-
oa Adaptive Learning Inversion with Gaussian Uncertainty
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, 2nd EAGE Workshop on Quantifying Uncertainty in Depth Imaging, Nov 2023, Volume 2023, p.1 - 3
Abstract
An adaptive learning inversion scheme was developed for geophysical data based on Gaussian Process and Active Learning, which goal is the reduction of the pool of data for training. The adaptive, physics-driven learning process uses evidence from field data as generated by an inversion process to provide real-world proxies to the machine learning training and ensure optimal generalization for field data applications. The dynamic learning model is obtained by an iterative feedback loop between the inversion process and an AI system progressively adapting to the characteristics of the field data. Convergence metrics is developed to monitor the flattening of the learning curve while ensuring the convergence of the data misfit from the physics-based procedure. The developed workflow enhances the generalization properties of machine learning for field data applications while ensuring only a small and statistically selected dataset are used for the task. The developed approach is tested on helicopter-borne transient EM and on seismic Full Waveform Inversion.