Hybrid Seismic Meta-Attributes and Machine Learning Approach for Robust Reservoir Property Prediction

This study evaluates the efficacy of an innovative approach that combines seismic meta-attributes with machine learning (ML) techniques to predict Acoustic Impedance (AI) and subsequently enable the prediction of reservoir properties, such as Porosity. Furthermore, it investigates the potential of ML as a faster and more scalable alternative to traditional inversion methods, examining its capacity to efficiently handle large datasets, reduce processing times, and streamline complex workflows. This study combines seismic meta-attributes with ML algorithms to accurately predict reservoir properties, specifically porosity from Acoustic Impedance. Our two-stage ML workflow fills gaps in log data to maximize well data coverage at reservoir level and integrates log and 3D seismic attributes to predict reservoir properties. Eight wells are utilized, with six for training and two for blind testing. Our ML models use various inputs, including conditioned well logs, 3D post-stack and pre-stack seismic volumes, and twelve seismic attributes. The Fully Convolutional Neural Network with Transformer (FCT++) outperforms XGBoost, Random Forest, and conventional Neural Networks in predicting reservoir properties, demonstrating its superior effectiveness for this application. Our machine learning workflow showed excellent predictive power, achieving high accuracy in forecasting acoustic impedance and porosity. Among four algorithms tested, FCT++ performed best, with correlation coefficients of 0.97 for acoustic impedance and 0.77 for porosity at blind wells. Our cascading workflow effectively tackled multi-step prediction challenges, estimating porosity using predicted acoustic impedance as input. We employed Monte Carlo dropout and ensemble stacking to quantify uncertainty, boosting prediction reliability. The workflow efficiently handled large datasets, reducing turnaround times and providing reliable property estimates. Its scalability and accuracy make it a promising alternative to traditional methods. This study introduces a novel seismic meta-attribute workflow that integrates machine learning and uncertainty estimation to predict reservoir properties. By leveraging the capabilities of advanced algorithms like FCT++ and integrating cascading processes, this approach can augment and enhance traditional physics-based methods, offering a more comprehensive and accurate solution. This innovative approach offers enhanced scalability, accelerated processing, and robust property prediction, marking a significant advancement in seismic reservoir characterization using machine learning.