Abstract: The liquidity index is a critical parameter for studying soil stability, deformation, strength, and related issues. Accurate prediction of the liquidity index is crucial. In this study, we evaluated the liquidity index using support vector regression (SVR), particle swarm optimization-based SVR (PSO-SVR), genetic algorithm-based SVR (GA-SVR), and simulated annealing-based SVR (SA-SVR) algorithms. The assessment utilized the piezocone penetration test (CPTU) dataset from Nanjing and Hefei regions, with the liquidity index derived from liquid limit and plastic limit tests as a reference. Predicted results were compared with laboratory tests and the CPTU empirical formula. Single-hole prediction analysis was conducted to align with engineering practice, and sensitivity analysis explored input parameter effects. Results showed that both the SVR model and optimized SVR models effectively predicted the liquidity index of cohesive soil, with the optimized models outperforming the original. Among these, the SA-SVR model excelled in wave smoothing and moderate peak values, enhancing prediction accuracy. For improved engineering predictions, normalized parameters (cone tip resistance, frictional resistance, pore pressure parameter ratio), overburden stress and effective overburden stress should be used as input variables. Sensitivity trends of PSO-SVR, GA-SVR, and SA-SVR models aligned with theoretical expectations, with SA-SVR exhibiting narrower span and greater consistency, confirming its accuracy. Thus, the proposed SA-SVR model offers superior prediction of cohesive soil liquidity index and guidance for engineering practice.

Key words: piezocone penetration test, liquidity index, support vector regression, particle swarm optimization, genetic algorithm, simulated annealing