GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 (Grombacher et al., 2021), and groundwater vulnerability assessment (Sandersen et al., 2021). When carrying out large-scale exploraton tasks, it is important for equipment operators to view the imaging interpretaton results of the detecton data, which is conducive to their on-site judgment of geological characteristcs and instrument working status, so as to make high-quality decisions. In recent years, a large number of studies have demonstrated the feasibility of deep learning algorithms in real-tme imaging of TEM data (Colombo et al., 2021; Chen et al., 2022). These studies established a specifc mapping between the exploraton data and the resistvity parameter space based on deep neural network (DNN) frameworks. However, due to the ill-posedness and mult-soluton nature of the geophysical electromagnetc inverse problem, the same exploraton data can have multple or infnite diferent geological model solutons. This one-to-many mapping relatonship brings great training difculty to the deep learning network and also afects the reliability of applying deep learning networks to interpret TEM data. In this context, the development of probabilistc neural networks (PNN) has provided an efectve soluton to the nonlinear inversion problems in geophysics. The currently typical PNN structures include mixture density networks (MDN) and invertble neural networks (INN). MDN can learn to map a vector to an n-dimensional conditonal probability distributon and parameterize it as a Gaussian mixture model (GMM) to learn arbitrary probability distributons (Mosher et al., 2021). INN can learn the bidirectonal mapping between inputs and outputs, and it can estmate the posterior probability density functon (PDF) by introducing additonal latent variables on the output side. Both of these network structures can efectvely simulate Bayesian posterior inference and have been successfully applied in geophysical inversion methods. Inspired by the outstanding research mentoned above, we propose a composite probabilistc neural network (cPNN) structure that incorporates the LSTM autoencoder network with both DNN and MDN network structures. This design allows for simultaneous deterministc imaging and probabilistc estmaton of tTEM data. Furthermore, we are able to evaluate the depth of investgaton (DOI) through the resistvity Gaussian distributon output by the cPNN network. II Bayesian imaging framework As shown in Fig. 1, the Bayesian imaging framework based on the cPNN mainly includes three stages: data generaton, network constructon and training, and imaging result output. In the data generaton stage, input data and label data for the entre network structure need to be prepared, including TEM response data and the corresponding theoretcal resistvity model. In this study, considering the computatonal complexity caused by high-dimensional layer models and the superior shallow subsurface detecton resoluton of ground-based TEM compared to ATEM, the number of model layers is set to 30. We generated 30 depth interfaces within a range of 120 m underground using a log increasing with depth method. The last layer is assumed to be a semi- infnite half-space. In the stage of imaging result output, the diference between MDN and conventonal neural networks is that MDN outputs a conditonal probability distributon, and it can learn arbitrary