Inverse design of ideal pre-stress distribution in assembly interface based on service performance

Abstract

This paper proposes an inverse design method (IPIDM) that integrates deep learning with FEM for assembly interfaces under extreme service conditions. IPIDM provides a universal framework to inversely derive the contact stress distribution at the assembly stage (referred to as the ideal pre-stress distribution) from a uniform stress distribution on the assembly interfaces during the service state. This distribution is designed to ensure uniform contact stress under extreme service conditions. Concurrently, IPIDM predicts morphology layouts to guide manufacturing. In IPIDM, the displacement of mesh nodes is utilized to extract the mapping relationship between interface morphology and stress. A U-Net-based deep learning network is developed and trained on this mapping model to simultaneously output the static contact stress distribution and the corresponding morphology layout. Compared with open-source neural networks, the proposed model demonstrates superior capabilities in global feature extraction and training efficiency. The training stability and predictive accuracy of IPIDM are verified, and the optimization effects of the predicted morphology layouts are verified through FEM. Results indicate that IPIDM significantly outperforms mainstream assembly interfaces optimization methods in optimization efficiency, particularly for 3D interfaces subjected to complex stress states. This makes IPIDM a promising tool for fast FEM simulations and digital twin applications.