture is less sensitive to model hyperparameters, while in the Gent-type benchmark it generalized more reliably with sparse data in the material-descriptor space. These results suggest that representing a material family through a small set of shared convex neural potentials provides a useful structural prior for learning descriptor-dependent constitutive behavior from limited data.

Sources

Related papers