To meet the market and environmental demands of the coming decades, the aerospace industry must accelerate the design and development cycles for next generation low/zero emission aircraft. However, methods for damage initiation prediction and failure prediction of large composite structures remain slow, repetitive and expert dependant.
To address these challenges, we integrate machine learning into virtual testing methods. For the first time in literature, we integrate machine learning into the one-way uncoupled global-local submodelling process to predict 3D stress distributions around design features of interest, and within each step of the characteristic length method to predict failure at the early design stage of composite airframes. Our methods are demonstrated for open-hole investigations, and considerations are made to extend the application of our methods towards a more complex feature: the bolted composite joint.
To enable our machine learning assisted virtual testing frameworks, we invent: a design of experiment method that results in well-distributed sampling of laminates across multiple parameter spaces; a work-equivalent homogenisation method that effectively reduces the input parameter space for training data; an interrogation method that permits the use of analytical solutions to inform our machine learning methods; a bearing-bypass displacement decomposition method to allow linear superposition of bolted joint stresses.
Furthermore, during the development of our machine learning assisted virtual testing frameworks, we discover: the benefit of using bi-directional neural networks over image-based neural networks for 3D stress prediction of symmetrical composite laminates; a limited benefit of pre-load on bearing strength of countersunk bolted composite joints; a promising route to approximate stress distributions of such joints given varying preload, without the use of additional training data.
The developments in this thesis contribute to an attractive compromise between model prediction accuracy, (online) model prediction speed, and (offline) model training resources for virtual testing of design features in composite airframes.