Characterising the Effect of Void Morphology on Composites Strength using Deep Learning

Voids are one of the most common, and arguably most critical, manufacturing defects in composites. They are difficult to avoid and detrimental to composite performance. Research has indicated that void morphology significantly influences failure, but difficulties in understanding these influences has resulted in industry implementing conservative part rejection criteria based on average void content. To maximise material utilisation, and reduce material waste, it is crucial to adopt an approach that takes these factors into account when assessing the impact of voids on composite strength.
Currently, finite element (FE) methods are widely employed to predict the performance of laminates with voids. FE modelling of samples with voids can be at the micro-scale or meso-scale, and is generally computationally expensive due to the need for refined meshes around voids.
In this study, deep learning, specifically convolutional neural networks (CNN), was applied to experimental data in order to predict the short beam shear strength of the composite materials with voids. A total of 230 samples were manufactured using two commonly used material systems (Hexcel’s IM7/8552 and IMA/M21). All samples were CT-scanned prior to testing. Two CNN approaches were investigated: the first approach involved deep CNN networks with varying architectures and complexities, the second approach involved CT-Scan parametrisation using autoencoders that are subsequently coupled with a Gaussian Process surrogate to predict material strength.
Several parameters were analysed to optimise the performance of the CNN, including learning rate, mini-batch size and CT-Scan resolution. A 5-Fold cross-validation approach was used to evaluate the network performance, and the results demonstrated that deep learning holds significant potential in strength prediction of composites with voids.