Expansive soils pose significant challenges to infrastructure, often requiring costly repairs, due to their susceptibility to expansion and contraction in response to moisture variations. Such soils are often stabilised prior to construction. Traditional stabilisation methods rely on cement and lime, which are effective but incur considerable environmental cost, including high carbon emissions. Repurposing cementitious agro-industrial byproducts, such as fly ash and slag offers, under some conditions, a promising alternative for improving subgrades while reducing waste. However, traditional experimental methods for soil stabilisation remain laborious and costly. Solutions developed for one site are often not transferable to others, given the inherent variability in soil conditions. Consequently, there is a growing need for predictive modelling to aid stabilisation design for diverse soil, mixture design, and curing conditions while minimising reliance on extensive laboratory testing. To date, no predictive models exist for the swelling of clays stabilised with aluminosilicates, presenting a critical research gap. Building on a program of swell tests conducted by the authors (and reported elsewhere), two such models are proposed, based on deep learning (DL) and genetic programming (GP). Results indicate that both approaches effectively learn the complex patterns in the data, with reasonably high R2 scores (80%) and low errors when tested on unseen data. GP initially exhibited lower performance uncertainty, however, the Wilcoxon test showed that there is no statistically significant difference in the performance of the two models. These models help optimise mix designs, reducing the need for extensive laboratory testing.