Deep learning-based inversion framework for fractured media characterization by assimilating hydraulic tomography and thermal tracer tomography data: Numerical and field study

Accurate characterization of fractured media is fundamental in the geological and geotechnical engineering applications such as coal mine production, deep geological disposal and enhanced geothermal systems (EGS). However, traditional inversion strategies are limited in their ability to characterize high-dimensional and non-Gaussian fractured media. Furthermore, a significant amount of observation well was employed during the inversion process in the previous studies. In this work, we proposed a joint inversion framework based on deep learning technique to overcome the limitations of the traditional strategies and the challenge of excessive use of observation wells. The convolutional variational autoencoder (CVAE) network was trained to parameterize the fractured media. After that, the ensemble smoother with multiple data assimilation (ESMDA) combined with the CVAE to characterize fractured media assimilating the hydraulic tomography (HT) and thermal tracer tomography (TT) data. A numerical study using four observation points validates the framework's reliability. The characterization errors for single-data cases are 16.9 % (HT) and 18.1 % (TT), decreasing to 16.7 % when both types of data are incorporated, demonstrating the synergies of multisource data. Sequentially, the framework is extended to the real-world scenario. The results show that our framework can effectively characterize the fractured media, capturing more features while addressing the challenge posed by excessive use of observation wells through the integration of multisource data. Our framework provides valuable insights into the characterization of fractured media in the practical engineering applications and highlights the benefits of multisource data assimilation.