Training a CNN network with powerful but simple models
Ciro Messina, Luigi Bianco, Maurizio Fedi · 2025
This study wants to show how a Convolutional Neural Network may be trained by models built on a simple but strong a priori information—in this case, the gravitational field of a fault—can allow a good reconstruction of complex 3D structures. The key innovation is to train the algorithm with elementary source models. These elementary blocks consist of fault models with varying parameters such as dip, density contrast, thickness, and depth to the top. For each anomaly, profiles are extracted from the anomaly map, subdivided into two sub-profiles, and interpreted using the fault-based ML algorithm. This workflow follows the idea that gravimetric anomalies, when analysed along a profile crossing the source, can be seen as composed by the constructive interference of anomalies generated by the edges of the source bodies reducible to faults. The interpreted sections are then interpolated to create a reference 3D model, which yields a strong information, as a reference model, for a final 3D inversion process, which refines the model and yields a good data-misfit. To validate the method, we applied it to two different cases: a synthetic diapir-shaped source and a real geological structure, the Caltanissetta basin in Sicily (Italy). In both cases, the method successfully reconstructed the different structures.