Probabilistic Inversion of Aeroelectromagnetic Data

The interpretation of AEM data with a model of the earths spatial distribution of electrical conductivity whose theoretical response is as close as possible to the observed data is an efficient tool for geological mapping, geotechnical questions, groundwater and mineral exploration, and environmental mapping. Due to modern AEM surveys may well contain tens of thousands of line kilometers of multichannel data sampling a full 2-D or 3-D solution to the AEM inverse problem is currently beyond the capabilities of all but the most powerful computers. Therefore frequency-domain AEM data have often been interpreted by 1-D inversion procedures based on the layered halfspace model. This has been proven to be a fast and approximative tool to locate anomalies in the subsurface even in the presence of lateral changes in resistivity. In this work we will use a probabilistic description of homogeneous halfspace models and horizontally layered homogeneous halfspace models with 2 or 3 layers. The mapping from the measured AEM data to a horizontally layered halfspace model is not unique. If one of the possible models is selected by applying extra constraints, the information about all other models which also could explain the data is lost. This can often lead to interpretation results which do not correctly approximate the real structure of subsurface at the measuring position. We will describe the set of models which can explain the data by a probability distribution of the model parameters. The modeling of the probability densities will be performed with Mixture Density Networks. This procedure will ensure that all available information about the relation between measured data and model parameters is maintained until the process of decision making. There it will be combined with extra knowledge, which can vary for different geological settings, and thus lead to different final model parameters for the same measured data vector at different locations. This will lead to a more realistic and more accurate interepretation of AEM data.