Moisture Ingress Effects on Polymer-Based Composite Structures Using Machine Learning-Assisted Stress Prediction

This study investigates the effects of moisture ingress on carbon-fibre-reinforced polymer (FRP) and glass- fibre-reinforced polymer (GFRP) laminates and sandwich structures exposed to a 3.5% salinity environment for 16 weeks. The moisture exposure was carried out using a salt-spray chamber and moisture uptake recorded using gravimetric method until saturation for all samples. Tensile and flexural tests were conducted on unexposed and salt-exposed specimens, and the measured stress-response data were subsequently used for condition-specific ML-assisted stress prediction. Specimen-independent validation was performed for 15 condition-specific datasets containing multiple independently identifiable specimen files. The experimental results showed higher moisture uptake in sandwich structures than in laminates, with seawater exposure reducing the load-bearing capacity, stiffness, strength, and modulus of all specimen types. Under specimen-independent nested validation, the condition-specific models achieved a mean held-out-specimen R² of 0.896 across 72 outer-fold evaluations, with a median R² of 0.951. Descriptively, specimen-level predictive performance was higher for laminate than sandwich datasets and for tensile than flexural datasets, while greater specimen-to-specimen variability was observed for several sandwich and flexural conditions. Model selection was condition-dependent, with KNN, ANN, HistGBR, and SVR selected most frequently, while linear regression was also preferred in several outer folds, indicating that increased model complexity was not universally required. The validation demonstrates within-condition generalisation to previously unseen specimens while retaining condition-specific prediction models. The findings provide a specimen-independent assessment of condition-specific stress-response predictability in moisture-exposed FRP laminates and sandwich structures.