EE Seminar: Deep Learning for Seismic Inversion
הרישום לסמינר יבוצע בתחילת הסמינר באמצעות סריקת הברקוד למודל (יש להיכנס לפני כן למודל, לא באמצעות האפליקציה)
Registration to the seminar is done at the beginning of the seminar by scanning the barcode for the Moodle (Please enter ahead to the Moodle, NOT by application)
Electrical Engineering Systems Seminar
Speaker: Maayan Gelboim
M.Sc. student under the supervision of Prof. Raja Giryes and Prof. Amir Adler
Wednesday, 22nd December 2025, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Deep Learning for Seismic Inversion
Abstract
Seismic inversion is a fundamental tool in exploration geophysics, involving the reconstruction of 2D and 3D subsurface earth models using seismic data, with applications in hydrocarbon exploration, CO2 sequestration, and shallow hazard assessment, among others. Seismic inversion is an ill-posed problem that typically requires Terabytes of recorded seismic data for high-quality imaging. Seismic inversion is commonly solved by full waveform inversion (FWI), which iteratively searches by gradient-descent an earth model that minimizes the misfit between observed and synthetic seismic data. However, FWI often converges to low-quality solutions due to its high sensitivity to initial conditions, and it also suffers from very high complexity leading to processing times on the order of one day per iteration for 3D inversion.
Our research addresses key questions regarding the feasibility of deep learning for seismic inversion: (I) can encoder-decoder architecture accelerate the reconstruction of complex 3D geological models, achieving significant time savings compared to FWI? (II) what is the impact of reducing the dimensionality of the seismic input data on the encoder-decoder inversion quality? (III) how robust is the encoder-decoder architecture to synthetic and field noise? (IV) how can compressed sensing and representation learning be utilized to identify the most important subsets of seismic data for deep-learning-based inversion? (V) are these tools for identifying the most important subsets of seismic data, effective also for accelerating FWI?
