Experimental and machine learning investigation of flexural strength and failure mechanisms in EPS-bottom ash-blast furnace slag lightweight geomaterials

The flexural performance of a new lightweight geomaterial (LWG) made of bottom ash (BA), expanded polystyrene beads (EPS), blast furnace slag (BFS) and ordinary Portland cement (OPC) was studied. Five EPS–BA mix ratios (0.2–1.0), three BFS contents (4%, 8% and 12%), two cement contents (20% and 30% of the BA weight) and three curing durations (7, 14 and 28 days) were considered for the experimental program. The flexural strength tests were performed to quantify the resistance to bending and subsequently polynomial regression, random forest regression and gradient boosting regression models were created based on the test data. The results showed that flexural strength was higher with age of curing, amount of cement and amount of BFS added and lower with higher EPS–BA mix ratio. Under the same conditions, the maximum flexural strength of 20% cement and 12% BFS and mix ratio 0.2 was 712.363kPa, while 30% cement had a flexural strength of 914.182kPa. The feature importance analysis showed that the curing age (38.6%) and cement content (27.4%) were the most influential variables and that the EPS–BA mix ratio (22.1%) and BFS content (11.9%) were the second most influential variables. Among the evaluated models, gradient boosting regression achieved the highest coefficient of determination (R 2 = 0.98) and the lowest RMSE (1.16%) and MAE (0.81%), whereas random forest regression produced the lowest MAPE (1.14%). Gradient boosting was therefore identified as the best overall model based on its combined goodness-of-fit and absolute-error performance. The study shows that the experimental and machine learning combined approach can be used for optimal design of lightweight geomaterials both in terms of mechanical properties and minimized density, which has promising applications in embankments, pavements and sustainable construction materials.