GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.1 GNGTS 2023 NeuroPick: Efficient LWD Image logs segmentation using Deep Learning A. Molossi, G. Roncoroni, M. Pipan Department of Mathematics and Geosciences, University of Trieste, via Weiss 2, Trieste, Italy. Introduction The need for greater efficiency to support the sustainability of drilling operations, whether related to the attainment and exploitation of fossil or renewable resources, is also a function of reducing the time taken to acquire crucial information for decision-making processes. An example of such decision-making steps in the drilling field concerns precision geosteering operations, where the well takes specific preferential directions other than vertical that optimise access to the area of interest (Chen et al., 2012). The deviation of wells, however, affects the risk of encountering borehole stability problems, increasing it by virtue of the possible interference between well direction and azimuth and dip directions of fault or fracture planes. Modern wellbore Logging While Drilling (LWD) technologies offer the ability to capture 360° images of the wellbore during drilling. The interpretation of this type of petrophysical data is functional for the recognition of surfaces and the geometric and structural reconstruction of the subsurface portions drilled, resulting in information to support the above-mentioned decision-making processes, as well as providing the basis for the calibration of seismic data also in view of 4D applications. The time required for a human operator to interpret the data is, however, relatively long and depends on the complexity of the drilled area. In any case, such times, in addition to the inherent subjectivity of the interpretation process, can undermine the benefits and advantages of such data. In recent decades, thanks in part to the widespread increase in computational capabilities and resources, Deep Learning (DL) has attracted increasing attention within the geophysical community, with the goal of gathering useful information from raw geophysical data for terrestrial characterization, which as we know can have considerable social and economic impact. Given the impressive rate at which observed geophysical data are growing, the development of DL-based models applicable to Big Data is crucial and involves both academia and industry, preparing the discipiline for a cultural shift focused on interdisciplinary expertise (Chen et al., 2012; Yu & Ma, 2021).