Neural Network-Based Tensor Model for Nematic Liquid Crystals with Accurate Microscopic Information

Baoming Shi, Apala Majumdar, Lei ZhangarXiv 2024 · DOI · arXiv

This research paper, recently published by researchers from Peking University and the University of Strathclyde, introduces a novel machine-learning approach to modeling liquid crystals. By bridging the gap between microscopic accuracy and macroscopic efficiency, the authors provide a new tool for designing the next generation of optical and biotechnological devices.

TL;DR

The authors present the NN-Tensor model, a neural network-based framework that brings molecular-level accuracy to macroscopic liquid crystal simulations. It outperforms the industry-standard Landau-de Gennes model by accurately capturing complex phase transitions and defect structures while remaining computationally efficient and "mesh-free."

What Problem Does This Paper Address?

Nematic Liquid Crystals (NLCs)—the materials behind your laptop screen and many biological sensors—are notoriously difficult to model. Scientists typically have to choose between two extremes:

  1. Molecular Models: These are highly accurate but computationally "expensive" because they track the orientation of every molecule (the "curse of dimensionality").
  2. Macroscopic Models (like Landau-de Gennes/LdG): These are fast because they use a simplified "Q-tensor" to describe the average orientation. However, they are "phenomenological"—meaning their parameters are often educated guesses that fail at low temperatures or near critical phase transitions.

The gap is a "trade-off dilemma": you can have speed or you can have physical truth, but rarely both. This paper aims to provide both.

Methods

The researchers developed a two-step neural network approach:

  1. The NN-Tensor Model: Instead of using a simple polynomial to calculate energy (as LdG does), they trained a neural network to act as a surrogate for the complex molecular energy.
  2. Input: The macroscopic $\mathbf{Q}$-tensor.
  3. Output: The bulk energy density derived from the microscopic Maier-Saupe molecular model.
  4. Innovation: To keep the model efficient, they used discrete spherical harmonics to represent the molecular distribution and ensured frame indifference (the physics stays the same even if you rotate the coordinate system) by encoding the $\mathbf{Q}$-tensor via its eigenvalues and invariants.
  5. The Configuration Solver: They embedded this energy model into a second neural network. This network learns to find the most stable arrangement of liquid crystals in a given space. Because it uses a neural network to represent the field, it is mesh-free, meaning it doesn't require the complex geometric grids used in traditional simulations.

Datasets

The model was trained and validated using synthetic data generated from the Maier-Saupe (M-S) molecular model.

  • They used the Bingham closure method to create a mapping between macroscopic $\mathbf{Q}$-tensors and their corresponding microscopic Orientation Distribution Functions (ODFs).
  • Evaluation was performed across various 2D and 3D geometries, including regular polygons, discs with square holes, cuboids, and asymmetric spherical shells.

Key Insights and Ideas

  • Learning the "Why": The NN-Tensor model doesn't just predict behavior; it incorporates physical parameters derived directly from molecular information. This allows it to succeed where LdG fails—specifically at low temperatures where molecular constraints become rigid.
  • Capturing the Transition: The model correctly identifies the Isotropic-Nematic (I-N) phase transition. In 3D, it accurately predicts a first-order transition, matching the molecular theory's energy curves almost perfectly.
  • Geometry Independence: Because the solver is mesh-free, it can easily handle "non-regular confinements" (weirdly shaped containers) that would be a nightmare to program using traditional finite-difference methods.

Key Results

  • Precision: The NN-Tensor model achieves energy precision comparable to full molecular models but at a fraction of the computational cost.
  • Phase Accuracy: In 2D, it correctly identified the transition at $\kappa^* = 4$. In 3D, it matched the molecular model's transition temperature ($\kappa^* \approx 6.7-7.5$), whereas the LdG model only works near the transition point and fails as the temperature drops.
  • Defect Discovery: The model successfully identified complex topological defects, such as the "radial hedgehog" in spheres and the migration of defects in polygons as the system size ($\lambda$) changes.
  • Speed: Numerical tests showed the neural network approach reaches energy minima faster than classical numerical methods.

Limitations and Future Work

  • Elastic Simplification: The current study uses a "one-constant" elastic energy density. The authors acknowledge that real-world liquid crystals often have more complex elasticity, which the NN-Tensor model could be expanded to learn in the future.
  • Phase Variety: While this paper focuses on the Nematic phase, the authors suggest the framework is flexible enough to be applied to other phases, such as Smectic (layered) liquid crystals or materials made of complex, "bent-core" molecules.
  • Beyond LCs: The methodology could potentially be exported to other physical systems described by Landau-style phase transition theories, such as magnetic materials or superconductors.

Summary updated

Abstract

The phenomenological Landau-de Gennes (LdG) model is a powerful continuum theory to describe the macroscopic state of nematic liquid crystals. However, it is invariably less accurate and less physically informed than the molecular-level models due to the lack of physical meaning of the parameters. We propose a neural network-based tensor (NN-Tensor) model for nematic liquid crystals, supervised by the molecular model. Consequently, the NN-Tensor model not only attains energy precision comparable to the molecular model but also accurately captures the Isotropic-Nematic phase transition, which the LdG model cannot achieve. The NN-Tensor model is further embedded in another neural network to predict liquid crystal configurations in a domain-free and mesh-free manner. We apply the NN-Tensor model to nematic liquid crystals in a number of two-dimensional and three-dimensional domains to demonstrate it can efficiently identify rich liquid crystal configurations in both regular and non-regular confinements.
BibTeX
@misc{shi2024neural,
  author = {Baoming Shi and Apala Majumdar and Lei Zhang},
  title = {{Neural Network-Based Tensor Model for Nematic Liquid Crystals with Accurate Microscopic Information}},
  year = {2024},
  eprint = {2411.12224},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.12224},
}
Figure