🔬 materials science

Concurrent enforcement of polyconvexity and true-stress-true-strain monotonicity in incompressible isotropic hyperelasticity: application to neural network constitutive models

This paper establishes that polyconvexity implies true-stress-true-strain monotonicity in incompressible isotropic hyperelasticity, a finding used to design and calibrate four physics-augmented neural network architectures that reveal significant differences in extrapolation performance despite satisfying identical constitutive constraints.

Original authors: Maximilian P. Wollner, Dominik K. Klein, Herbert Baaser, Gerhard A. Holzapfel, Patrizio Neff

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maximilian P. Wollner, Dominik K. Klein, Herbert Baaser, Gerhard A. Holzapfel, Patrizio Neff

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where materials like rubber, silicone, and soft biological tissues could be perfectly predicted by a computer. Engineers and scientists have long sought to write the "rules" that govern how these substances stretch, twist, and bounce back. These rules, known as constitutive models, are essentially mathematical recipes that tell a simulation how a material will react when pulled or squeezed. For decades, researchers have relied on fixed formulas, often named after the scientists who discovered them, to describe this behavior. However, these traditional formulas can be rigid; they struggle to capture the complex, messy reality of real-world materials, and choosing the right one requires deep expertise. In recent years, a new approach has emerged: using artificial intelligence, specifically neural networks, to learn these rules directly from experimental data. These "physics-augmented" networks are flexible and powerful, but they come with a risk. If not carefully guided, they can learn the training data perfectly but then make wild, physically impossible predictions when asked to guess how a material behaves in situations they have never seen before.

A team of researchers set out to solve this problem by building a safety net into the very architecture of these neural networks. They focused on a specific type of material behavior called incompressible hyperelasticity, which describes materials like rubber that change shape without changing their volume. The team investigated two fundamental physical principles that any realistic material model must obey. The first is a mathematical guarantee of stability, ensuring the material doesn't spontaneously collapse or behave erratically under stress. The second is a simple, intuitive rule: as you stretch a material further, the force required to stretch it should always increase. The researchers wanted to know if they could force their neural networks to obey these rules by design, and if doing so would guarantee that the networks would behave reasonably even when pushed far beyond the data they were trained on.

The researchers began by revisiting a classic mathematical proof from the 1970s that established conditions for material stability. They demonstrated that for rubber-like materials that do not change volume, satisfying the strict stability condition automatically ensures that the force required to stretch the material always increases as the stretch increases. This was a significant theoretical finding because it meant that by building a neural network to satisfy the complex stability rule, they could be confident the simpler "more stretch equals more force" rule would follow naturally. With this theoretical foundation laid, they constructed four different types of neural network architectures. Each was designed to respect these physical laws from the very beginning, but they used different mathematical ingredients to do so. Some networks were built using standard measures of deformation, while others used more complex, specialized measures of how the material stretches.

To test these models, the team fed them real experimental data from three different types of soft rubber-like materials. They calibrated the networks so that they could accurately reproduce the stress and stretch measurements from the lab tests. In the range where data existed, all four networks performed well, successfully learning the behavior of the materials. However, the true test came when the researchers asked the models to predict what would happen outside the range of the experimental data, a process known as extrapolation. This is where the models diverged sharply. Even though all four networks were built to obey the same physical laws and fit the training data equally well, they predicted very different behaviors when stretched to extreme lengths. Some models predicted that the material would continue to stiffen indefinitely, while others suggested the material would eventually soften or behave in ways that, while mathematically stable, felt physically counterintuitive.

The study revealed a crucial insight: satisfying the known physical constraints is necessary but not sufficient to guarantee a model will predict the future correctly. The researchers found that the specific way a neural network is structured—the choice of mathematical ingredients it uses to describe the material—has a profound impact on how it behaves when it has to guess. Two models might look identical when tested against known data, yet one might be a reliable predictor for extreme conditions while the other fails to capture the true nature of the material. This suggests that simply adding more data or making the network larger is not the answer. Instead, the choice of the underlying mathematical framework is just as important as the data itself. The authors conclude that while these physics-guided networks are powerful tools, they do not automatically possess the ability to predict the unknown with certainty. To truly model the future behavior of complex materials, scientists may need to incorporate even more physical intuition into the design of these networks, moving beyond just stability and monotonicity to capture the deeper, more subtle ways real materials respond to extreme forces.

Technical Summary: Concurrent Enforcement of Polyconvexity and True-Stress–True-Strain Monotonicity in Incompressible Isotropic Hyperelasticity

Problem Statement
The design of physics-augmented neural networks (PANNs) for constitutive modeling has gained significant traction due to their flexibility in approximating complex material behaviors. However, a critical challenge remains: ensuring that these data-driven models satisfy fundamental physical constraints, particularly when extrapolating beyond the calibration regime. In the context of incompressible isotropic hyperelasticity, two primary constitutive constraints are widely considered physically reasonable: polyconvexity (which implies material stability and rank-one convexity) and true-stress–true-strain monotonicity (TSTS-M), the requirement that stress increases monotonically with strain.

While polyconvexity is often enforced in PANN architectures, its relationship with TSTS-M in incompressible materials has been a subject of theoretical ambiguity. In the compressible case, polyconvexity does not guarantee TSTS-M. The authors investigate whether this relationship simplifies under the assumption of incompressibility and how these theoretical insights impact the practical performance and extrapolation capabilities of different PANN architectures.

Methodology
The paper proceeds in two distinct phases: theoretical derivation and numerical application.

Theoretical Framework:
- The authors revisit the classical framework of isotropic incompressible hyperelasticity, focusing on the interplay between polyconvexity and TSTS-M.
- They provide a traceable proof of Ball's (1976) sufficient conditions for polyconvexity using the theory of majorization, offering a clearer explanation of the underlying mechanics.
- They demonstrate that for a large class of incompressible strain-energy functions, Ball's sufficient conditions for polyconvexity automatically imply TSTS-M (and Hill's inequality). This is established by showing that the strict convexity of the strain-energy function in the logarithmic principal stretches (Hencky strain) is a consequence of these conditions.
- The authors clarify that while these conditions are sufficient, they are not necessary, and they discuss open questions regarding whether polyconvexity generally implies TSTS-M in incompressibility without relying on Ball's specific conditions.
PANN Architecture and Calibration:
- Four distinct PANN architectures are constructed, all designed to satisfy the derived constitutive constraints a priori:
  - PANN– $I$ : Based on principal invariants of the right Cauchy–Green tensor ( $I_1, I_2$ ).
  - PANN– $\sqrt{I}$ : Based on the square roots of these invariants.
  - PANN– $\lambda$ : Based on principal stretches ( $\lambda_k$ ) with permutation invariance enforced via group symmetrization.
  - PANN– $\nu$ : Based on signed singular values with $\Pi(3)$ -invariance.
- These models utilize convex-monotonic neural networks (CMNNs) or input-convex neural networks (ICNNs) with softplus activation functions to ensure the mathematical properties of the strain-energy function.
- The models are calibrated against experimental data from three soft rubber-like materials: Treloar's vulcanized rubber, an EPDM polymer, and a 3D printed DLP material.
- Performance is evaluated based on interpolation accuracy (mean squared error) and, crucially, extrapolation behavior in uniaxial, equibiaxial, and pure shear loading scenarios.

Key Contributions

Theoretical Link: The paper establishes that for incompressible isotropic hyperelasticity, Ball's sufficient conditions for polyconvexity are sufficient to guarantee TSTS-M. This provides a rigorous theoretical basis for constructing PANNs that satisfy both stability and monotonicity constraints simultaneously.
Architectural Comparison: The study constructs and compares four distinct PANN parametrizations. It demonstrates that satisfying the same constitutive constraints does not guarantee identical approximation power; the choice of input variables (invariants vs. stretches vs. signed singular values) significantly influences the model's ability to fit specific material behaviors.
Extrapolation Analysis: A central finding is that while different PANN architectures may achieve comparable interpolation accuracy (fitting the training data well), they exhibit pronounced and divergent behaviors when extrapolating to large strains.
Curvature Limitation: The authors observe that while the enforced constraints ensure monotonic stress increase, they do not constrain the curvature of the stress-strain response. Consequently, some PANN models predict a stress response that increases monotonically but with negative curvature (softening behavior) at large strains, which contradicts the intuitive expectation of stiffening for ideal elastic bodies.

Results

Interpolation: All four PANN models generally provided good fits to the experimental data, though performance varied by material. For the EPDM and DLP materials, the signed singular value-based model (PANN– $\nu$ ) outperformed the invariant-based models (PANN– $I$ and PANN– $\sqrt{I}$ ), suggesting that invariant-based parametrizations may have inherent limitations in flexibility for certain material responses.
Extrapolation: Significant deviations were observed in the extrapolation regime. Even models with nearly identical interpolation errors produced vastly different stress predictions at large stretches.
- Invariant-based models (PANN– $I$ ) showed a tendency for the stress response to flatten or exhibit negative curvature at high stretches due to the specific derivatives of the invariants with respect to principal stretches.
- In contrast, stretch-based models (PANN– $\lambda$ ) and signed singular value models (PANN– $\nu$ ) maintained different extrapolation trends.
Stability: All models maintained monotonicity and stability (no rank-one convexity violations) due to the architectural constraints, confirming the theoretical derivation.

Significance and Conclusions
The paper concludes that while physics-augmented neural networks successfully enforce fundamental constitutive constraints like polyconvexity and TSTS-M, these constraints alone are insufficient to guarantee physically reasonable extrapolation. The choice of parametrization (the "ansatz function") acts as a hidden bias that dictates the model's behavior outside the training domain.

The authors highlight a critical gap in current constitutive modeling: the set of constraints (polyconvexity + TSTS-M) restricts the slope of the stress response but fails to constrain its curvature. This leads to a fundamental ambiguity where multiple models fit the data equally well but predict different physical realities upon extrapolation. The paper suggests that for "idealized elasticity," additional constraints regarding the curvature of the Cauchy stress might be necessary, though formalizing such constraints remains an open challenge.

Ultimately, the work serves as a cautionary note for the PANN community: high approximation power and adherence to basic physical laws do not automatically equate to definite predictive power. The selection of a specific PANN architecture involves a modeling choice that significantly impacts the reliability of predictions in unobserved deformation regimes.

More like this