Machine learning-driven digital twin for strength prediction of dissimilar adhesive joints under environmental aging

Abstract

This study investigates the integration of machine learning (ML) models with a digital twin (DT) framework to predict the performance of CFRP-to-aluminum adhesive joints under hygrothermal aging. The joints were aged naturally for one to three years and subjected to accelerated aging under hygrothermal conditions for four to 50 days. Three-point bending tests were conducted to evaluate joint performance. Five machine learning algorithms were employed to correlate natural aging periods with accelerated aging data. The results demonstrated that XGBoost achieved near-perfect prediction accuracy across all three rounds of real-time updates, indicating exceptional adaptability and reliability. In contrast, models such as SVR and linear regression exhibited limited adaptability, with higher error margins and less consistent predictions. These findings underscore the importance of selecting flexible and robust models for dynamic environments where real-time adaptation is critical. The integration of the digital twin framework with machine learning proved to be a powerful approach for real-time model adaptation and accurate performance prediction, ultimately enhancing the durability and reliability of composite structures.