Deep-Learning Inversion of Seismic Data (1901.07733v2)

Abstract: We propose a new method to tackle the mapping challenge from time-series data to spatial image in the field of seismic exploration, i.e., reconstructing the velocity model directly from seismic data by deep neural networks (DNNs). The conventional way of addressing this ill-posed inversion problem is through iterative algorithms, which suffer from poor nonlinear mapping and strong nonuniqueness. Other attempts may either import human intervention errors or underuse seismic data. The challenge for DNNs mainly lies in the weak spatial correspondence, the uncertain reflection-reception relationship between seismic data and velocity model, as well as the time-varying property of seismic data. To tackle these challenges, we propose end-to-end seismic inversion networks (SeisInvNets) with novel components to make the best use of all seismic data. Specifically, we start with every seismic trace and enhance it with its neighborhood information, its observation setup, and the global context of its corresponding seismic profile. From the enhanced seismic traces, the spatially aligned feature maps can be learned and further concatenated to reconstruct a velocity model. In general, we let every seismic trace contribute to the reconstruction of the whole velocity model by finding spatial correspondence. The proposed SeisInvNet consistently produces improvements over the baselines and achieves promising performance on our synthesized and proposed SeisInv data set according to various evaluation metrics. The inversion results are more consistent with the target from the aspects of velocity values, subsurface structures, and geological interfaces. Moreover, the mechanism and the generalization of the proposed method are discussed and verified. Nevertheless, the generalization of deep-learning-based inversion methods on real data is still challenging and considering physics may be one potential solution.

• The paper presents SeisInvNet, a novel deep-learning approach that enhances seismic inversion by integrating enriched trace features for accurate velocity model reconstruction.
• The method employs an embedding encoder, spatial alignment module, and decoder with advanced loss functions to overcome full-waveform inversion limitations.
• Experimental results demonstrate substantial improvements in MAE, MSE, and SSIM, promising robust subsurface imaging for applications like hydrocarbon exploration.

Deep-Learning Inversion of Seismic Data: An Analysis

The paper "Deep-Learning Inversion of Seismic Data" explores a novel approach to the seismic data inversion problem utilizing deep neural networks (DNNs). Traditional methods of seismic inversion, particularly full-waveform inversion (FWI), are frequently hampered by their dependence on iterative algorithms. These methods often struggle with nonlinear mapping complexities and nonuniqueness challenges, which may result in suboptimal solutions unless a suitable starting model is provided. Additionally, noise-contaminated data can further complicate conventional approaches. The authors address these limitations by proposing end-to-end seismic inversion networks (SeisInvNets), which effectively leverage the entire seismic dataset without necessitating intensive pre- and post-processing steps.

Methodology

The proposed SeisInvNet represents an architectural innovation aimed at addressing several core challenges inherent in seismic inversion. Fundamentally, the seismic inversion problem is specified by the weak spatial correspondence between time-series seismic data and the spatial velocity model. The authors mitigate this by ensuring each seismic trace contributes to the entire velocity model's reconstruction. This is achieved by enhancing each trace with neighborhood information, observation setup, and global profile context to create enriched, spatially aligned features. Through the learning of these feature maps and subsequent concatenation, a more accurate velocity model reconstruction is achievable.

The network employs an embedding encoder to augment information from each seismic trace, a spatially aligned feature generator to ensure the data aligns with the velocity model, and a decoder to produce the final model through common convolutional neural networks. Loss functions incorporating MSSIM and squared L2 norm are used, aligning reconstructed models with expected geological features and maintaining consistency in velocity values and structure.

Results and Comparison

The performance of SeisInvNet is rigorously assessed against baseline models—fully convolutional network architectures similar to InversionNet—and conventional FWI. The proposed method yields consistently superior outcomes in terms of both the accuracy of velocity values and the delineation of geological interfaces. The assessment leverages metrics such as Mean Absolute Error (MAE), Mean Square Error (MSE), Structural Similarity Index Measure (SSIM), and a novel Soft F measure for edge persistence.

The paper details experiments on a synthetic SeisInv dataset, which encompasses varied velocity models simulating realistic geological conditions. In these tests, SeisInvNet demonstrated a notable enhancement in resolution and accuracy of subsurface images over FWI and baseline models, indicating strong results in interface detection and value alignment across the depth profile.

Implications and Future Research

SeisInvNet's capabilities in seismic data inversion portend significant practical and theoretical advancements. Practically, these networks promise higher efficiency in delineating subsurface properties, which has direct applications in fields such as hydrocarbon exploration and engineering geology. The network's ability to generalize to new and noise-embedded datasets also suggests robustness in real-world applications. Theoretical implications extend toward refining our understanding of how DNNs can be structured to reconcile weakly correlated inputs with spatial models.

Future extensions of this work could include the integration of physical constraints to further enhance results on field data, addressing the persistent challenge of generalization in deep-learning-based inversion methods. Cross-domain adaptation, leveraging pre-trained models across varied geophysical scenarios, presents another promising avenue for research. Moreover, exploration into optimizing network architectures and hyperparameters could yield further performance gains, extending the application of such approaches to increasingly complex geological settings.

In conclusion, the paper introduces an innovative use of DNNs to overcome the limitations of traditional seismic inversion methodologies. By enhancing seismic traces with comprehensive contextual information, SeisInvNets hold substantial promise for improving both the accuracy and reliability of subsurface velocity modeling.