AI (ANN, GP, and EPR)-based predictive models of bulk density, linear-volumetric shrinkage & desiccation cracking of HSDA-treated black cotton soil for sustainable subgrade

ABSTRACT AI-based bi-input predictive models have been executed to forecast the bulk density, linear and volumetric shrinkages and desiccation cracking of HSDA-treated black cotton soil (BCS) for sustainable subgrade construction purposes. The BCS was characterised and classified as A-7 group soil with high plasticity and poorly graded condition. Sawdust ash was obtained by combusting sawdust and sieving through 2.35 mm aperture sieve. It was further activated by blending it with pre-formulated activator material (a blend of 8 M NaOH solution and NaSiO2 in 1:1 ratio) to derive waste-based HSDA. The HSDA was further used in wt % of 3, 6, 9, and 12 to treat the BCS. The treated samples were compacted in the standard proctor moulds, cured for 24 h and extruded. The desiccation tests were then performed on the prepared specimens by drying them at a temp of 102°C for 30 days and behavioural changes in weight, height, diameter, average crack development, etc., were taken throughout the period. Multiple data sets were collected for the references test, and treated specimens of 3, 6, 9, and 12% wt HSDA of the soil for 30 drying days. XRF and SEM tests were also conducted to determine the pozzolanic strength via the chemical oxide composition, three chemical moduli (TCM) and the microstructural arrangement of the experimental materials and the treated BCS. The XRF tests showed that the experimental materials had less pozzolanic strength, which improved with the treated blends thereby forming stabilised mass of BCS. Also, it showed the silica moduli of the TCM dominated the stabilisation of the soil with waste-based HSDA. SEM tests showed increased formation of ettringite and gels with the addition of the HSDA. The data collected was subjected intelligent models’ prediction using ANN, GP and EPR for the four outcomes; BD, CW, LS and VS of the HSDA-treated BCS. The models’ performance showed that EPR outclassed the other techniques in predicting BD and CW with accuracies of 98.2% and 92.7% and minimal error, while ANN outclassed the other techniques in predicting LS and VS with accuracies of 98.8% and 99.3% and minimal error, respectively.