On predicting crack length and orientation in twill-woven CFRP based on limited data availability using a physics-based, high fidelity machine learning approach

Predicting crack length and orientation in twill woven CFRP plates is notoriously non-trivial due to the interplay of complex physics over multiple length scales. Furthermore, on an industrial scale, a timely yet accurate nondestructive prediction based on a limited amount of data is critical for successful industrial adoption. This paper proposes a physics-assisted surrogate approach, by first measuring electrical conductivities using the Electrical Resistance Charge (ERC) and then developing high-fidelity through-thickness crack models resulting in 42 datasets, and processing the resulting data with a polynomial regression coupled with the leave one out cross validation using physics-based engineered features. The proposed method achieved averaged error of 0.09% and 0.48% for crack length and orientation, respectively, exhibits advantages over the Artificial Neural Network (ANN) in terms of both accuracy and tendency to overfit. This proposed approach paves the way for the maturing of the sought-after real-time Structural Health Monitoring (SHM) of crack length and orientation.