Spontaneous Wrinkle Collapse in Anisotropic Condensed Matter Predicted by Deep Learning

Abstract A deep learning framework is presented for predicting equilibrium nematic liquid crystal configurations, trained on datasets generated by a finite element Landau–de Gennes solver. The model, a 3D U‐Net, accurately captures both global orientational order and local defect structures, providing predictions in milliseconds compared to conventional simulations. Reliability is validated by comparing simulated director fields with experimental observations on photoaligned wrinkle substrates fabricated using azobenzene‐based photoalignment solution and reactive mesogens. Wrinkle orientations accurately report local molecular alignment, enabling direct comparison with predicted structures and POM images. The network demonstrates strong fidelity across diverse boundary conditions, including complex defect patterns and high‐order topological charges, where the spontaneous collapse of high‐energy defects and their subsequent spontaneous splitting is accurately reproduced. This approach establishes a robust, data‐driven surrogate for nematic systems, bridging theory and experiment, and offers a generalizable route toward designing and controlling topological textures in condensed matter for applications in photonics, reconfigurable optics, and defect‐engineered materials.