Low-Code Automl Solutions for Predicting Bond Strength and Failure Modes of Cfrp-Steel Joints

Abstract

Carbon fibre-reinforced polymer (CFRP) has become the predominant material for strengthening steel structures, where the performance of CFRP-steel joints, crucial for effectiveness, can be influenced by various geometric and material variables. This paper introduces a low-code automated machine learning framework to investigate both the bond strength and failure models of CFRP-steel joints using popular regression and classification models separately. A dataset comprising 255 single-lap shear experimental results was utilised for model development. The ensemble learning models demonstrated exceptional predictive ability. The optimally fine-tuned Extra Trees Regressor model exhibited high accuracy in predicting bond strength, achieving an R2 of 0.819, while the optimally fine-tuned Random Forest Classifier model showed high accuracy in predicting failure modes, achieving an F1 score greater than 0.933 for all classes. Furthermore, the built-in SHapley Additive exPlanations method was applied to interpret the constructed predictive models within the low-code environment. The adhesive's Young’s modulus was identified to be the greatest impact on the bond strength, while the adhesive's tensile strength was the greatest impact on the failure mode. A web-based application was further developed to enable real-time batch forecasting, which can be used for extended research and also provide practical insights for optimised design strategies.

Keywords: Automated machine learning, CFRP-steel joints, bond strength, failure modes, Streamlit web-app