Abstract
In oil and gas exploration, accurate fluid prediction is crucial for strategic decisions. Our study introduces the BiGRU–Transformer, a deep learning model merging bidirectional gated recurrent units (BiGRUs) with the Transformer. The BiGRU interprets well logging data in a bidirectional manner, ensuring a comprehensive understanding of temporal factors at different depths and time intervals, enhancing fluid prediction precision. The model's sensitivity to local temporal dynamics is heightened by its recurrent neural network structure and gating mechanisms. This feature is particularly beneficial in identifying essential details during brief geological occurrences and minor temporal shifts. The bidirectional approach of the model is tailored to accommodate geological variations across assorted depths and timelines, thus boosting its capability to adapt to diverse geological formations. In the Transformer, the self-attention mechanism dynamically adjusts information importance in geological sequences, enabling flexible handling of diverse relationships and a deeper understanding of intricate structures. Primary data come from key well log curves, providing essential features for analysis. Fed into the BiGRU–Transformer, it links fluid attributes with logging parameters. Benchmarking against state-of-the-art models shows our BiGRU–Transformer's superior accuracy and impressive generalization in fluid prediction across various scenarios. This innovation marks a significant advancement in machine learning for well logging, offering a precise tool for geologists and engineers, enhancing exploration and development quality. The BiGRU–Transformer highlights the transformative impact of advanced machine learning in geosciences, paving the way for novel approaches in oil and gas resource exploration and utilization.