Data assimilation for geotechnics – exploring the possibilities

Paper in proceeding, 2022

With the increasing amount of data being gathered in geotechnics, new methods that quantify and reduce the uncertainties of the numerical forecasting models become available. In geotechnics several researchers have already started to employ Data Assimilation (DA) combining forecasting and monitoring data for parameter inference and state estimation, however, a systematic survey of DA methods in terms of numerical stability and computational efficiency is still lacking. Furthermore, in most geotechnical problems only a limited number of independent physical quantities is measured whilst many model parameters need to be assessed. This high order dimensionality makes parameter inference using augmented system state approach (combining measured state + parameters) a challenging task. In this paper we present a comparison on the potential of different DA methods in geotechnics. A classical problem of soil water coupled one-dimensional settlement of an embankment on soft soil is used as a benchmark demonstration. A general framework incorporating different DA algorithms, namely, the Unscented Kalman Filter, Ensemble Kalman Filter, Ensemble Square Root Filter and Particle filter will be presented and demonstrated for an example case. The initial results show that the Unscented Kalman Filter shows better convergence of parameters for simple elasto-plastic models, however the Ensemble Kalman Filter with perturbed observation performs substantially better compared to its square root variant and all other filters, even for complex elasto-viscoplastic models. Particle Filter is less suitable for geotechnical applications due to the curse of dimensionality although it’s accuracy can be increased with an increased sample size or using an appropriate proposal distribution to sample the posterior, but this becomes computationally demanding. This study also shows that the model parameter sensitivity is a function of observation strategy and regardless of the Data Assimilation methodology used, its efficiency depends on an optimally designed monitoring network i.e., optimal sensor location and measurement interval.