Застосування технології збільшення роздільної здатності мігрованих сейсмічних зображень для 2D- і 3D-зйомок

• Datum: 2025
• Hauptverfasser: Okhariev, Viacheslav, Noskov, Oleksiy
• Format: Artikel
• Sprache: English
• Veröffentlicht: V.M. Glushkov Institute of Cybernetics of NAS of Ukraine 2025
• Schlagworte: модель машинного навчання; нейронна мережа; сейсміка; U-net архітектура; функція втрат
• Online Zugang: https://jais.net.ua/index.php/files/article/view/280
• Назва журналу: Problems of Control and Informatics

The quality of geological exploration works is directly proportional to the quality of seismic data processing, in particular the migration procedures that will be used for modeling and forecasting the distribution of reservoirs. To increase the quality of seismic interpretation workflow, a mathematical model of machine learning based on the neural network of the U-net architecture was developed and software implemented to increase the resolution and increase the signal / noise ratio for 2D and 3D seismic survey fields. An algorithm was built for the preparation of migrated seismic data in the standard SEGY format for processing with the help of a model and reverse conversion into the input format, which allows working with data of any geometry and physical size of the input file. To work with a three-dimensional dataset, an algorithm for transforming a set of 2D images into pseudo 3D volumes was described, as well as sufficiency criteria for training datasets for 2D and 3D variants of the of neural networks structure. In a simplified view, the quantitative characteristics of the input data for training the network in both formats are explained and the policies, in particular the mixed-precision mechanism, involved to reduce the load on the graphics card performing the model training are explained. The need to transition from 2D to 3D layout of the artificial intelligence network is substantiated, a number of forced measures to save PC resources for working out the given task are presented. The optimal loss function and layout of neural network blocks were selected for both 2D and 3D implementation options. As a result, neural networks were obtained whose loss function values are identical, i.e. the task is qualitatively performed for all proposed seismic data formats.