Two‐Stage Neural Network for Modeling Ductile Fracture in Structural Steel

ABSTRACT This study presents a data‐driven surrogate to the Gurson–Tvergaard–Needleman (GTN) fracture model, designed as an uncoupled formulation for computationally efficient prediction of ductile fracture initiation under high triaxiality in structural steels. Although the GTN model is widely used to simulate void nucleation, growth, and coalescence in metals, its coupled nature incurs significant computational overhead, limiting its applicability to finite element simulations of steel structures. To address this challenge, we propose a two‐stage artificial neural network (TS‐ANN) trained on 49,875 single‐element GTN simulations generated by varying six micromechanical parameters: initial void volume fraction, void nucleation volume fraction, nucleation strain, nucleation strain standard deviation, stress triaxiality, and void volume fraction at fracture. The TS‐ANN architecture decouples the material parameter mapping and stress‐state dependence into two subnetworks: Subnet‐1, a larger network evaluated once per analysis, and Subnet‐2, a compact network with only 25 parameters, evaluated at each integration point. The model achieves a mean absolute error of 0.00143 and predicts failure strain within 5% of GTN results for over 99% of the data points. When implemented in ABAQUS as an uncoupled fracture criterion, the proposed model predicts fracture displacement in ASTM A992 steel specimens with less than 5% deviation from GTN‐based results, while achieving a 35–95‐fold reduction in computational time. This approach enables accurate and scalable simulation of ductile fracture in structural‐scale systems.