Reducing Computational Costs in Phase-Field Fracture Simulations Using Machine Learning Techniques

Resumo

The study of crack propagation is critical in Structural Engineering to prevent material failure and structural collapse. Among the various approaches available, phase-field modeling stands out by transforming the sharp crack into a smooth transition between intact and damaged material. This eliminates the need for explicit tracking of the crack surface. In this method, an additional field variable is introduced to quantify the damage level at each material point, with values ranging from zero for intact material to one for fully damaged material.Phase-field models are particularly suitable for modeling heterogeneous materials, such as concrete or composite structures, due to their ability to naturally capture complex crack patterns, including branching and merging, without the need for predefined crack paths. Despite these advantages, phase-field modeling is computationally expensive and limits its applicability in large-scale simulations or in real-time scenarios.To address this limitation, the present work proposes replacing the incremental iterative procedure used for computing the phase-field variable with a machine learning model, which has been previously trained to directly predict the phase-field. Machine learning offers a promising alternative by learning complex mappings from input fields (such as stress or strain) to damage evolution patterns, potentially bypassing the need for iterative numerical solutions. This integration of traditional numerical methods with artificial intelligence aims to reduce computational time while maintaining accuracy, enabling more efficient simulations of fracture in heterogeneous materials. All simulations will be conducted using the open-source INSANE software.