With the growing demand for reservoir characterization, accurate prediction of petrophysical properties, such as clay content and porosity, attracts increasing attention. For this problem, Bayesian rock-physics inversion (BRI) methods demonstrate success in theoretical and practical applications. The methods provide optimal estimations of petrophysical properties by effectively integrating the constraints of physical models and data-driven prior distributions. In recent years, deep neural network (DNN) models, recognized for their strong ability to extract complex features and map nonlinear relationships, show significant potential for petrophysical property predictions. However, in practice, the scarcity of petrophysical property data (typically well-log data) severely limits the size of the training data set, leading to the weak generalization of DNN models. In other words, although DNN models can be effectively trained to produce results that closely align with training well-log data, their predictions may still be unreliable when the trained DNN models are applied to regions distant from the training wells. To address this issue, we develop an unsupervised DNN model that leverages the strengths of deep learning to enhance the performance of BRI methods. Importantly, the new method operates under the same conditions as the original BRI without requiring additional training sample data. The application case demonstrates that our method produces more refined and clearer inversion results compared with the original BRI method. Furthermore, when petrophysical property data are available at the wells, these data can be incorporated to further improve inversion accuracy.

You do not have access to this content, please speak to your institutional administrator if you feel you should have access.