<p>In this research, manufactured geopolymer coarse aggregates using cut-blade mechanisms have been assessed based on transport and chemical degradation durability indicators, using physics-informed machine learning models for prediction. The aggregates are first cured in an oven at 60⁰ c for 48&#xa0;h to achieve proper geo-polymerisation and then cured in hot water for 28–90 days. Extensive preliminary trials were conducted to determine the optimal molarity of the alkali activator, as the concentration of NaOH significantly influences geopolymer gel formation and the stability of the final aggregate. After finalising the suitable molarity and activator ratio, the glass powder content was varied from 0% to 20% to examine its individual influence on durability. The results show that the FA80GP20 mix is having the best performance, with porosity reduced to about 9.85–10.1%, water absorption to 4.63–4.8%, and sulphate mass loss to 1.65–1.70%, compared to the control FA100 mix which showed porosity of 15.8–16.5%, water absorption of 8.0–8.5%, and mass loss of 3.8–4.0%. This improvement of approximately 35–40% in durability-related properties. To assess long-term behaviour, a Physics-Informed Machine Learning (PIML) model was developed by combining experimental data with established material relationships and comparing the results with a simple machine learning model. The PIML model achieved high predictive accuracy (R² = 0.94) while maintaining physical consistency, with a deviation of less than 2.1% from expected material trends. SHAP analysis showed that porosity, glass powder content, and activator ratio were the most influential features controlling durability. The findings demonstrate that integrating durability experiments with physics-guided machine learning offers a dependable framework for designing high-performance geopolymer aggregates, particularly for time-dependent durability indicators.</p>