Deep learning based surrogate modeling for crystal-plasticity driven predictions of fracture in aluminum alloys

A key utility of micromechanical models is as calibration tools for more computationally efficient phenomenological counterparts. This paradigm is primarily driven by the high computational cost of running these micromechanical models. In this paper, a deep operator network was employed to act as an efficient surrogate of a micromechanical model to predict ductile fracture in arbitrary age-hardenable aluminum alloys. The micromechanical model was used to generate training data by evaluating the fracture behavior of randomly generated alloys considering variable hardening parameters and crystallographic texture. The operator network was intentionally designed to require only input data obtained from a uniaxial tension test and a volumetric texture measurement, such as by X-ray diffraction or electron backscatter diffraction. The surrogate model was able to predict full material fracture surfaces under proportional loading using these data. The surrogate model was also tested using four commercial aluminum alloys; AA6014-T4, AA6061, AA6082, and AA7075-T6, showing excellent prediction accuracy. Finally, the model also achieved a four-order-of-magnitude increase in computational speed relative to the reference micromechanical model.