Overpressure Prediction of Source Rocks By Physics-Informed Machine Learning

Significance 

Overpressures is one of the factors with a significant influence on the safety of oil/gas exploration, drilling and production. While pressure prediction in petroleum exploration is well articulated in the literature, most existing studies have concentrated on limited to moderately overpressure formations attributed to disequilibrium compaction. The Xihu depression, located in the East China sea basin, is a prime oil/gas production region. This depression has several vertically placed overpressure systems, one of the main indicators of the presence of natural gas in the region. Thus, overpressure prediction is of great significance to oil/gas exploration not only in the Xihu depression but in other oil fields as well.

Reliable pressure prediction requires a thorough knowledge of the origin and distribution of the overpressure, which is classified into five categories: fluid expansion, pressure transfer, tectonic compression, disequilibrium compaction and diagenesis. Fluid expansion caused by natural gas generated from the source rocks is the dominant overpressure mechanism in the Xihu region. Compared with conventional overpressures in uncompacted formations, the resulting overpressures caused by gas expansion manifest as thick mud-coated thin sand (MCS) high-pressure structure, which is still not fully addressed in prediction. Such complex MCS structures, gas-producing overpressures, high porosities and the influence of high gas content on seismic velocities render traditional prediction methods unsuitable.

Rock physics templates are widely used to predict overpressures. However, the main challenge to this approach is elastic nonlinearity. Since attenuation and acoustic velocity from rock physics exhibit uncertain dependence on formation pressures, it could be more useful if pressure predictions were based on nonlinear seismic inversions based on applying a combination of statistical and deterministic methods. To address these challenges, Dr. Yifan Cheng and Professor Li-Yun Fu from China University of Petroleum (East China) employed some petrophysical models into non-traditional Caianiello convolutional neural networks (CNNs) to combine deterministic and statistical overpressure prediction mechanisms. The work is currently published in the journal, Journal of Petroleum Science and Engineering.

In their approach, a frequency-dependent quality factor Q-pressure petrophysical model was employed owing to the high sensitivity of acoustic attenuation to overpressures. First, the physics-informed CCNNs were established for overpressure prediction, followed by the geological setting of the Xihu region based on the available information reading the origin and distribution of the overpressures. The well-based empirical petrophysical model was constructed by correlating pressures and well quality factors. The reliability of the seismic data was established and the model was validated. Finally, the overpressure of the Xihu depression was predicted using the proposed method.

The authors demonstrated that by relating overpressures to seismic properties, petrophysical model could improve the predictive ability of physics-informed CCNNs. Such convolution neurons allowed strong extraction of features as well as powerful learning ability desirable to enhance overpressure prediction for Xihu depression. The presented scheme successfully validated the applicability of the physics-informed CCNNs inversion scheme. The reliability of the prediction away from the training wells was dependent on the geological complexity of the areas being studied. However, it was possible to improve the representation information in the trained CCNNs by feeding new well data gradually during oil field development.

In summary, Professor Li-Yun Fu and Dr. Yifan Cheng investigated the origin and distribution of overpressures in the Xihu depression to improve the reliability and accuracy of overpressure prediction. The presented CCNNs-based seismic inversion scheme comprised deep learning for neural wavelets, petrophysical modeling reservoir, deconvolution-based inversion, well-seismic correlation analysis and input-signal reconstruction to enhance the reliability of the initial pressure model. In a statement to Advances in Engineering, Professor Li-Yun Fu pointed out that the study would pave the way for improved overpressure prediction in oil/gas exploration and development.

Overpressure Prediction of Source Rocks By Physics-Informed Machine Learning - Advances in Engineering
Flow diagram of CCNNs-based joint inversion scheme by integrating seismic data and logging data to estimate seismic impedance (Fu, 2004) using seismic convolutional model (Robinson, 1957), porosity/clay-content (Fu, 2002) using traditional petrophysical models (Raymer et al., 1980; Han et al., 1986), and formation pressure in this study using the frequency-dependent Q – pressure petrophysical model as the activation function of Caianiello neurons, respectively. First, reservoir petrophysical modeling is conducted for the physics-informed activation function of Caianiello neurons. Nest, neural wavelet (NW) extraction and nonlinear factor (NF) optimization are implemented by integrating logging and seismic data at wells. Then, deconvolution-based inversion away from wells is performed for initial model estimation. Finally, input-signal reconstruction is used to improve the initial model.

About the author

Li-Yun Fu received the B.S. degree in geophysics from the Chengdu College of Geology, Chengdu, China, in 1985, and the M.S. and Ph.D. degrees in geophysics from the China University of Petroleum, Beijing, China, in 1992 and 1995, respectively.

He commenced his research career with the China Offshore Oil Exploration & Development Research Center, CNOOC, China. From 1995 to 1997, he was a Post-Doctoral Fellow of Engineering Mechanics with Tsinghua University, Beijing. In 1997, he joined the Institute of Tectonics, University of California at Santa Cruz, CA, USA, as a Researcher. In 1999, he joined Australia CSIRO, Perth, WA, Australia, as a Scientist Staff. In 2004, he joined the Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing, and headed the Seismology Group. In 2017, he joined the China University of Petroleum (East China), Qingdao, China, where he is currently a Professor of Geophysics, with the webpage link: https://orcid.org/0000-0001-8692-8405 and http://geori.upc.edu.cn/2018/0526/c10380a151967/page.htm.

He pioneered Caianello convolutional neural networks (CCNNs) since his doctoral dissertation (1995) by recovering a visual nervous model (Caianello, Journal of Theoretical Biology, 1961, 2, 204–235). He proposed physics-informed CCNNs by taking some deterministic physical models as the activation function of Caianiello convolutional neurons. Both the CCNNs-based feature extraction from training data and the physics-informed activation function for guided learning result in various powerful algorithms that provide fast solvers for geophysical applications. His current researches focus on the CNN learning architecture based on dual-domain Fourier integral operators, which shuttles information flow in the space and wavenumber domains for a machine learned partial differential equation.

He has authored or coauthored more than 300 articles in academic journals. His research interests include wave propagation, Caianello convolutional neural networks, physics-informed CCNNs, dual-domain Fourier neural operators, complex structure imaging, nonlinear inverse problem, and high-temperature/pressure rock physics.

Reference

Cheng, Y., & Fu, L. (2022). Nonlinear seismic inversion by physics-informed Caianiello convolutional neural networks for overpressure prediction of source rocks in the offshore Xihu Depression, East China. Journal of Petroleum Science and Engineering, 215, 110654.

Go To Journal of Petroleum Science and Engineering

Check Also

Advancing Fusion Energy: High-Field REBCO Superconducting Magnets in the SPARC TFMC Program - Advances in Engineering

Advancing Fusion Energy: High-Field REBCO Superconducting Magnets in the SPARC TFMC Program