A new artificial-intelligence system could help geoscientists locate underground oil and gas reservoirs by turning subtle patterns in seismic data into detailed maps of likely reservoir zones. The method, developed by researchers in China and tested on seismic data from the South China Sea, combines two of the most widely used deep-learning architectures—convolutional neural networks and bidirectional long short-term memory networks. The hybrid, called HCBL, is designed to recognize both the spatial structures hidden in seismic images and the way those structures change through geological time. The researchers report that the approach outperformed standalone deep-learning models in accuracy, generalization and interpretability, potentially giving exploration teams a faster way to plan wells and assess complex subsurface formations. The study appears in Earth Science Informatics, where it addresses one of the most persistent problems in petroleum geophysics: predicting the location and character of reservoirs when direct measurements from wells are sparse, unevenly distributed and heavily biased toward the most common rock types.
Reservoir prediction is difficult because the underground target is not a simple cavity filled with fluid. A petroleum reservoir is typically a porous and permeable rock body whose geometry, composition and fluid content vary over distances that may be difficult to resolve. Seismic surveys provide broad coverage by sending acoustic energy into the Earth and recording echoes from underground interfaces, but the resulting signals are indirect. Seismic reflections primarily reveal contrasts in acoustic impedance—the product of rock density and seismic-wave velocity—rather than directly displaying hydrocarbons or pore space. Interpreters must therefore infer lithology, stratigraphic boundaries and reservoir continuity from complicated waveforms. Well logs offer more direct information, recording properties such as density, resistivity, gamma radiation and sonic velocity, but wells are expensive and sample only narrow vertical columns. The mismatch between dense seismic coverage and sparse well control creates an ideal but challenging application for machine learning, provided that algorithms can avoid mistaking local patterns for geological rules.
The HCBL system begins by transforming seismic measurements with a technique known as inversion spectral decomposition, or ISD. Conventional seismic sections are often examined in the time domain, where reflection amplitudes are plotted against the arrival time of returning waves. Spectral decomposition instead separates the signal into its constituent frequencies, producing a time-frequency representation. This can reveal features that are blurred in the original data: thin beds, channel margins, discontinuities and changes in depositional architecture may respond differently at different frequencies. Inversion spectral decomposition seeks to improve that representation by estimating a higher-resolution time-frequency spectrum from the observed seismic signal. The result is not merely a single seismic image but a sequence of spectral frames, each emphasizing a different aspect of the subsurface response. These multi-frame spectra become the input to the neural network, giving it a richer description of the geological scene than a single attribute or raw trace might provide.
The first major component of HCBL is a time-distributed convolutional neural network. Convolutional neural networks, or CNNs, are particularly effective at extracting spatial features because they apply learnable filters across neighboring pixels or data points. Early layers may identify simple patterns such as edges, amplitude changes or localized textures; deeper layers combine those signals into larger structures, such as dipping reflectors, channels or geological boundaries. In the HCBL design, the CNN processes each time-frequency frame while preserving the relationship among successive frames. This time-distributed arrangement matters because the spectral images are not independent photographs. They are different views of the same subsurface volume, sampled across time and frequency. The network can consequently learn spatial signatures within each frame before passing the resulting feature sequence to the second part of the architecture. In effect, the CNN acts as a geological pattern detector, compressing complex seismic textures into a set of features that can be analyzed over their broader sequence.
That sequence is interpreted by a bidirectional long short-term memory network, or BiLSTM. LSTM networks are a form of recurrent neural network created to model dependencies across ordered data, while reducing the tendency of ordinary recurrent networks to lose information over long sequences. Their internal gates regulate which information should be retained, updated or discarded. A bidirectional version reads the sequence in both directions, allowing the model to use contextual information from earlier and later positions rather than relying only on a one-way progression. For seismic interpretation, this gives the system a way to connect features that occur at different depths or times along a trace and to identify coherent patterns extending across adjacent spectral frames. A reservoir signature may be weak in one portion of the data but become recognizable when considered alongside neighboring layers or frequency responses. By combining CNN-derived spatial features with BiLSTM-derived sequential context, HCBL attempts to model the subsurface as a connected spatial-temporal system rather than a collection of isolated measurements.
The researchers also targeted a statistical problem that can quietly undermine automated geological interpretation: class imbalance. In a training dataset, some categories—such as common surrounding rocks—may appear far more frequently than reservoir-bearing intervals or less abundant lithologies. A model trained with ordinary cross-entropy loss can achieve an apparently high overall accuracy simply by favoring the majority class. That performance may conceal poor recognition of the rare categories that matter most for exploration decisions. HCBL addresses this by using a weighted cross-entropy loss function. During training, errors involving minority classes receive greater weight, increasing the penalty when the network overlooks them. The objective is not to force the model to label everything as a reservoir, but to make the optimization process pay sufficient attention to underrepresented geological classes. This balancing strategy can improve sensitivity to subtle reservoir signatures while also reducing the risk that the network will learn a distorted picture of the subsurface from uneven training examples.
The combined system was applied to seismic data from the South China Sea, a region where offshore exploration depends on extracting as much information as possible from geophysical surveys before drilling. According to the study, the HCBL results were more accurate than those from independent deep-learning models and remained more reliable when applied beyond the precise examples used for training. This ability to generalize is crucial. Deep-learning systems can be exceptionally good at recognizing patterns in familiar data while failing when acquisition conditions, geological settings or signal quality change. A model that performs well only in one survey area would have limited value. The reported improvement in interpretability is also significant, although the source material does not provide a single numerical score or a complete set of benchmark metrics in its abstract. In practical terms, a more interpretable prediction can help geoscientists compare the network’s output with seismic reflections, well information and established geological models rather than treating an automated classification as an unquestionable answer.
If validated across additional basins and acquisition campaigns, the method could alter how exploration teams screen large seismic volumes. Modern surveys can contain millions of traces, and manual interpretation requires specialists to inspect horizons, faults, amplitude anomalies and depositional patterns over long periods. An algorithm able to convert time-frequency spectra into reservoir predictions could rapidly highlight zones for closer examination, reducing the initial search space. The technology might also support infill-well planning, reservoir delineation and development decisions by identifying probable continuity between wells. Yet the system is not a substitute for drilling, petrophysical analysis or geological judgment. Seismic signals remain non-unique: different combinations of rock type, porosity, fluid and thickness can produce similar responses. A prediction is therefore a probability-informed interpretation, not a direct observation of oil or gas. The strongest use case is likely to be a human-machine workflow in which artificial intelligence performs exhaustive pattern recognition while experts test the results against physical constraints and independent evidence.
The study’s data limitations underline why such validation will matter. The authors state that no datasets were generated or analyzed during the current study and that the seismic data used are not publicly available because of confidentiality restrictions imposed by the provider. That prevents outside researchers from independently reproducing the reported South China Sea experiment from the published data alone. The researchers do provide details of a code package called CNN_BiLSTM, written in Python and Matlab, with modest stated hardware requirements and a program size of 23.5 megabytes, although the source listing does not provide a working public download link. Future studies could strengthen the case for HCBL by testing it on openly accessible datasets, different sedimentary environments and wells that were withheld from training. They could also report class-by-class precision, recall, uncertainty estimates and performance against established geophysical inversion methods. For now, the work illustrates both the promise and the boundaries of deep learning beneath the Earth: when carefully designed to combine frequency, space, sequence and data balance, neural networks can make hidden geological patterns easier to see—but their most important predictions still require the discipline of independent verification.
