Reservoir parameter prediction based on spatial attentive neural processes

For the task of reservoir parameter prediction using seismic attributes,the sparsity of logging samples significantly constrains the application of deep neural network models. To address this issue, we introduce the Neural Processes (NPs) model to establish an effective mapping between well-logging reservoir parameters and seismic attributes. To further enhance the model’s expressive capability,we incorporate dual attention mechanisms inspired by spatial geostatistics into the NPs framework. These mechanisms are respectively designed as spatial and attribute attention modules to achieve the dependency of parameters on spatial locations and attribute correlations. This leads to the development of the Spatially Attentive Neural Processes (SANPs) model. During the optimization process,a context-based meta-learning strategy is employed to effectively learn inter-sample relationships, thereby enabling few-shot learning. In the inference phase,by integrating spatial and attribute correlation representations derived from known input samples as constraints,the model achieves stable predictive ability. Both synthetic data experiments and field data applications demonstrate that our proposed model can accurately predict reservoir parameters within the study area, characterize their spatial distribution patterns, and provide quantitative reliability assessments through variance estimation.