An Integrated Machine Learning–Enabled Workflow for Seismic-Driven Reservoir Characterization Using AVO Attributes, Extended Elastic Impedance, Colored Inversion, and Non-Stationary Property Modeling

Abstract

We applied machine learning to enhance reservoir characterization by predicting elastic and petrophysical properties from seismic AVO attributes. Using synthetic gathers and a representative wavelet, we trained models at well locations and applied them directly to intercept and gradient volumes, bypassing traditional seismic inversion. Auto-selection of optimal machine learning algorithms minimized prediction error. Final outputs included reservoir properties cubes, improving rock facies identification and reservoir understanding through Quantitative Interpretation (QI) workflows. Traditional neural network methods for reservoir characterization often lack deep learning capabilities, regression metric selection, and AVO integration. We applied a QI-based machine learning approach using FastTree and synthetic AVO angle gathers, supervised by petrophysical logs, to predict reservoir properties. The trained model was applied to seismic intercept and gradient cubes to generate acoustic impedance volumes. Advanced machine learning algorithms were then integrated with geostatistics to predict porosity and permeability with high fidelity and uncertainty quantification. This approach handles non-stationary data, reduces pre-conditioning, and enables automated stochastic modeling, including sweet spot identification, with minimal manual input. Machine learning applied to reservoir characterization using AVO attributes has shown strong potential. By training models on synthetic intercept and gradient gathers at well locations, petrophysical properties like acoustic impedance and porosity were predicted directly from seismic data, eliminating the need for seismic inversion. Among tested algorithms, FastTree, a gradient boosting method, stood out for its accuracy and robustness. It builds regression trees iteratively, each correcting the previous using a loss function. The QI ML function performed well with conditioned well logs and minimum-phase seismic data, trained within intervals rich in AVO signals and validated against petrophysical logs. Auto-selection of machine learning algorithms and tuning via regression metrics (R², RMSE, MAE, MSE) helped identify optimal models. Additionally, rotating seismic and wavelet inputs assuming near-zero phase improved synthetic generation and prediction fidelity. In ADNOC Offshore's pilot, advanced machine learning workflows trained well log datasets using supervised learning, auto-selecting optimal algorithms and metrics. These models were applied to seismic volumes to derive elastic, porosity, permeability cubes, and facies probability distributions. FastTree and neural networks showed superior accuracy and spatial coherence, especially in complex carbonate and fluvial-deltaic reservoirs. This workflow eliminates seismic inversion by using synthetic AVO gathers and intercept–gradient volumes for direct prediction. It auto-selects machine learning models and regression metrics, adapting to reservoir-specific conditions. A single deterministic wavelet ensures consistency in synthetic data. The approach enhances QI workflows for complex reservoirs, reducing drilling risk and improving characterization. Deployed in ADNOC Offshore, it supports strategic goals of faster execution and digital innovation in reservoir modeling.