Deep learning aided inverse design of the buckling-guided assembly for 3D frame structures

Buckling-guided assembly of three-dimensional (3D) mesostructures from pre-defined 2D precursor patterns has arisen increasing attention, owing to the compelling advantages in developing 3D electronic devices and systems with novel functionalities and/or capabilities. Establishments of rational inverse design methods that allow accurate mapping of the target 3D configuration onto the initial 2D precursor pattern are crucial to the widespread application of buckling-guided assembly methods. While a few methods (e.g., those based on theoretical models and generic algorithms) have been reported for the inverse design of 3D frame structures with interconnected ribbons, limitations still exist in their applicable 3D geometries or computational efficiency. In this work, we report an effective inverse design method based on the point-cloud deep learning neural network (DLNN) model for the buckling-guided assembly of 3D frame structures. A structure-based database in the point-cloud form is established based on massive finite element analyses (FEA) of postbuckling deformations for diverse 2D precursor patterns with different numbers of intersections. The well-trained deep learning models assisted by transfer learning strategy utilizing datasets in the constructed database are verified to establish the end-to-end implicit mapping between the 3D frame structure and corresponding 2D precursor pattern. Computational and experimental demonstrations over a bunch of complexly shaped structures, including those resembling 3D shapes of real-world objects, illustrate the high efficiency and accuracy of the proposed deep learning aided inverse design method. In comparison to previously reported methods based on genetic algorithms, the proposed inverse design method can save much more computational efforts, and does not require the initial guess of the 2D precursor pattern. Furthermore, the proposed inverse design method offers an excellent extensibility, as the size and diversity of the structure-based database can be continuously expanded in a sustainable manner, with the future development of buckling-guided assembly methods.