A Micromechanics-Based Data-Driven Failure Criterion for Fiber-Reinforced Composites under Multiaxial Stress States

Reliable failure criteria for fiber reinforced composites (FRCs) under general multiaxial stress states remain challenging, particularly due to the complex interaction of different failure modes. As a result, considerable research effort has been devoted to the development of data-driven failure criterion (DDFC), combining insights from micro-mechanical modeling and machine learning (ML). However, general multiaxial stress states have largely been excluded from such studies. The present work addresses this gap by deriving a DDFC from synthetic failure data generated with a micro-scale representative volume element (RVE). A neural network (NN) is trained to distinguish FRC-typical failure modes using a dataset comprising both fully general three-dimensional stress states and reduced stress states in which only subsets of the six stress components are active, both from proportional monotonic loading. This dataset composition is essential for obtaining a criterion that is robust across different levels of stress state complexity. The proposed model achieves an overall accuracy of 88% on synthetic test data withheld throughout the training. A detailed analysis of the misclassifications shows that most erroneous predictions occur very close to the decision surface between no failure and inter fiber failure . When this uncertainty is accounted for by accepting a tolerance of ± 2% in the predicted failure stress state, the effective accuracy in distinguishing failure from no failure on the test set increases to 95%. The established framework demonstrates how synthetic data from micro-mechanical models can be used to train and critically assess ML-based failure criteria, providing a basis for further development of DDFC for uni-directional FRC and more complex microstructures.