2D Site Characterization by Mixture of Gaussian Processes

Gaussian process regression is an effective method for the stochastic interpolation of geotechnical site data. However, a significant drawback of this method is its stationarity assumption, which is unrealistic given the existence of different soil layers. This assumption results in higher uncertainty in interpolation and poorer performance. To address this limitation, a mixture of Gaussian processes model is investigated for simultaneous layer identification and spatial interpolation in 2D. The model is based on the probabilistic assessment of layer boundaries and Gaussian processes for defining the statistical properties of the layers as well as spatial interpolation. The accuracy of the model is tested with real CPT profiles. The performance of the model is evaluated based on interpolation accuracy.

Keywords: CPT, Gaussian process, Mixture of Gaussian processes, Site characterization.