Multi-level intelligent design of variable angle tow laminates via image-driven method

Compared with the traditional design of constant stiffness composite laminates, the fiber variable angle-tow (VAT) laminates as variable stiffness composites have more flexible design space. In order to obtain continuous fiber paths and high-precision analysis model, the design of VAT requires more detail features, yet the large number of design variables presents a significant obstacle to conventional optimization methods. In this study, an image-driven intelligent optimization method is proposed. The VAT fiber angle vector field is accurately described by the B-spline surface with different control points, and is mapped into a gray image with fixed size. Through the isogeometric analysis (IGA) method, the buckling load is gained as the label of the images, and the dataset is obtained to train convolutional neural networks (CNNs) for deep learning models. Then, a multi-level refinement of the local key design space effectively optimizes the VAT laminates while taking into account manufacturing constraints. Due to the images providing as a single unit of input data, the CNNs model can be applied in variable design space without resampling. Three numerical examples demonstrate the effectiveness, and obtain higher buckling load of VAT laminates in different level design spaces.

Automated summary

This study proposes an image-driven intelligent optimization method for the design of fiber variable angle-tow (VAT) laminates. By accurately describing the VAT fiber angle vector field and mapping it into gray images, as well as training convolutional neural networks (CNNs) for deep learning models, the design of VAT laminates is effectively optimized.