Learning constitutive relations from indirect observations using deep neural networks

Daniel Zhengyu Huang (黄政宇), Daniel Zhengyu Huang, Kailai Xu, Charbel Farhat, Eric DarveJournal of Computational Physics 2020 · DOI

TL;DR

This paper introduces a novel framework that uses deep neural networks to learn constitutive relations (material behavior models) from indirect observations like full-field displacement data. This approach bypasses the need for direct stress-strain measurements, which are often difficult to obtain, and demonstrates superior performance compared to traditional methods in complex solid mechanics problems.

What Problem Does This Paper Address?

Simulating complex engineering systems often requires accurate constitutive relations that describe how materials deform under stress. However, obtaining these relations directly, especially for heterogeneous or anisotropic materials, can be prohibitively expensive or experimentally challenging. Traditional methods often rely on simplified phenomenological models or require extensive, high-fidelity data. This paper addresses the challenge of learning these constitutive relations from indirect, full-field displacement data, which is more readily available from experiments or simulations. The goal is to build a coarse-grained model that can accurately predict material behavior and quantify uncertainties, even when direct stress-strain data is scarce.

Methods

The core of the proposed method is the use of deep neural networks (DNNs) as a flexible function approximator for the constitutive relation. The framework integrates several key components:

  • Finite Element Method (FEM): Used to discretize the physical system and establish the governing equations.
  • Deep Neural Networks (DNNs): Employed to represent the unknown constitutive relation, denoted as $M_\theta(u)$, where $\theta$ are the trainable parameters.
  • Automatic Differentiation (AD): Crucial for computing gradients of the loss function with respect to the neural network parameters, enabling efficient training.
  • End-to-End Training: The entire process, from input data to output prediction, is trained together. The loss function is defined as the norm of the residual force, which is minimized using optimization algorithms like L-BFGS-B.
  • Uncertainty Quantification (UQ): The framework incorporates methods to estimate uncertainties in the learned constitutive relations, stemming from factors like optimization error, data noise, and homogenization.

The key novelty lies in training the neural network based on the global response (e.g., displacement fields) rather than direct point-to-point stress-strain data. This allows the model to learn complex material behaviors from more accessible observations.

Datasets

The paper uses numerically generated data for training and evaluation. Specific datasets include:

  • Multiscale Fiber-Reinforced Thin Plate Problem: Data generated from simulations of a thin plate with varying fiber volume fractions.
  • Nonlinear Rubber Membrane Problem: Data generated from simulations of a Mooney-Rivlin hyperelastic membrane under pressure loads.

The authors do not introduce new benchmark datasets but rather use these examples to demonstrate the effectiveness of their framework.

Key Insights and Ideas

  • Learning from Indirect Data is Powerful: The framework demonstrates that full-field displacement data, which is often easier to obtain than direct stress-strain measurements, is sufficient to learn complex constitutive relations.
  • DNNs as Universal Approximators for Constitutive Models: Deep neural networks prove to be highly effective and flexible function approximators for capturing intricate material behaviors, outperforming traditional methods like piecewise linear functions and radial basis functions in certain scenarios.
  • End-to-End Training Enhances Robustness: Training the entire system (FEM + NN) in an end-to-end fashion leads to better generalization and robustness, especially in the presence of noise or complex material behavior.
  • UQ is Crucial for Reliability: Quantifying uncertainties in the learned constitutive relations is essential for predictive modeling and provides confidence intervals for predictions.
  • Discretization vs. Optimization Error Trade-off: For certain problems, the error can shift from being dominated by discretization to optimization as the mesh is refined, suggesting that overly fine meshes might not always be necessary.

Key Results

  • Superior Performance of DNNs: In the nonlinear rubber membrane problem, DNNs demonstrated better generalization and robustness to noise compared to piecewise linear functions (PL) and radial basis functions (RBF).
  • Accurate Recovery of Constitutive Relations: In the linear fiber-reinforced plate problem, the learned constitutive tensor was recovered almost exactly.
  • Effective Uncertainty Quantification: The UQ analysis showed that confidence intervals for predictions often contained the exact solutions, indicating the reliability of the framework.
  • Efficiency: The framework, leveraging automatic differentiation and efficient optimizers, allows for training with relatively small datasets.

Limitations and Future Work

  • Extrapolation Limitations: The paper notes that the supervised learning framework is incapable of extrapolating beyond the range of the training data. Predictions for stretches outside the convex hull of training data may not be accurate.
  • Space-Invariant Assumption: In some UQ analyses, the model assumes a space-invariant constitutive relation, which might not hold for heterogeneous materials.
  • Future Work:
    • Extending the framework to partially observed data.
  • Applying the method to time-dependent physical systems (e.g., viscoelastic or elasto-plastic materials).
  • Exploring alternative methods for forward propagation and gradient derivation.
  • Applying the framework to other coarse-grained models like damage mechanics and structural internal damping.

Summary updated

BibTeX
@article{2020learning,
  author = {Daniel Zhengyu Huang (黄政宇) and Daniel Zhengyu Huang and Kailai Xu and Charbel Farhat and Eric Darve},
  title = {{Learning constitutive relations from indirect observations using deep neural networks}},
  journal = {Journal of Computational Physics},
  year = {2020},
  doi = {10.1016/j.jcp.2020.109491},
}
Figure