Rapid urbanisation and population growth have increased demand for infrastructure and sustainable waste management. Geotechnical engineering can help in addressing this challenge by repurposing cementitious waste materials to stabilise low-strength subgrades beneath pavements and railway embankments. While previous studies show the effectiveness of such byproducts in enhancing soil properties, practical application is often hindered by soil variability and a lack of robust predictive models. Predicting dry density is essential for assessing strength, compaction characteristics, and the long-term performance of stabilised subgrades. This study uses previously unexplored experimental data as a basis to compare the performance of tree-based ensembles and genetic programming in predicting the dry density. A dataset of 213 rows of data is compiled from ten experimental studies on soil stabilisation, covering a wide range of soil conditions and mix designs. Results indicate that both models generalise well to unseen data, but the ensemble extra trees regressor outperforms with a higher R2 score (0.98), lower errors, and better alignment with empirical evidence. Results also indicate that the compaction method, sand/fines ratio, and alkali activator content are critical factors influencing dry density in both models. The influence of sand/fine and ash/binder ratios is found to be nonlinear and often governed by interaction effects.