Many of the complex patterns seen in nature occur when symmetry is broken. As a system moves from a highly symmetric state to a more ordered state, small but stable irregularities may appear. These features, known as topological defects, appear across vastly different scales, from the structure of the universe to matter in general. Because they appear wherever order forms, they provide scientists with a powerful way to understand how complex systems are organized.
Nematic liquid crystals provide a particularly useful environment to study these defects. In this type of material, the molecules are free to rotate while remaining roughly oriented in the same direction. This combination makes the liquid crystals easier to control and observe, allowing researchers to track how defects appear, move, and reconfigure over time. Traditionally, scientists use the Landau de Gennes theory to explain these structures. This is a mathematical framework that explains how molecular order breaks down within defective cores where the orientation no longer has a well-defined definition.
Accelerating defect prediction with AI
Researchers led by Professor Jun-Hee Na from Chungnam University in South Korea have introduced a method to predict stable defect patterns faster using deep learning. Their work replaces time-consuming and computationally expensive numerical simulations with an AI-based approach that provides results much more quickly.
The technique, published in the journal Small, can generate predictions in milliseconds, rather than the hours typically required by traditional simulations.
“Our approach complements slow simulations with fast and reliable predictions and facilitates systematic exploration of defect-rich regions,” says Professor Na.
Inside a deep learning model
The team built the system using a 3D U-Net architecture, a type of convolutional neural network commonly used in scientific and medical image analysis. This design allows the model to recognize both large-scale alignment and local details associated with defects. Instead of performing step-by-step simulations, the framework directly ties boundary conditions to the final equilibrium state. Boundary information is fed into the network to predict the complete molecular alignment field, including defect shape and location.
To train the model, the researchers used data from traditional simulations covering a variety of alignment scenarios. After training, the network was able to accurately predict completely new configurations that it had never encountered before. These predictions were in close agreement with the results of both simulations and laboratory experiments.
Handling complex and merged defects
The model learns material behavior directly from data, rather than relying on explicit physical equations. This gives us the flexibility to handle particularly complex cases, such as higher-order topological defects where defects are combined, split, or rearranged. Experiments confirmed that the AI accurately captured these behaviors and operated reliably under a wide range of conditions.
A faster path to advanced materials
This approach allows scientists to quickly explore many design possibilities and also creates new opportunities to design materials with carefully controlled defect structures. These features are particularly valuable for advanced optical devices and metamaterials.
“By significantly shortening the material development process, AI-driven design has the potential to accelerate the creation of smart materials for a variety of applications, from holographic and VR or AR displays to adaptive optical systems and smart windows that react to the environment,” said Professor Na.
