Abstract
We introduce a deep learning framework designed to train smoothed elastoplasticity models with interpretable components, such as a smoothed stored elastic energy function, a yield surface, and a plastic flow that are evolved based on a set of deep neural network predictions. By recasting the yield function as an evolving level set, we introduce a machine learning approach to predict the solutions of the Hamilton-Jacobi equation that governs the hardening mechanism. This machine learning hardening law may recover classical hardening models and discover new mechanisms that are otherwise very difficult to anticipate and hand-craft. This treatment enables us to use supervised machine learning to generate models that are thermodynamically consistent, interpretable, but also exhibit excellent learning capacity. Using a 3D FFT solver to create a polycrystal database, numerical experiments are conducted and the implementations of each component of the models are individually verified. Our numerical experiments reveal that this new approach provides more robust and accurate forward predictions of cyclic stress paths than these obtained from black-box deep neural network models such as a recurrent GRU neural network, a 1D convolutional neural network, and a multi-step feedforward model.
Introduction
In computational mechanics, predicting the elasto-plastic behavior of materials—how they stretch and permanently deform—is a classic challenge. Traditional models are often "hand-crafted" by experts using simplified mathematical expressions. However, these struggle to capture the complex, history-dependent reality of microstructures like polycrystals. While modern machine learning (ML) offers a data-driven alternative, standard "black-box" models (like Recurrent Neural Networks) often fail due to physical inconsistency (violating thermodynamics) and numerical instability (producing jagged predictions unsuitable for Finite Element Method solvers).
This research bridges the gap by creating a "gray-box" model that combines the learning capacity of deep learning with the interpretability and physical rigor of classical mechanics.
Methodology
The framework employs three novel strategies to ensure neural networks behave like real physical systems:
Sobolev Training ($H^1$ and $H^2$ norms)
Instead of just training the network to match a single value (like stress), the authors train it to match derivatives as well. By minimizing the error of the gradient (stress) and the Hessian (stiffness), the model becomes much more stable and accurate for use in numerical solvers.
Multiply Layers
Standard neural networks use activation functions like ReLU, which have "sharp" corners where the second derivative is zero. The authors introduce "Multiply layers" (element-wise products) to increase local non-linearity, ensuring the network can represent the smooth, higher-order curvatures required for elasticity and plasticity.
Hamilton-Jacobi Level Set Hardening
The "yield surface" (the boundary where a material starts to deform permanently) is treated as a level set of a higher-dimensional function. Its evolution is governed by a Hamilton-Jacobi equation, allowing the ML model to "discover" complex hardening mechanisms directly from data.
Datasets and Experimental Results
The framework was validated using a high-fidelity 3D Polycrystal Database generated with a 3D Fast Fourier Transform (FFT) solver. This database contains the homogenized responses of various Representative Volume Elements (RVEs) with different crystal orientations.
Key results include:
- Superior Generalization: The proposed model accurately predicted unloading and reloading paths even when only trained on monolithic loading data. Standard black-box models (GRU, CNN) failed, predicting continued plastic deformation during unloading.
- Numerical Robustness: Sobolev-trained networks produced smooth tangent operators (stiffness matrices), allowing seamless integration into standard Finite Element Method (FEM) software.
- Accuracy: In cyclic loading tests, the Hamilton-Jacobi hardening framework significantly outperformed GRUs and 1D CNNs in tracking the stress-strain path.
Discussion and Future Work
By embedding thermodynamic constraints (like convexity and non-negative plastic work) directly into the loss function, the model is prevented from making "unphysical" guesses. This acts as a powerful form of regularization.
While successful, the current framework assumes material isotropy and focuses on infinitesimal (small) strains. Future work will extend these level set methods to anisotropic materials and finite (large) deformation plasticity. Additionally, while inference is fast, Sobolev training is more computationally intensive due to the requirement of calculating higher-order gradients during the loss calculation.
Summary updated