A Globally Oriented and Lightweight Seismic Interpolation Network: More Global than Convolutional Neural Network and More Efficient than Transformer

Abstract Due to certain environmental limitations and economic considerations, seismic interpolation is occasionally used to compensate for sparse and incomplete geometries. In recent years, a number of deep learning (DL)-based methods have shown great promise in seismic data interpolation. Most DL-based interpolation methods adopt architectures that rely on convolutional neural networks, which are inherently limited in capturing global features owing to their local receptive fields. Transformer-based architectures have been recently proposed in several attempts to capture such global features and thus improve the interpolation performance. However, the core operation, multi-head self-attention (MSA), imposes a substantial computational burden. The purpose of this study is to explore a lightweight and globally oriented DL architecture for seismic data interpolation that achieves performance comparable to, or even surpassing, Transformers while substantially reducing the heavy computational cost. To this end, the Mamba module is, for the first time, adopted in the context of seismic data interpolation. Different from Transformers, the Mamba uses the selective state space model (SSM) to capture the global features instead of the MSA. The selection mechanism and hardware-aware algorithm in SSM can significantly alleviate the heavy computational burden by compressing features and reducing memory swaps. Therefore, several Mamba modules are incorporated into a feed-forward neural network, with dense connections introduced to enhance the feature interactions among different layers. In both synthetic and field examples, the proposed dense Mamba-based interpolation network achieves a superior trade-off between the computational cost and interpolation performance, compared with two existing architectures: U-Net and an encoder-stacked Transformer.