NEW PROBABILISTIC FRACTURE MECHANICS APPROACH WITH NEURAL NETWORK-BASED CRACK MODELING: ITS APPLICATION TO MULTIPLE CRACKS PROBLEM

摘要

This paper proposes a new method to incorporate sophisticated crack models, such as interaction and coalescence of multiple surface cracks, into probabilistic fracture mechanics (PFM) computer programs using neural networks. First, hundreds of finite element (FE) calculations of a plate containing multiple surface cracks are performed by parametrically changing crack parameters such as sizes and locations. A fully automated 3D FE analysis system is effectively utilized here. Second, the back-propagation neural network is trained using the FE solutions, i.e. crack parameters vs. their corresponding stress intensity factors (SIFs). After a sufficient number of training iterations, the network attains an ability to promptly output SIFs for arbitrary combinations of crack parameters. The well trained network is then incorporated into the parallel PFM program which runs on one of massively parallel computers composed of 512 processing units. To demonstrate its fundamental performances, the present computer program is applied to evaluate failure probabilities of aged reactor pressure vessels considering interaction and coalescence of two dissimilar semi-elliptical surface cracks.