Automatic reconstruction of 3D geological models based on recurrent neural network and predictive learning

The spatiotemporal evolution of sedimentary bodies is difficult to model with traditional geological modeling tools due to its non-stationarity nature. Deep learning algorithms, based on Convolutional Long-Short Term Memory (ConvLSTM) networks, allow to alleviate these limitations as the spatial and temporal dynamics of the sedimentary environment can be explicitly modeled, with structural and attribute information being constructed layer-by-layer. However, due to memory flow limitations and hierarchical visual representations of ConvLSTM, both low-level and high-level semantic features cannot be simultaneously captured. Consequently, small-scale geological features are often overlooked. In addition, long-term modeling and predicting capabilities of ConvLSTM are insufficient during geological sections encoding and forecasting processes. All these challenges might impact the application of ConvLSTM for geo-modeling. To overcome these limitations, we propose herein a geological modeling Recurrent Neural Network (GM-RNN) framework. Specifically, we use zigzag transition path of spatiotemporal memory flow, which allow spatial dynamics at different recurrent layers to interact with each other. Besides, Spatiotemporal LSTM (ST-LSTM) units with memory decoupling are introduced, in which long-term and short-term modeling capabilities for complex spatiotemporal variations can be improved. Finally, Reverse Schedule Sampling (RSS) strategies are used to improve the long-term prediction performances of GM-RNN. Two kinds of Training Images (TIs) are used to assess the simulation performance of GM-RNN. Numerical experiments show that diverse simulations match the corresponding TI in terms of spatial variability, channel connectivity, facies type proportion and spatial distribution patterns. Additionally, we show that 2D geological sections with different scales can be the input of a trained GM-RNN and geobodies are predicted at these scales without compromising the quality of the models.