Interaction-driven improvement of problematic soils using lime and nano-zeolite: experimental investigation and machine learning prediction

Abstract

This study investigates the improvement of problematic soils through a controlled five-factor experimental program incorporating fine-content category (F), lime content (L), nano-zeolite content (NZ), curing temperature (T), and curing time (CT). Soil resistance (q) increased markedly from 193 to 7027 kPa, with the mean value rising from 225 kPa for untreated soil to 3402 kPa when the total lime–nano-zeolite dosage reached 20%. Higher curing temperatures and longer curing periods significantly enhanced strength development, increasing mean resistance from 1538 kPa at 20°C to 3082 kPa at 40°C and from 1354 kPa after 7 days to 2856 kPa after 90 days. Optimal performance was achieved with 10–15% lime and 8–12% nano-zeolite, while excessive nano-zeolite replacement reduced efficiency. To interpret these nonlinear interactions, six machine learning regression models were evaluated. The Gradient Boosting model achieved the highest predictive accuracy (R² ≈ 0.95, MAE ≈ 253 kPa, RMSE ≈ 379 kPa). Feature importance analysis identified lime content and curing temperature as the dominant variables, whereas fine-content category had minimal influence. Overall, the integration of systematic experimental analysis with machine learning provides a robust and interpretable framework for capturing nonlinear interactions and optimising stabilised soil performance in geotechnical engineering applications.