Accelerated Prediction of Clay-Bound Water in Complex Reservoirs using Sequential Learning Techniques

Clay-rich carbonate reservoirs are characterized by highly heterogeneous pore structures, including both intraparticle and interparticle porosities. Neglecting the volume of clay-bound water (V_CBW) in such systems can lead to significant errors in estimating hydrocarbon recovery factors. Conventional well logging methods often struggle to quantify V_CBW in such reservoirs, while advanced techniques like nuclear magnetic resonance (NMR) are expensive and not widely available. This study proposes a novel, cost-effective sequential machine learning (ML) framework for predicting V_CBW using standard well log data. A dataset comprising 5,668 measurements from a clayrich carbonate reservoir was used to train and validate six sequential ML models, including long short-term memory (LSTM), recurrent neural network (RNN), gated recurrent units (GRU), bidirectional LSTM (BiLSTM), convolutional LSTM (ConvLSTM), and multihead attention ConvLSTM models. Model performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE). Among all tested models, ConvLSTM achieved the best performance, with an R² of 0.998 and an RMSE of 0.09 (vol%) on the training dataset, and an R² of 0.958 and an RMSE of 0.49 (vol%) on the testing dataset. BiLSTM, multihead attention ConvLSTM, and LSTM also showed strong predictive capabilities (testing R² > 0.945), while GRU and RNN exhibited lower accuracy (testing R² < 0.90). The study demonstrates that accurate V_CBW prediction from standard well logs enables more precise estimation of producible pore volume (PV) and effective water saturation, minimizes the need for expensive laboratory and NMR measurements. This approach offers a practical tool for reservoir engineers and geoscientists working in clay-rich, complex carbonate systems.