Partial automation of the seismic to well tie with deep learning and Bayesian optimization

Tying the seismic to the well logs is an important operation in the processing and interpretation workflow. It is performed multiple times during the exploration phase to understand the wavelet and the time-depth relation at the well position. It is an iterative process where the geoscientist needs to adjust a number of key parameters in order to converge towards an accepted solution. Potential noise in the data as well as various sources of uncertainties can complicate the task and lead to a time consuming and sometimes frustrating experience for the interpreter. In this work, we automate the overall workflow using deep learning and Bayesian search. First, we build a variational convolutional neural network and train it to solve the wavelet extraction problem. The network learns the deconvolution process from synthetic pairs of reflectivity and seismic traces. Once trained, it can robustly estimate wavelets given input series of arbitrary lengths. The variational nature of the network allows to quantify the uncertainties of the process, in particular for the phase of the wavelet. Secondly, we resort to a global Bayesian optimizer to automatically tune the parameters. The search space is composed of a number of key parameters, such as the time-depth table bulk shift, and the optimizer iteratively tries new combinations in order to maximize the quality of the tie. The user controls the bounds of the space, making the overall algorithm more interpretable. We also extend the method to perform prestack well-ties and explore the possibilities for joint ties. We validate our approach on two challenging real datasets and show that it yields accurate results using reasonable computing resources.