Recently, intelligent data processing and interpretation based on deep learning (DL) have received considerable attention. Training data are vital for DL-based approaches. In geosciences, researchers have been facing a significant obstacle, i.e., the absence of authoritative and representative open data for training and testing artificial neural networks (ANNs). Although open-source works in geosciences are increasing, the quantity is currently limited. With the aid of the Society of Exploration Geophysicists (SEG) 2020 Machine Learning (ML) Blind-Test Challenge Data, we open source a package for DL-based seismic facies classification (SFC). We regard SFC as translating the seismic data into the facies by training a flexible end-to-end encoding–decoding-style ANN “BridgeNet” revised from U-Net. Inspired by the residual network (ResNet), we further enhance the BridgeNet by inserting the identity shortcut connections, which can theoretically ease the notorious problem of vanishing/exploding gradients. The evaluated framework, however, is not restricted to SFC. We hope that it can provide some insights that help researchers to construct and train ANNs that yield reliable and robust results in their own tasks. We carry out benchmarking experiments to investigate some crucial factors impacting the ANN’s performance to elucidate how we obtained our current optimal results on the SEG 2020 ML Challenge Data. The accuracy and continuity of predicted facies along the training and testing sections indicate that the results are consistent with geologic sedimentation, verifying the generalization capability of the enhanced flexible end-to-end encoding–decoding-style BridgeNet. For SFC, we recommend a memory-saving sparse categorical cross-entropy (CC) loss function to improve the efficiency. The codes are available at https://doi.org/10.5281/zenodo.5787673 .

Low frequencies are vital for full-waveform inversion (FWI) to retrieve long-scale features and reliable subsurface properties from seismic data. Unfortunately, low frequencies are missing because of limitations in seismic acquisition steps. Furthermore, there is no explicit expression for transforming high frequencies into low frequencies. Therefore, low-frequency reconstruction (LFR) is imperative. Recently developed deep-learning (DL)-based LFR methods are based on either 1D or 2D convolutional neural networks (CNNs), which cannot take full advantage of the information contained in 3D prestack seismic data. Therefore, we present a DL-based LFR approach in which high frequencies are transformed into low frequencies by training an approximately symmetric encoding-decoding-type bridge-shaped 3D CNN. Our motivation is that the 3D CNN can naturally exploit more information that can be effectively used to improve the LFR result. We designed a Hanning-based window for suppressing the Gibbs effect associated with the hard splitting of the low- and high-frequency data. We report the significance of the convolutional kernel size on the training stage convergence rate and the performance of CNN’s generalization ability. CNN with reasonably large kernel sizes has a large receptive field and is beneficial to long-wavelength LFR. Experiments indicate that our approach can accurately reconstruct low frequencies from bandlimited high frequencies. The results of 3D CNN are distinctly superior to those of 2D CNN in terms of precision and highly relevant low-frequency energy. FWI on synthetic data indicates that the DL-predicted low frequencies nearly resemble those of actual low frequencies, and the DL-predicted low frequencies are accurate enough to mitigate the FWI’s cycle-skipping problems. Codes and data of this work are shared via a public repository.

Low-frequencies (LF) are vital for full-waveform inversion (FWI) to retrieve accurate long-scale features and then reliable subsurface properties from seismic data, which are missing in the acquired seismic data due to limitations in acquisition steps. Low-frequency reconstruction (LFR) is imperative. The recently developed deep learning (DL) based LFR methods are based on 1D/2D convolutional neural networks (CNNs), which cannot take full advantage of the information in the 3D seismic data. We develop a DL-based approach for LFR in which the high-frequencies (HF) (e.g., ≥ 10 Hz) are transformed into the LF counterparts (i.e., < 10 Hz) by training an end-to-end 3D CNN. Benchmarking experiments on LFR for FWI indicate the FWI result of the DL-predicted LF nearly resembles that of the true LF, and the DL-predicted LF are accurate enough to solve the cycle-skipping problem of FWI.