Effects of soil layer discretisation on embankment settlement using Bayesian back analysis

Bayesian updating is a machine learning tool that has been used to assess the performance of an embankment constructed on soft ground, which incorporates measured data to evaluate key geotechnical parameters. The method is computationally expensive and hence reducing run times while retaining accuracy in the modelling is a desirable goal. A model utilising four soil layers with 13 random variables is compared with a model utilising nine layers having 37 random variables. An initial case is run using only surface settlement plate data to predict surface settlement followed by a second case incorporating both downhole magnetic extensometer data and vibrating wire piezometer data for comparison. The run time for the analyses did not reduce when a four-layer model was used. Both the four- and nine-layer models can accurately match the monitoring data but unrealistic posterior parameters are generated when a four-layer model is used with only surface settlement data. More realistic posterior parameter sets are obtained from the nine-layer model with surface settlement data and by both four- and nine-layer models when using magnetic extensometer data.