.png){kind=icon}
.png){kind=icon}
Return
A data-driven machine learning framework for predicting the mechanical performance and failure mode of polyurethane adhesive joints
.png){kind=icon}
PRE
AI
L
S
P
DOI：10.1080/01694243.2026.2683015.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Polyurethane (PU) adhesives are widely used in aerospace and automotive industries for bonding dissimilar materials. However, accurately predicting the lap-shear strength of such joints is challenging due to complex, non-linear interactions among parameters like bondline thickness, curing time, substrate types, and aging behaviour. This study presents a data-driven machine learning (ML) framework to predict both the mechanical performance (lap-shear strength and extension) and the dominant failure mode of PU adhesive joints. An experimental dataset comprising 330 data points was generated using aluminium (Al), galvanized steel, and glass fiber-reinforced polymer (GFRP) substrates bonded with PU adhesive. Seven input features (three numerical, four categorical) were used to train and validate multiple ML models, including random forest (RF), gradient boosting (GB), extreme GB (XGB), and multilayer perceptron (MLP). For regression tasks, GB and XGB models demonstrated superior accuracy and robustness. On the final test set, GB and XGB achieved the highest R2 scores for predicting lap-shear strength (0.82) and extension (0.93), both with the lowest mean absolute error. For classification of dominant failure modes, which involved an imbalanced dataset, logistic regression (LR) model emerged as the most suitable model. Although GB attained the highest test accuracy (0.89), LR model was uniquely capable of correctly identifying all critical substrate/substrate (S/S) failures. These results demonstrate the potential of ML-based approaches to efficiently and accurately predict the complex behaviour of adhesive joints.
Keywords:
Polyurethane
adhesive bonding
machine learning
lap-shear strength
failure mode analysis
gradient boosting
Journal
J
IF:
3.7
Papers:
340
Citations:
6.8K
Organization
Cited Papers