GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 An innovatve machine learning algorithm for gravity modelling C. Messina¹, L. Bianco¹, M. Fedi¹, ¹ Department of Earth, Environment and Resources Sciences, University of Naples “Federico II”, Naples, Italy We describe a Machine Learning algorithm for interpretaton of gravity data generated by rather complex structures. We choose a Convolutonal Neural Network (CNN) with a U-Net architecture. This architectural design of the network has been recently applied in gravity modelling scenarios, in which the training dataset was built introducing strong prior informaton about the source without obtaining a generalized training set. To overcome this limit we train the network through examples composed by labels consttuted by simple elements, here called building blocks, with features being their corresponding gravimetric anomalies. Next to the training we test our method frst analysing gravity anomalies produced by simple structures (e.g., prisms, horizontal cylinders), and then with those generated by increasingly complex sources with irregular shapes, such as salt diapirs. We show examples of 2D-3D of real cases. We assume that a gravimetric anomaly can be seen as composed of the constructve interference of anomalies generated by the edges of the source associated to building blocks. Moreover, this method streamlines decision-making and reduces computatonal eforts involved in assembling a suitable dataset. Corresponding author: ciromessina631@yahoo.it