MD-inferred neural network monoclinic finite-strain hyperelasticity models for $β$-HMX: Sobolev training and validation against physical constraints

Nikolaos N. Vlassis, Puhan Zhao, Ran Ma, Thomas D. Sewell, WaiChing Sun

TL;DR

Researchers have developed a machine learning framework that predicts the complex, directional elastic behavior of $\beta$-HMX—a primary ingredient in high explosives. By combining Molecular Dynamics (MD) data with Sobolev training and physical constraints, the team created neural networks that don't just "fit the data" but strictly obey the laws of continuum mechanics and crystallographic symmetry.

What Problem Does This Paper Address?

$\beta$-HMX is a monoclinic organic crystal used in plastic-bonded explosives (PBXs). When these materials are subjected to mechanical shock, their behavior is incredibly complex: it is anisotropic (properties change depending on the direction of the force) and exists in a geometrically nonlinear regime (large deformations).

Traditional engineering models struggle here. Most "hand-crafted" hyperelastic models are designed for isotropic materials (like rubber) or simple symmetries. $\beta$-HMX’s low-symmetry monoclinic structure makes it nearly impossible to write down a perfect mathematical formula for its energy. While Molecular Dynamics (MD) simulations can provide the data, they are "noisy" due to thermal fluctuations. If you train a standard Neural Network (NN) on this noisy data, the resulting model might predict the stress correctly but fail miserably at predicting the stiffness (the derivative of stress), leading to unstable simulations.

Methods

The authors propose a multi-step machine learning pipeline to bridge the gap between atomistic simulations and continuum mechanics:

  1. Spectral Filtering: To handle MD noise, they don't just smooth the raw data. They perform a spectral decomposition of the Cauchy stress tensor and apply a moving average filter to the eigenvalues and Euler angles. This preserves the physical meaning of the stress state while removing thermal "jitter."
  2. Sobolev Training: Instead of a standard loss function that only looks at energy, they use a Sobolev norm. This forces the NN to match both the energy and its derivatives (stress). This ensures the "slope" of the learned energy functional is physically accurate.
  3. Transfer Learning for Physics: To ensure the model respects Material Frame Indifference (the model shouldn't change just because you rotate the coordinate system) and Crystallographic Symmetry, they use transfer learning. They take a pre-trained model and fine-tune it with additional loss terms that penalize any violations of these symmetries.
  4. Stability Validation: They implement a post-training check for Strong Ellipticity. Using a "Hill Climbing" optimization algorithm, they search for "worst-case" deformation scenarios to ensure the model remains numerically stable and doesn't predict non-physical material collapse.

Datasets

  • MD Database: Generated using the LAMMPS simulator with an all-atom flexible force field.
  • Loading Paths: The data includes uniaxial tension/compression, pure shear, and biaxial compression.
  • Scale: 163,400 MD data points for training and 70,030 for validation.
  • Benchmarks: The model was validated against previously published elastic constants for $\beta$-HMX (e.g., Pereverzev and Sewell, 2020) at specific pressure states (10⁻⁴ GPa and 5 GPa).

Key Insights and Ideas

  • The Derivative is King: In structural mechanics, the "tangent stiffness" (how much more force is needed for more deformation) is critical for simulation stability. Sobolev training is the "secret sauce" that makes NNs usable in implicit finite element solvers.
  • Symmetry as a Teacher: Rather than forcing a specific NN architecture to be symmetric (which is mathematically restrictive), the authors taught the network symmetry. This allows the model to capture the subtle, direction-dependent shear behavior of $\beta$-HMX that simpler models miss.
  • Anisotropy Evolves: The paper highlights that the degree of anisotropy isn't constant. As the crystal is compressed, its direction-dependent behavior becomes much more pronounced, jumping significantly once deformation exceeds 8%.

Key Results

  • High Fidelity: The NN models ($M_1$ and $M_2$) accurately reproduced MD stress-strain curves across all loading paths, including complex shear behaviors that differ in positive vs. negative directions.
  • Physical Admissibility: The models passed the strong ellipticity tests near the reference state, proving they are stable for standard engineering simulations.
  • Stiffness Accuracy: The predicted elastic coefficients matched literature values closely at low pressures. At high pressures (5 GPa), the NN models provided more accurate descriptions of the material because they accounted for geometric nonlinearities that infinitesimal strain models ignore.
  • Anisotropy Index: The researchers calculated a general anisotropy index ($A_I$), finding that for uniaxial compression along certain axes, the anisotropy can jump from a value of ~10 to over 100, a critical insight for shock physics.

Limitations and Future Work

  • The Singularity Problem: While the model shows energy growth as volume decreases (det F → 0), it does not reach the theoretical "infinite energy" required by continuum theory for a total volume collapse. The authors suggest adding "artificial" high-energy data points at the singularity in future versions.
  • Purely Elastic: The current model does not account for plastic yielding or damage. In real-world shock events, $\beta$-HMX will eventually break or deform permanently.
  • Future Direction: The team aims to integrate these hyperelastic NNs into larger Crystal Plasticity frameworks to simulate the full life cycle of an explosive from impact to detonation.

Summary updated

Abstract

We present a machine learning framework to train and validate neural networks to predict the anisotropic elastic response of the monoclinic organic molecular crystal $β$-HMX in the geometrical nonlinear regime. A filtered molecular dynamic (MD) simulations database is used to train the neural networks with a Sobolev norm that uses the stress measure and a reference configuration to deduce the elastic stored energy functional. To improve the accuracy of the elasticity tangent predictions originating from the learned stored energy, a transfer learning technique is used to introduce additional tangential constraints from the data while necessary conditions (e.g. strong ellipticity, crystallographic symmetry) for the correctness of the model are either introduced as additional physical constraints or incorporated in the validation tests. Assessment of the neural networks is based on (1) the accuracy with which they reproduce the bottom-line constitutive responses predicted by MD, (2) detailed examination of their stability and uniqueness, and (3) admissibility of the predicted responses with respect to continuum mechanics theory in the finite-deformation regime. We compare the neural networks' training efficiency under different Sobolev constraints and assess the models' accuracy and robustness against MD benchmarks for $β$-HMX.
BibTeX
@misc{vlassis2021md,
  author = {Nikolaos N. Vlassis and Puhan Zhao and Ran Ma and Thomas D. Sewell and WaiChing Sun},
  title = {{MD-inferred neural network monoclinic finite-strain hyperelasticity models for $β$-HMX: Sobolev training and validation against physical constraints}},
  year = {2021},
  eprint = {2112.02077},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.02077},
}
Figure