A Rock Physics Inversion Method Based on Physics-guided Autoencoder Network

Rock physics inversion is a technique used to determine the physical properties of reservoir rocks. Typically, inversion methodologies are divided into two primary groups: one that depends on data-driven statistical or machine learning approaches, which often suffer from a lack of interpretability, and another that utilizes rock physics models to align rock properties with the observed elastic responses but it requires constraints from low-frequency models. We propose a physics-guided autoencoder network approach to solve the rock physics inversion problem by embedding rock physics models that conform to geological characteristics into the neural network. The network's output serves as the input to the rock physics model and is used to generate synthetic elastic parameters. The loss function calculates the difference between these synthetic and observed elastic parameters for gradient computation and weight updates. The training process of the model does not require labeled data and is entirely guided by rock physics theory. Through comparison with a purely data-driven neural network model for a dataset from a tight sandstone reservoir, we establish the efficacy of our approach in developing an accurate, interpretable, and generalizable framework for rock physics inversion. Moreover, the method is independent of the choice of an initial model, which allows for a more flexible rock physics inversion approach.