Probabilistic prediction of 2D geotechnical parameter distributions considering sparse logging data and geophysical imaging

We illustrate a new work flow for probabilistic prediction of engineering target parameters in 2D or 3D. Traditionally, these target parameters are measured in laterally sparse boreholes or by means of direct push technology. We use Artificial Neural Networks (ANN) to find the optimal prediction model between the geophysical tomograms and the logging data of the target parameter. During the training phase of ANNs we consider the uncertainty of logging data and geophysical tomographic ambiguity to avoid data overfitting by ANNs. This can greatly improve the prediction results. We exemplary illustrate this approach employing cross-borehole tomographic data acquired at a field site South of Berlin, Germany, and link it to tip resistance logging data emanating from cone penetration tests. We achieve a probabilistic 2D tip resistance model which could be used for geotechnical site characterization and risk assessment.