A generalized machine learning framework for brittle crack problems using transfer learning and graph neural networks

Despite their recent success, machine learning (ML) models such as graph neural networks (GNNs), suffer from drawbacks the need for large training datasets and poor performance unseen cases. In this work, we use transfer (TL) approaches to circumvent retraining with datasets. We apply TL an existing ML framework, trained predict multiple crack propagation stress evolution in brittle materials under Mode I loading. The new ACCelerated Universal fRAcTure Emulator (ACCURATE), is generalized a variety of problems by using sequence update steps including (i) arbitrary lengths, (ii) orientations, (iii) square domains, (iv) horizontal (v) shear loadings. show that small 20 simulations each step, ACCURATE achieved high prediction accuracy Mode-I Mode-II intensity factors, paths these problems. demonstrate ACCURATE’s ability growth cases involving combination boundary dimensions lengths orientations both tensile also significantly accelerated simulation times up 2 orders magnitude faster (200x) compared XFEM-based fracture model. framework provides universal computational mechanics model can be easily modified or extended future work.