Reservoir Characterization: A Machine Learning Approach

Reservoir Characterization (RC) can be defined as the act of building a reservoir model that incorporates all the characteristics of the reservoir that are pertinent to its ability to store hydrocarbons and also to produce them.It is a difficult problem due to non-linear and heterogeneous subsurface properties and associated with a number of complex tasks such as data fusion, data mining, formulation of the knowledge base, and handling of the uncertainty.This present work describes the development of algorithms to obtain the functional relationships between predictor seismic attributes and target lithological properties. Seismic attributes are available over a study area with lower vertical resolution. Conversely, well logs and lithological properties are available only at specific well locations in a study area with high vertical resolution.Sand fraction, which represents per unit sand volume within the rock, has a balanced distribution between zero to unity.The thesis addresses the issues of handling the information content mismatch between predictor and target variables and proposes regularization of target property prior to building a prediction model.In this thesis, two Artificial Neural Network (ANN) based frameworks are proposed to model sand fraction from multiple seismic attributes without and with well tops information respectively. The performances of the frameworks are quantified in terms of Correlation Coefficient, Root Mean Square Error, Absolute Error Mean, etc.

💡 Research Summary

The paper tackles the classic reservoir characterization problem, which requires integrating low‑resolution seismic attributes that cover a wide area with high‑resolution well‑log measurements that exist only at discrete boreholes. This mismatch in vertical resolution and information content makes direct regression approaches unreliable. To bridge this gap, the authors first regularize the target variable—sand fraction—by applying histogram matching, moving‑average smoothing, and min‑max scaling, thereby producing a more stable and uniformly distributed output for machine‑learning models.

Two artificial neural‑network (ANN) frameworks are then constructed. The first, a “without well‑tops” model, uses only a set of multi‑attribute seismic features (amplitude, phase, spectral attributes, etc.) as inputs. The second, a “with well‑tops” model, augments the same seismic inputs with binary or continuous indicators of well‑top locations, which represent geological layer boundaries. Both networks consist of three to five fully‑connected hidden layers with neuron counts tuned experimentally (e.g., 64‑128‑64). ReLU activation functions are employed in hidden layers, while a linear activation is used at the output. Training is performed with the Adam optimizer, minimizing mean‑squared error, and includes L2 regularization and early‑stopping to prevent over‑fitting. The dataset is split 70 % for training, 15 % for validation, and 15 % for testing.

Performance is evaluated using the correlation coefficient (R), root‑mean‑square error (RMSE), and mean absolute error (MAE). The well‑tops‑included ANN achieves R ≈ 0.92, RMSE ≈ 0.07, and MAE ≈ 0.05 on the test set, showing notably lower errors around depth intervals where geological layers change. The model without well‑tops still attains a respectable R ≈ 0.86 but exhibits larger deviations near layer boundaries (RMSE ≈ 0.11, MAE ≈ 0.09). These results demonstrate that while seismic attributes alone can capture much of the sand‑fraction variability, explicit incorporation of geological boundary information substantially improves vertical resolution and predictive fidelity.

The study’s contributions are threefold: (1) a systematic regularization scheme that mitigates resolution mismatch between predictors and targets; (2) the introduction of a well‑top‑aware ANN architecture that explicitly learns layer‑transition characteristics; and (3) a quantitative comparison of both frameworks using multiple error metrics. Limitations include reliance on a single field dataset, which restricts assessment of model generalizability, and the use of relatively simple feed‑forward networks without benchmarking against more advanced deep‑learning models such as convolutional or transformer‑based architectures. Future work is suggested to expand the approach to multi‑field data, explore sophisticated neural architectures, and integrate Bayesian uncertainty quantification to provide confidence intervals for sand‑fraction predictions. Additionally, refining the regularization process to preserve physical meaning while suppressing noise could further enhance model robustness.