LAUSR

Abstract As for a typical machine-learning method, large volume of labeled examples is required to train and update neural parameters that are later used on new datasets. Well logs usually provide true values of elastic rock properties and enough amount of their correct answers may not often be available, due to budget limit and consumed time. In order to solve this problem, an unsupervised machine-learning approach is proposed to estimate rock properties based on pre-stack seismic data. The pre-designed Convolutional Neural Networks is capable to transform input gathers to relative rock properties, which are going to be added with a low-frequency component. Simulation of wavefields is then incorporated in this novel system, and neural weights and biases are altered to minimize a data-to-data loss function. The synthetic model of Book Cliffs is used for an application of this new method, in which both subsurface structures and property values are recovered very well. The proposed approach is further applied to a real dataset for the geothermal exploration in Denmark. Attributed to a relatively large distance between the well location and seismic line, inverted results are suboptimal.