A machine learning enhanced characteristic length method for failure prediction of open hole tension composites

• Prediction of load-displacement profile improves prediction of ultimate load. • Long-short term memory neural network accurately predicts stress field. • Indirect prediction of characteristic length is accurate and interpretable. • Orders of magnitude time saving with machine learning for characteristic length. The characteristic length method is a non-local approach to predicting the failure of open and closed-hole composite features. This method requires the determination of the linear elastic stress field of the composite laminate at its failure load. Typically, this requires computationally expensive progressive damage and linear elastic modelling and simulation with finite element analysis (FEA). In this study, we demonstrate the benefit of machine learning methods to efficiently and accurately predict characteristic lengths of composite laminates with open holes. We find that the prediction of the load-displacement profile usefully informs ultimate failure load prediction. We also find that linear elastic stress fields are more accurately predicted using a long-short term memory neural network rather than a convolutional decoder neural network. We show indirect prediction of characteristic length, via prediction of failure loads and linear elastic stress fields independently, results in more flexible, interpretable and accurate results than direct prediction of characteristic length, given sufficient training data. Our machine learning-assisted characteristic length method shows over five orders of magnitude of time-saving benefit compared to FEA-based methods.