Polyconvex neural network models of thermoelasticity

Jan N. Fuhg, Asghar Jadoon, Oliver Weeger, Daniel Seidl, Reese E. Jones

TL;DR

This paper introduces a novel framework for modeling thermoelasticity using physics-constrained neural networks. By embedding fundamental thermodynamic and material theoretic requirements into the network architecture, they achieve accurate predictions of material behavior across varying temperatures, even for complex phenomena like thermal inversion. This work paves the way for more robust and generalizable constitutive models in computational solid mechanics.

What Problem Does This Paper Address?

Accurately simulating the behavior of materials under varying temperatures (thermoelasticity) is crucial for many engineering applications, especially with the rise of advanced manufacturing and micro-scale components. Traditional constitutive models, while grounded in physics, often rely on pre-defined functional forms that can limit their accuracy and flexibility. Machine learning, particularly neural networks (NNs), offers a powerful alternative due to its ability to learn complex, nonlinear relationships from data. However, simply applying NNs to material modeling can lead to unphysical predictions or poor generalization. The challenge lies in developing NN models that not only fit experimental data but also inherently satisfy fundamental physical principles like thermodynamic consistency and polyconvexity, which are essential for ensuring stable and well-posed simulations. Existing work has made progress in hyperelasticity, but a general, physics-constrained framework for thermo-hyperelasticity, which captures temperature-dependent behaviors beyond simple scaling, has been lacking.

Methods

The authors extend a previously developed polyconvex hyperelastic neural network framework to handle thermo-hyperelasticity. The core idea is to define the material's free energy function as a neural network that explicitly incorporates thermodynamic and material theoretic constraints.

Key methodological aspects include:

  • Physics-Augmented Neural Networks (PINNs): The neural networks are designed to satisfy specific physical requirements. For thermoelasticity, this involves ensuring the Helmholtz free energy is polyconvex with respect to deformation invariants and concave with respect to temperature.
  • Polyconvexity and Concavity Constraints: The paper details how to enforce these constraints within the NN architecture. This is achieved by carefully designing the network's activation functions and weight structures, drawing inspiration from Input Convex Neural Networks (ICNNs) and related architectures.
  • Separable Free Energy Representation: The free energy function is represented as a sum of products of functions of temperature and deformation invariants. This separable form allows for independent modeling of thermal and mechanical contributions and facilitates the imposition of constraints.
  • Lº Regularization for Sparsification: To prevent overfitting and improve generalization, especially with limited experimental data, the authors employ a smoothed Lº regularization technique. This prunes the number of trainable parameters in the neural networks, leading to more interpretable and parsimonious models.

Datasets

The authors evaluate their framework using a combination of synthetic and experimental datasets:

  • Synthetic Data: Generated using analytic free energy models to demonstrate the framework's ability to capture expected thermomechanical phenomena.
  • Experimental Data:
    • Temperature-dependent uniaxial tension and tension-torsion experimental datasets from the literature.
  • Specific datasets include:
  • Moshin and Treloar [74] for tension-torsion of butyl rubber.
  • Zhang et al. [73] for tensile data of pig ear tissue.
  • Fu et al. [75] for uniaxial tension of carbon-filled black rubber.

Key Insights and Ideas

  • Embedding Physics for Robustness: The most significant insight is that explicitly embedding thermodynamic and polyconvexity constraints into the neural network architecture is crucial for developing reliable and generalizable material models, rather than relying solely on data fitting.
  • Capturing Complex Thermal Behavior: The framework successfully models non-self-similar temperature dependencies, including thermal expansion, thermal softening, and even thermal inversion, which are challenging for simpler models.
  • Sparsification for Interpretability and Generalization: The use of Lº regularization not only prevents overfitting but also leads to models with a reduced number of effective parameters, enhancing interpretability and improving generalization to unseen data.
  • Separable Free Energy as a Powerful Ansatz: Representing the free energy in a separable form (product of temperature and deformation functions) simplifies the imposition of constraints and allows for independent learning of mechanical and thermal contributions.

Key Results

The paper demonstrates the efficacy of the proposed framework through several examples:

  • Synthetic Data: The model accurately reproduces the behavior generated by analytic thermoelastic models.
  • Experimental Data:
    • Rubber Tension-Torsion: The model achieved less than 4% error in predicting forces and moments for butyl rubber, with the purely mechanical term being eliminated as redundant.
  • Pig Tissue Tension: The model accurately represented uniaxial tension data and predicted plausible equibiaxial responses at various temperatures, generalizing well outside the training temperature range. The Lº regularization resulted in models with approximately 20 trainable parameters.
  • Carbon-Filled Rubber Tension: The model successfully captured the stress inversion phenomenon observed in the experimental data, a complex behavior not easily modeled by simpler approaches. The resulting models were also plausible in equibiaxial loading.

Limitations and Future Work

  • Limitations:
    • The paper notes that for some datasets, the stress response was self-similar with temperature, which simplified the modeling. The true power of the framework is demonstrated in cases with non-self-similar behavior.
  • The current framework primarily focuses on thermoelasticity.
  • Future Work:
    • Anisotropic Materials: The formalism is extensible to anisotropic thermo-hyperelasticity by incorporating additional invariants related to structure tensors.
  • Time and History Dependence: Extending the framework to include time- and history-dependent processes influenced by temperature effects, such as viscoelasticity and elastic damage, is a key future direction. This would involve incorporating non-equilibrium potentials, internal variables, and their flow, potentially using neural ordinary differential equations.

Summary updated

Abstract

Machine-learning function representations such as neural networks have proven to be excellent constructs for constitutive modeling due to their flexibility to represent highly nonlinear data and their ability to incorporate constitutive constraints, which also allows them to generalize well to unseen data. In this work, we extend a polyconvex hyperelastic neural network framework to thermo-hyperelasticity by specifying the thermodynamic and material theoretic requirements for an expansion of the Helmholtz free energy expressed in terms of deformation invariants and temperature. Different formulations which a priori ensure polyconvexity with respect to deformation and concavity with respect to temperature are proposed and discussed. The physics-augmented neural networks are furthermore calibrated with a recently proposed sparsification algorithm that not only aims to fit the training data but also penalizes the number of active parameters, which prevents overfitting in the low data regime and promotes generalization. The performance of the proposed framework is demonstrated on synthetic data, which illustrate the expected thermomechanical phenomena, and existing temperature-dependent uniaxial tension and tension-torsion experimental datasets.
BibTeX
@article{fuhg2024polyconvex,
  author = {Jan N. Fuhg and Asghar Jadoon and Oliver Weeger and Daniel Seidl and Reese E. Jones},
  title = {{Polyconvex neural network models of thermoelasticity}},
  journal = {Journal of the Mechanics and Physics of Solids},
  year = {2024},
  doi = {10.1016/j.jmps.2024.105837},
}
Figure