SplineNet Integrates CAD/CAE for Complex Shell Design

The design and analysis of complex shell structures in engineering often involve a cumbersome process of converting Computer-Aided Design (CAD) models into formats suitable for Computer-Aided Engineering (CAE) simulations, leading to time-consuming data exchange and potential loss of geometric fidelity. A new method, SplineNet, aims to bridge this gap by offering an isogeometric deep learning approach. SplineNet is built upon watertight spline representations, such as analysis-suitable unstructured T-splines, allowing for exact geometric descriptions of CAD models directly within the neural network architecture. It leverages Bézier extraction to construct the network, using Bernstein polynomials as non-linear activation functions. This unique integration enables a seamless workflow between design and analysis. The method supports both data-free and data-driven applications. In the data-free scenario, energy-based formulations (like the Kirchhoff-Love model for shells) can be incorporated as loss terms, allowing for accurate mechanical behavior calculations directly within the network. For data-driven applications, SplineNet can serve as the trunk network for Deep Operator Networks (DeepONet), providing interpretability and enabling immediate results for unseen input data without retraining. Numerical examples demonstrate its effectiveness, especially for real-world complex geometries.