Bayesian Model Calibration Using Geotechnical Centrifuge Tests

The predicted performance using a geotechnical prediction model is expected to deviate from reality. A practical approach to assess the model error is through calibration with observed performances in physical model tests. In this paper, a Bayesian framework of model calibration using centrifuge modeling tests is proposed and the procedure of model calibration is illustrated. Two centrifuge tests conducted to investigate the performance of soil slopes under rainfall conditions are used to calibrate a coupled hydromechanical analysis model. It is found that for centrifuge tests with different levels of soil variability, the test with a smaller variability of soil properties is more efficient for model calibration. According to the concept of random field, a centrifuge model with a larger model size and accelerated to a lower acceleration is better for model calibration. When the discrepancy between the performance interpreted from the centrifuge model and the field performance is small, the improvement of the reliability estimation for a new slope is significant. However, when there is little information about the discrepancy, the reliability estimation cannot be significantly improved by the information from centrifuge modeling. The proposed procedure is shown to be able to quantify the calibration effects of centrifuge tests and may be used to achieve a more reliable calibration.