AI-Driven Utilisation of Alkali-Activated Aluminosilicate Wastes for Sustainable Soil Stabilisation

Abstract

This thesis addresses critical gaps in stabilising low-strength and expansive soils using aluminosilicate-rich by-products such as fly ash and slag. Conventional stabilisation depends on extensive mix-design testing, which is costly, time-consuming, and difficult to generalise across varying soils and curing conditions. To improve efficiency, this thesis develops predictive models for unconfined compressive strength (UCS) and one-dimensional free swelling.

The methodology has three components. First, two UCS models are developed using genetic programming (GP) and Extra Trees (ET), trained on the largest aluminosilicate stabilisation dataset to date (1,600 records from 23 studies). Second, an experimental program of one-dimensional swelling tests under monotonic and cyclic wet–dry conditions is undertaken, supported by X-ray fluorescence, X-ray diffraction, mercury intrusion porosimetry, scanning electron microscopy, UCS testing, and pH measurements to assess swelling behaviour, mechanisms, durability, and environmental impacts. Third, 260 records from the experimental program are used to develop swelling models using ET and GP.

The UCS models show considerable accuracy improvements over existing models. Total CaO content strongly increases UCS, while combined Si–Al–Fe oxides show a non-linear influence. Water content has an exponential negative effect. For swelling, alkali-liquid content is the dominant factor, reducing swelling, whereas ash content increases swelling at low activator dosages. Experiments show swelling can be reduced from 56% to under 3% within 24 hours using 1 M NaOH. Microstructural analyses show denser matrices and fewer pores, while high pH indicates environmental considerations.

Overall, the thesis advances predictive modelling and demonstrates that aluminosilicate waste is an effective, low-carbon binder. Further work should investigate alternative activators, dry-density effects, and field calibration for practical implementation.