Global sensitivity analysis and Bayesian calibration of an advanced cyclic constitutive model for sands via surrogate modelling

Purpose This paper aims to propose a probabilistic calibration of the SANISAND-MS model under cyclic loading conditions by integrating polynomial chaos expansion (PCE) metamodels with Bayesian inference. The primary objective is to demonstrate the effectiveness of this approach compared to conventional calibration processes and to highlight the benefits of obtaining posterior probability distributions. Design/methodology/approach The paper presents a framework that combines PCE metamodels with Bayesian inference. This framework is initially applied to the monotonic component of an advanced constitutive model, followed by the cyclic component. Additionally, the study explores the interrelationships among various parameters. Findings In addition to demonstrating improved optimization of parameter calibration, understanding the interrelationships among these parameters clarifies various aspects of the selected constitutive model. The effectiveness and importance of this process are supported by comparisons with previously published calibration data. Research limitations/implications The exercise has been carried out on a single set of laboratory data to demonstrate the proposed methodology’s potential and value. Practical implications The practical implications of the proposed framework may include: improvement of constitutive modelling calibration workflow. The combination of PCE-based global sensitivity analysis and Bayesian parameter updating provides a reproducible, data-driven path to identify influential constitutive parameters, quantify uncertainty and obtain posterior parameter sets for advanced cyclic sand models. This supports objective, transferable constitutive model calibrations instead of trial-and-error tuning. Laboratory test planning. Sensitivity results help prioritize laboratory experimental programmes by revealing which parameters and response features drive predictive variance, so limited testing resources can target the most informative loading paths, densities and drainage conditions. Constitutive model selection and use. Posterior parameter distributions with correlations allow engineers to propagate uncertainty in boundary value analyses, improving credibility of predictions for components and systems that face cyclic loading. Performance-based assessment. Surrogate-enabled ensembles make large simulation sets feasible in practical time, supporting probabilistic checks at serviceability and ultimate states in design studies. Originality/value For the first time this type of ML approach is applied for the calibration of advanced constitutive modelling for cyclic loading. The findings can help future uncertainty propagation studies at system level, particularly towards offshore energy structures.