Machine Learning–Assisted Reservoir Characterization With Synthetic Well Log

ABSTRACT

Reservoir characterization is the process of understanding the ability of a reservoir to store and produce hydrocarbons (HCs). It plays a pivotal role in the efficient extraction of HC reserves, demanding accurate analysis of subsurface formations and successful field development. Well logs play an important role in reservoir characterization. Conventional well logs often contain missing, corrupted, or incomplete borehole logs across target wells. Generating synthetic well logs provides continuous, standardized petrophysical features, enabling robust predictive modeling of critical reservoir properties without time‐consuming, expensive re‐logging operations. This work investigates the potential of machine learning (ML) models for reservoir characterization using synthetic well logs. In this work, we proposed applying the light gradient boosting machine (LightGBM) and compared its performance with other ML algorithms, such as support vector regression, artificial neural networks, a multilayer perceptron, sparse Gaussian process regression, and a basis transformer‐inspired model. The synthetic well logs were generated using different combinations of inputs and outputs for the ML algorithms and then used for reservoir characterization. The synthetic well‐log generation models were developed to produce shale volume, porosity, and water saturation. The LightGBM predicted reservoir properties at different depths based on patterns inherent in the synthetic log data within the subsurface formations. We evaluated LightGBM using the coefficient of determination (97.33%), mean absolute error (0.0230), root mean square error (0.0218), and computational time (0.086 s) across the training, testing, and validation sets and compared its performance with other ML algorithms. It has been shown that using the LightGBM algorithm for reservoir characterization from synthetic well logs improves petroleum reservoir characterization and insight generation, as it is accurate, computationally efficient, and cost‐effective.