Free-Fall Penetrometer Data Interpretation Through Bayesian Inference and Gaussian Process Regression

The interpretation of free-fall penetrometer (FFP) data for offshore site investigations is associated with substantial uncertainty, arising from factors such as the inherent variability of subsurface properties, measurement error, and the use of simplified empirical models for FFP analysis. In this study, Bayesian inference, combined with Markov Chain Monte Carlo sampling, is employed to optimize a commonly used formulation for interpreting FFP data, based on paired FFP and constant-rate penetrometer laboratory test data. Additionally, Gaussian Process regression (GPR) is utilized to develop a predictive model for quasi-static tip resistance (q_s). Results show that the uncertainty in model parameters, particularly the strain-rate coefficient, can significantly influence the estimated range for quasi-static tip resistance. Bayesian inference not only significantly improves the mean prediction of q_s, with R2increasing from 0.21 to 0.75, but also provides uncertainty measure. When the GPR is applied, the prediction performance improves, reaching an R² of 0.95 for the training and 0.91 for the testing dataset.

Keywords: Free-fall penetrometer, Bayesian inference, Gaussian Process regression, Offshore site investigation.