Hybrid Xanthan Gum–Nano-Silica Stabilization of Expansive Clay: A Metaheuristic-Optimized Machine Learning Framework for Strength and Swelling Prediction

Swelling and shrinking of expansive clays with changes in moisture content are one of the largest problems in geotechnical engineering, causing damage to foundations, pavements, and other infrastructure when moisture changes. While widely used stabilizers like cement and lime work, they have significant carbon footprints, so sustainable options are growing in interest. Biopolymer Xanthan gum and nanomaterial (NM) silica, which have separately been proven to enhance the strength and reduce the swelling of soil, have not been used in combination and no hybrid machine learning (ML) model has been built for simultaneous prediction of both properties for this pair of nanomaterials. In this study, this gap has been filled by a set of experimental studies on a high-plasticity expansive clay (LL = 58.6%, PI = 32.3%, free swell index = 82%) using a complete 4 × 4 factorial matrix of xanthan gum (0, 1, 3 and 5% by dry mass of soil) and nano-silica (0, 2, 4 and 6% by dry mass of soil), giving 16 mix combinations, each cured for 7, 28 and 90 days and tested in triplicate, for a total of 144 specimens. Four bio-inspired metaheuristic algorithms (Firefly, Cuckoo Search, the Grey Wolf Optimizer and Elephant Herding Optimization) were used to tune the kernel hyperparameters of a single Gaussian process regression (GPR) base learner, yielding four optimized variants of that learner for each of two prediction targets: unconfined compressive strength (UCS) and swelling potential (expressed as the free swell index, FSI). These variants were benchmarked against a standalone GPR baseline and against random forest, support vector regression, artificial neural network and extreme gradient boosting models trained on identical data partitions, and all models were validated using grouped cross-validation in which the three replicates of any mix–curing combination were kept together, so that no validation record is a near-duplicate of a training record. The experimental results demonstrated that the dosages of xanthan gum and nano-silica led to an increase in UCS at all curing periods, with the mean unconfined compressive strength of the treated mixes rising from approximately 455 kPa at the lowest dosage combination to approximately 1235 kPa at the highest dosage combination after 90 days of curing, and the complete specimen-level database spanning 215–1843 kPa; whereas the dosages of FSI and swelling pressure led to a decrease with increasing dosage of xanthan gum and nano-silica and curing time, with all the treated specimens remaining lower than the untreated soil's baseline swelling behavior. The GPR-Firefly model that was optimized for the Firefly gave the best predictive performance for both targets on the validation dataset: R 2 =0.984, root mean square error=52.5 kPa for UCS; R 2 =0.949, root mean square error=3.69% for FSI. The interpretability analysis using SHAP showed that the dosage of xanthan gum and nano-silica was the more influential variable in increasing strength, while the dosage of xanthan gum and dry density was the more influential variable in suppressing swelling, suggesting that while both strength gain and swelling suppression are controlled by the same four input variables, they are driven by different mechanisms. Benchmarked against machine learning frameworks previously reported for xanthan gum-treated and for nano-silica-treated soils considered separately, the present framework achieved accuracy within the range reported for those single-material studies, indicating that introducing a second additive and a second target property did not degrade predictive tractability; the comparison is indicative rather than controlled, since the studies differ in soil type, database size and partitioning strategy. The results presented here highlight the effectiveness of the hybrid system for both xanthan gum and nano-silica as stabilizer for expansive clays, and of the proposed metaheuristic-optimized machine learning system as a practical data-driven aid to mix design. The framework reduces rather than replaces laboratory testing: it is validated for interpolation within the soil type, dosage ranges, dry density range and curing ages represented in the training data, and application to a different expansive clay would require re-training or confirmatory testing.

Hali tarjima qilinmagan