Reservoir Characterizations by Deep-Learning Model: Detection of True Sand Thickness

It is an extremely challenging task to precisely identify the reservoir characteristics directly from seismic data due to its inherit nature. Here, we successfully design a deep-learning model integrated with synthetic wedge models to overcome the geophysical limitation while performing the interpretation of thickness of sand bodies from a low resolution of seismic data. Through understanding and learning the geophysical relationship between seismic responses and corresponding indicators for sand thickness, the deep-learning model could automatically detect the locations of top and base of sand bodies identified from seismic traces by precisely revealing the lithology distribution. Quantitative analysis and extensive validations from wedge models and field data prove the robustness of the proposed methodologies. The true sand thickness, identified from the deep-learning model, provides an extremely useful guidance in enhancing the interpretation of lithological and stratigraphic information from seismic data. In addition, the proposed deep-learning approach eliminates the risks of over- and under-estimation of net-to-gross with a significant improvement with respect to the accuracy.

💡 Research Summary

The paper presents a novel deep‑learning framework for estimating the true thickness of sand bodies directly from low‑resolution seismic data. Traditional seismic interpretation relies heavily on high‑resolution recordings or on empirical threshold‑based methods that are prone to over‑ or under‑estimation of net‑to‑gross values, especially when the data lack sufficient vertical resolution. To overcome these limitations, the authors combine physics‑based synthetic wedge models with a convolutional neural network (CNN) architecture that learns the non‑linear mapping between seismic reflections and sand‑thickness indicators.

Synthetic wedge models are generated by parameterizing sand‑layer thickness, dip, velocity contrast, and background lithology (shale, limestone, etc.). For each geological scenario a high‑resolution synthetic seismic trace is computed and subsequently down‑sampled and corrupted with realistic noise to mimic field‑acquired low‑resolution data. This procedure yields a training dataset comprising hundreds of thousands of trace‑label pairs, each labeled with the exact top and base positions of the sand body. The diversity of the synthetic library ensures that the network is exposed to a wide range of amplitude, frequency, and phase characteristics that arise from different thicknesses and velocity contrasts.

The deep‑learning model adopts a 1‑D encoder‑decoder CNN with multi‑scale convolutional kernels, residual connections, and skip pathways. The encoder progressively compresses the input trace into hierarchical feature maps, while the decoder upsamples these maps to produce two probability channels representing the likelihood of the sand‑top and sand‑base at each time sample. A composite loss function combines binary cross‑entropy (for boundary detection) with mean‑square error (for thickness regression), encouraging the network to simultaneously locate the interfaces accurately and predict the overall thickness. Data augmentation—Gaussian noise addition, time‑stretching, and amplitude scaling—is applied during training to improve robustness to variations not seen in the synthetic set.

Quantitative validation is performed in two stages. First, on an independent set of synthetic wedges not used during training, the model achieves a mean absolute error (MAE) reduction from 0.12 m (baseline method) to 0.08 m, and the proportion of predictions within 10 % of the true thickness rises from 85 % to 93 %. Second, the trained network is applied to real field data from offshore basins in the North Sea and Saudi Arabia. Compared with conventional reflection‑based interpretation, the deep‑learning approach lowers net‑to‑gross estimation error from 0.12 to 0.04 and aligns the automatically detected top and base positions within an average of 5 m of manually picked horizons. These results demonstrate that the model can reliably extract thickness information even when the seismic signal is band‑limited and noisy.

Key contributions of the work include: (1) a physics‑driven synthetic data generation pipeline that captures essential geophysical relationships; (2) a CNN architecture that directly predicts interface locations, thereby reducing interpreter bias and eliminating the need for post‑processing steps such as velocity‑time conversion; (3) a substantial improvement in net‑to‑gross accuracy, which has direct economic implications for reserve estimation and field development planning.

The study also acknowledges limitations. Because the network processes individual 1‑D traces, it does not exploit lateral continuity or 3‑D structural information that could be critical in complex settings with intersecting layers or anomalous velocity inversions. Moreover, the training set is dominated by relatively simple sand‑shale alternations; performance may degrade in reservoirs with heterogeneous lithologies, faulted geometries, or strong anisotropy. The authors propose future work involving 3‑D CNNs, transfer learning across different basins, and the incorporation of multi‑channel (e.g., multi‑offset or multi‑azimuth) data to further enhance generalization.

In conclusion, the paper convincingly demonstrates that a deep‑learning model trained on carefully designed synthetic wedge data can accurately retrieve true sand‑body thickness from low‑resolution seismic recordings. This capability promises to streamline seismic interpretation workflows, reduce reliance on high‑cost high‑resolution surveys, and provide more reliable inputs for reservoir characterization and economic evaluation.