GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 3.1 GNGTS 2023 Discussion In this work, we have demonstrated the applicability of DL-based methods for the optimisation of the LWD well image interpretation process to support and streamline decision-making steps and reducing the operational risk associated with drilling for fossil or renewable natural resource exploitation (hydrocarbons, geothermal reservoir), injection as in the case of modern CCS applications, or applications in interplanetary missions, e.g., on rover-mounted LWD acquisition systems. Qualitative considerations that naturally emerge from the comparison of the results of the two strategies discussed, CL and single-step, draw attention to the fact that in the single-step case the pixels activated by the model during prediction are fewer in number and thus allow for better recognition of the sharper contrasts visible in the original image. Similarly, however, the model from the CL strategy activates more pixels even where there are no sharp contrasts in the original data. This allows us to immediately consider a complementarity of the results from the two strategies. Later on, in fact, this complementarity can be exploited for the recognition of surfaces that are more difficult to interpret in a poor-quality data as in this case, such as faults or fractures, which do not necessarily correspond to sharp contrasts in the image, but may coincide with rather hidden sinusoids within generally homogeneous thicknesses. Future developments of the method concern the implementation of a further network that estimates the geometric parameters of the discontinuities from the results of the segmentation carried out by the U-Net. The basic idea is to add a regression step for the estimation of the planar approximation of the geological discontinuity, at the end of the classification task corresponding to the segmentation, in order to obtain a true geometric model of the subsurface as the ultimate result of the entire workflow of automatic interpretation of this LWD data. Imagining the real-life application of such a method, it is intuitive that the added value created by the model concerns the interpretation of the perforated formations in real time, and this can considerably reduce the time related to the same task performed by a human alone. The latter may have an advantage in interpretation time if, supported by a model whose segmentation results are qualitatively acceptable as in Fig.3, his task is limited to validating or discarding the surfaces found automatically. Not least, this speeding up and input of objectivity into the process can also benefit uncertainties and, therefore, the sustainability of drilling. References Bengio, Y., Louradour, J., Collobert, R., & Weston, J. ; 2009: Curriculum learning. Proceedings of the 26th Annual International Conference on Machine Learning - ICML ’09, 1–8. https://doi.org/10.1145/1553374.1553380 Chen, H., Chiang, R. H. L., & Storey, V. C.;2012: Business Intelligence and Analytics: From Big Data to Big Impact. MIS Quarterly, 36(4), 1165–1188. https://doi.org/10.2307/41703503 Ronneberger, O., Fischer, P., & Brox, T. ;2015: U-Net: Convolutional Networks for Biomedical Image Segmentation (arXiv:1505.04597). arXiv. http://arxiv.org/abs/1505.04597