Multi-scale information fusion of well-logging data-based deep learning 3D modeling method

Abstract Reservoir modeling based on interpolation and extrapolation of well-logging data provides initial model for seismic inversion. However, the conventional Kriging-based interpolation algorithm relies on the spatial correlation between different observation points in the same stratum, ignoring the correlation between sedimentary strata in different periods within the sedimentary cycle and the current stratum, resulting in unreasonable heterogeneous characteristics of the modeling results. To address this issue, this paper proposes a multi-scale information fusion of well-logging data-based deep learning 3D modeling method (MF-DTMM). This method first constructs the windows of different lengths to extract multi-scale information within the sedimentary cycle. Next, a network architecture combining fully connected neural networks and convolutional neural networks is adopted to establish a nonlinear mapping relationship between multi-scale feature information and spatial interpolation points, thereby constructing the weight coefficient more reasonably. Notably, MF-DTMM integrates the spatial weight mapping features at different scales through a feature fusion mechanism, ensuring that the inter-well interpolation results conform to the internal variation laws of sedimentary units, thereby effectively enhancing the rationality of modeling results. Application of the MF-DTMM method in area A demonstrates that, compared to conventional methods, the MF-DTMM method can construct more reasonable modeling results, providing a new technical approach for reservoir modeling.