Porosity is one of the most critical parameters for subsurface characterization, and is of vital interest to understand fluid flow properties (Bernabé, 1995;Carman, 1956), elastic and mechanics behaviors of rocks (McBeck et al., 2019), and pore pressure evolution and prediction (Obradors-Prats et al., 2017). Therefore, predicting porosity from seismic data has significant influences in geo-energy (oil, gas, geothermal, etc.) reservoir characterization and development (

Seismic characterisation of deep carbonate reservoirs is of considerable interest for reservoir distribution prediction, reservoir quality evaluation and reservoir structure delineation. However, it is challenging to use the traditional methodology to predict a deep-buried carbonate reservoir because of the highly nonlinear mapping relationship between heterogeneous reservoir features and seismic responses. We propose a machine-learning-based method (random forest) with physical constraints to enhance deep carbonate reservoir prediction performance from multi-seismic attributes. We demonstrate the effectiveness of this method on a real data application in the deep carbonate reservoir of Tarim Basin, Western China. We first perform feature selection on multi-seismic attributes, then four kinds of physical constraint (continuity, boundary, spatial and category constraint) transferred from domain knowledge are imposed on the process of model building. Using the physical constraints, the F1 score of reservoir quality and reservoir type can be significantly improved and the combination of the effective physical constraints gives the best prediction of performance. We also apply the proposed strategy on 2D seismic data to predict the spatial distribution of reservoir quality and type. The seismic prediction results provide a reasonable description of the strong heterogeneity of the reservoir, offering insights into sweet spot detection and reservoir development.

Lithology prediction from seismic data is of great significance for sweet-spot detection, reservoir structure delineation, geologic model building, and hence reducing the risk of exploration and development. Traditional lithofacies prediction methods often are limited by the seismic inversion accuracy and reliability of the rock-physics relationships, which are challenging to be applied in complex reservoirs (such as those containing coal-bearing strata or thin layers). Convolutional neural networks (CNNs) can represent the coupling relationship of seismic characteristics in the time domain through multilayer convolution and effectively manipulate multitype and multidimensional seismic data. Under the framework of a supervised CNN, we jointly integrate prestack seismic gathers (Pre-SGs), seismic inversion results (P-impedance and VP/ VS ratio), multiseismic attributes (amplitude-variation-with-offset [AVO] intercept, AVO gradient, instantaneous amplitude, instantaneous frequency, and instantaneous phase), and spectral decomposition attributes (SDA) to predict lithofacies in a complex clastic reservoir interbedded with thin-layer coal. We determine that the fusion model with multiseismic information containing different perspectives and complementary information of seismic data is capable of achieving better prediction performance than only using one type of input feature. In particular, using the proposed methodology, the angle-dependent Pre-SG is essential to decipher the rich information of lithologic details. The models using only poststack seismic data or inversion results cannot reliably describe lithologic details (especially the thin-coal layers). In addition, by including the SDA into model inputs, the model’s ability to recognize thin layers has been further improved but lead to the slight sacrifice of overall prediction accuracy.

Seismic fluids prediction under the machine-learning framework is of great significance for the exploration and development of oil and gas resources, geothermal energy exploitation, carbon dioxide sequestration monitoring, and groundwater management. Data-driven supervised machine-learning algorithms often rely heavily on the characteristics of the data (number of labels and data distribution). The disparity in the number of different labels for the majority and minority samples would hinder the generalization ability of the machine-learning model, especially weakening the predicting power for minority groups (e.g., hydrocarbon-bearing rocks) which are often of essential interest to us. For a clastic reservoir exhibiting a typical class imbalance (the ratio of gas sandstone to other lithofluids is significantly low), under the framework of a supervised convolutional neural network, we investigate and compare various class-rebalance methods to enhance the model’s prediction ability for gas-bearing sandstones. To achieve the purpose of class rebalance, we mainly use sampling methods to obtain class-balanced data sets and cost-sensitive learning methods to modify loss functions. The crosswell blind tests indicate that the ensemble-based undersampling method of BalanceCascade is found to be most effective in enhancing the prediction performance, increasing the F1 score of gas sandstone by as much as 15%. We also propose the combination of BalanceCascade and focal-loss (FL) methods, which can further improve the F1 score of gas-bearing sandstone in several wells compared with using BalanceCascade or FL alone. By incorporating class-rebalance strategies into model building, we finally obtain more reliable seismic prediction results for gas-bearing sandstone.

No abstract