A Review of Machine Learning for Progressive Damage Modelling of Fiber-Reinforced Composites

The accurate prediction of failure of load-bearing fiber-reinforced structures remains a challenge due to the complex interacting failure modes at multiple length scales. In recent years however, there has been considerable progress, in part due to the increasing sophistication of advanced numerical modelling technology and computational power. Advanced discrete crack and cohesive zone models enable interrogation of failure modes and patterns at high resolution but also come with high computational cost, thus limiting their application to coupons or small-sized components. Adaptively combining high-fidelity with lower fidelity techniques such as smeared crack modelling has been shown to reduce computational costs without sacrificing accuracy. On the other hand, machine learning (ML) technology has also seen an increasing contribution towards failure prediction in composites. Leveraging on large sets of experimental and simulation training data, appropriate application of ML techniques could speed up the failure prediction in composites. While ML has seen many uses in composites, its use in progressive damage is still nascent. Existing use of ML for the progressive damage of composites can be classified into three categories: (i) generation of directly verifiable results, (ii) generation of material input parameters for accurate FE simulations and (iii) uncertainty quantification. Current limitations, challenges and further developments related to ML for progressive damage of composites are expounded on in the discussion section.