RENJUN WEN

CEO, SUBSURFACEAI INC., CALGARY, AB

RENJUN@SUBSURFACEAI.CA

 

“From feature enhancement to machine interpretation, attributes remain the bridge

between seismic signals and geologic meaning.”

Abstract

Artificial intelligence (AI) is rapidly changing how seismic interpreters map faults, channels, facies, and other geobodies. Modern deep learning can often learn useful representations directly from raw amplitudes, yet in day-to-day projects interpreters still face non-stationary wavelets, variable signal-to-noise, acquisition footprint, stitching of vintages, and geologic complexity that complicate “end-to-end” approaches. Seismic attributes continue to matter because they are fit-for-purpose transformations that enhance specific aspects of the seismic response-discontinuity, texture, frequency content, geometry, and elastic sensitivity—making interpretation easier for both humans and machines. A critical point is that attributes are not geology by themselves: they highlight patterns, but the geological meaning of those patterns must be assigned by an interpreter or learned by an AI model trained with interpreter-provided labels. This article revisits why attributes remain essential in modern interpretation workflows and presents three practical examples: (1) using semblance/coherence to expose discontinuities while training AI to separate faults, channels, and volcanic boundaries, (2) using RGB spectral decomposition blends to illuminate channels while training AI to avoid confusing fault lineaments as channels, and (3) predicting pseudo-log volumes (e.g., gamma ray) by training well logs with seismic attributes, where attribute generation and selection control whether the resulting property volumes are trustworthy. The discussion concludes with practical guidelines for building robust attribute-plus-AI workflows that improve repeatability, reduce cycle time, and strengthen quality control.

Introduction

Artificial intelligence (AI) is now a practical part of many seismic interpretation teams’ toolkits. Where an interpreter previously spent days or weeks building fault frameworks, outlining channel systems, or classifying seismic facies, supervised machine learning and deep learning can often produce survey-wide “probability volumes” in hours once a training set has been generated. Those probability volumes are not a replacement for interpretation; rather, they are a new representation of interpretation: a model’s spatial estimate of where a feature is likely to occur, consistent across the full volume and easy to integrate with manual picks and geologic context.

With this momentum, an understandable question keeps resurfacing: if neural networks can learn complex features directly from raw amplitudes, do we still need seismic attributes? Some deep learning papers even highlight that a convolutional neural network can take a raw image as input, effectively learning its own internal “attributes” during training (e.g., Waldeland et al., 2018; Wu et al., 2018; Kaur et al., 2023). That is true in principle and sometimes true in practice.

In most real projects, however, seismic amplitudes are not a clean, stationary, single-domain dataset. Interpreters deal with acquisition footprint (Marfurt et al., 1998b), variable wavelets, processing differences between vintages, coherent noise, illumination changes, and the fact that the same geologic feature can express itself differently across a survey. In this environment, attributes remain essential because they deliberately enhance aspects of the seismic response that are diagnostic for specific interpretation tasks.

More importantly, attributes clarify a foundational point that is easy to overlook when AI results look impressive: seismic attributes are excellent at highlighting potential geologic features, but they do not provide geologic meaning by themselves. A semblance slice may show many discontinuities, yet the image does not tell you which discontinuities are faults, which are channel edges, and which are volcanic boundaries or non-geologic artifacts. The missing step is interpretation, and interpretation can be performed by a human interpreter or learned by an AI model trained with interpreter-provided labels.

This article expands on that idea and argues that attributes are not becoming obsolete in the age of AI; they are becoming more valuable. They serve two complementary roles: (1) they help humans see patterns and build geologic hypotheses, and (2) they provide fit-for-purpose inputs that help AI learn faster, generalize better, and produce probability volumes that are easier to trust and quality-control.

Seismic attributes: a short refresher

A seismic attribute is any measurable quantity derived from seismic data that helps visualize or quantify a characteristic of interest. Attributes can be computed on a single trace (trace attributes) or from multiple neighbouring traces (volume attributes), and they can be extracted along interpreted surfaces or computed throughout a 3D volume. Decades of industry experience show that the most useful attributes are those that enhance a specific geologic expression in a way that is interpretable and robust to noise (Chopra and Marfurt, 2005).

Although hundreds of named attributes exist, most fall into a few families that can be understood through the lens of what aspect of the signal they emphasize:

1) Amplitude and energy attributes: Examples include envelope (instantaneous amplitude), RMS amplitude, maximum/minimum amplitude, average absolute amplitude, and ratios that combine amplitude and frequency (e.g., sweetness). These attributes help highlight bright spots, tuning effects, and energy-related geomorphology when tied to a stratigraphic interval (Taner et al., 1979).

2) Phase, polarity, and waveform-shape attributes: Instantaneous phase, cosine of phase, reflection strength, and similar measures can help correlate events and interpret continuity where amplitude varies. Waveform classification methods (including clustering and self-organizing maps) can be viewed as higher-level waveform attributes often used to characterize seismic facies.

3) Frequency and spectral attributes: Instantaneous frequency, dominant frequency, bandwidth, and spectral decomposition volumes (computed using Fourier or wavelet methods) help interpreters “tune” into bed thickness, depositional elements, and interference patterns. RGB blends of spectral decomposition are widely used to illuminate channel systems and stratigraphic architecture (Barnes, 1993; Partyka et al., 1999).

4) Discontinuity and similarity attributes: Coherence, semblance-based coherency, eigenstructure coherence, variance, similarity, and related measures quantify how similar neighbouring traces are. They are widely used for mapping faults, fracture corridors, stratigraphic edges, karst collapse, and other features expressed as waveform discontinuities (Bahorich and Farmer, 1995; Bednar, 1998; Marfurt et al., 1998a; Gersztenkorn and Marfurt, 1999).

5) Geometric attributes: Dip, azimuth, curvature (most-positive, most-negative, Gaussian curvature, etc.), and measures derived from reflector geometry highlight subtle folding, flexures, fault damage zones, and fracture trends. They are often computed after dip estimation and sometimes benefit from dip-steered (structure-oriented) filtering (Chopra and Marfurt, 2007).

6) Elastic and rock-physics-related attributes: Pre-stack and partial-stack attributes, AVO/AVA intercept and gradient, elastic impedance, and inversion-derived impedance volumes are used when the target is not only geometry but also lithology, porosity, or fluid effects (Shuey, 1985; Castagna and Swan, 1997).

An important practical point is that attributes are not “free” information. Each attribute embeds assumptions: window length, frequency bandwidth, dip estimation method, smoothing, and whether the attribute is computed along structural dip or in fixed time. Bad parameter choices can smear features, create artifacts, or suppress the very signal one is trying to highlight. In other words, attributes can improve interpretability, but they can also amplify problems if the input-seismic is poorly conditioned or if an attribute is applied without a clear purpose.

For that reason, a modern best practice is to compute fit-for-purpose attributes—a small set that is physically and geologically motivated for the feature of interest, rather than generating dozens of volumes and hoping that something will work. This principle is just as important for AI workflows as it is for human interpretation.

Attributes are not geology: interpretation is the missing step

Seismic interpretation is fundamentally pattern recognition in the presence of uncertainty. Interpreters integrate multiple cues — continuity, amplitude, frequency, geometry, stratigraphic context, and regional geologic knowledge — to decide what a pattern means. Attributes were invented to make those cues more visible and measurable.

AI-based interpretation does not remove the need for meaning; it changes how meaning is propagated. In a supervised workflow, an interpreter labels examples (fault sticks, channel polygons, facies classes, etc.) on a small subset of slices or sections. An AI model learns a statistical mapping from input data (amplitudes and/or attributes) to labels, then applies that mapping across the full survey to produce probability volumes (fault probability, channel probability, facies probability, and so on).

Two observations follow immediately:.

• A model can only learn what is represented in the inputs: If the input representation makes a feature difficult to see (low contrast relative to background noise or competing patterns), the model must work harder and usually needs more training examples.

• A model can only assign geologic meaning if meaning exists in the labels: A discontinuity attribute highlights “something changed laterally”, but the labels teach whether that change corresponds to a fault, a channel edge, a volcanic feature, or a processing artifact.

Attributes help AI for the same reason they help humans: they enhance separability. In machine learning terms, they create feature spaces where the classes are easier to distinguish. Semblance improves the contrast of discontinuities; curvature emphasizes flexures; spectral decomposition isolates frequency components, that respond to thickness or interference; pre-stack attributes capture angle-dependent behavior linked to elastic contrasts. When those cues are explicit in the input data, a model often learns faster and generalizes better with fewer labels (Shuey, 1985; Partyka et al., 1999; Chopra and Marfurt, 2005, 2007).

This does not contradict the fact that deep networks can learn internal representations from raw amplitudes. Instead, it reflects a pragmatic workflow reality: interpreters rarely have infinite training labels, and the data rarely behave like a perfectly curated computer vision dataset. Attributes inject domain knowledge in a controllable way. They also support quality control, because interpreters can check whether a model’s probability volume is consistent with the attribute expressions they already understand.

In the sections that follow, three examples illustrate this interpretation principle: attributes highlight candidate features, but labels (human interpretation) teach an AI model to attach the correct geologic meaning and to generate useful probability volumes.

Figure 1. Conceptual attribute-plus-AI workflow. Seismic amplitudes are conditioned and transformed into fit-for-purpose attribute volumes. An interpreter labels a small number of representative slices/sections. A supervised model is trained to learn geologic meaning and generate survey-wide probability volumes (e.g., faults, channels, volcanics) or pseudo-log volumes (e.g., gamma ray). Outputs are quality-controlled against seismic and attributes, and labels/models are iteratively refined.

Example 1: Semblance shows discontinuities; AI learns faults vs channels vs volcanics

Discontinuity attributes are among the most widely used seismic attributes, and for good reason. Many structural and stratigraphic boundaries manifest as abrupt lateral changes in waveform, which is exactly what similarity measures are designed to quantify. Coherence and semblance-based attributes were originally popularized as a “coherence cube” approach that revealed faults and stratigraphic features directly in 3D volumes, giving interpreters a way to build structural frameworks before detailed horizon picking. Later work improved robustness and interpretability through dip correction and eigen-structure approaches (Bahorich and Farmer, 1995; Bednar, 1998; Marfurt et al., 1998a; Gersztenkorn and Marfurt, 1999).

A semblance (or coherence) stratal slice is often a dramatic image: faults, channel edges, levees, collapse features, reef margins, and other discontinuities can jump out, even when the corresponding amplitude slice looks subtle. Yet that dramatic visibility can be misleading if one forgets what the attribute actually measures. Semblance is agnostic: it reports “low similarity”, not “fault” or “channel.” Any process that produces waveform changes—geology, acquisition footprint, processing artifacts, multiples, residual statics—can create discontinuity signatures (Marfurt et al., 1998b).

Figure 2 illustrates this problem using a 3D seismic dataset from offshore New Zealand that contains extensional faulting, channelized stratigraphy, and volcanic features. On the semblance slice, discontinuities are abundant. Some are clearly linear and align with fault trends; others are curvilinear and bound sinuous channel forms; others outline volcanic edifice boundaries and related intrusive/extrusive features. Without additional context, the semblance image does not specify which discontinuity belongs to which geologic class.

This is where interpretation—human or AI—must enter. A practical approach is to treat the discontinuity attribute as an excellent “candidate generator,” and then teach an AI model to assign meaning. The training strategy can be surprisingly efficient: an interpreter labels a small number of time slices (or crossline/in-line sections) with separate classes, for example: Fault, Channel edge, Volcanic boundary, and Background/Other. Those labels are not meant to be exhaustive picks of every feature; instead, they are representative examples that capture the visual variability of each class in different parts of the survey.

Once training labels exist, a supervised segmentation model can be trained to output multiple probability volumes—one per class—rather than a single “discontinuity probability”. Inputs can include the semblance volume itself and a small set of complementary attributes that help disambiguate classes. For example, adding curvature or dip variance can reinforce structural fabric around faults (Chopra and Marfurt, 2007); adding spectral decomposition or amplitude texture attributes can help emphasize depositional geomorphology around channels (Partyka et al., 1999); and adding measures of reflectivity-character can help isolate volcanic facies with distinct waveforms. CNN-based segmentation and fault-detection workflows provide practical templates for these kinds of applications (Waldeland et al., 2018; Wu et al., 2018).

Figure 3 shows a stratal slice that is co-rendering of faults, channels, and volcano probability volumes predicted from AI models trained from the semblance slice shown in Figure 2b. . The same discontinuity field is decomposed into separate probability volumes that better match geologic meaning. The fault probability volume is expected to highlight linear, laterally continuous breaks and associated damage zones; the channel probability emphasizes sinuous margins and channel belts; and the volcanic probability outlines edifice boundaries and volcanic-related discontinuities. Importantly, the AI models are learning the interpreter’s definition of these classes, which can be tuned to the project objective (e.g., structural framework for trap analysis versus channel fairway delineation).

The practical value is twofold. First, the AI result is immediately more actionable than the raw semblance image because it separates competing hypotheses. Second, it accelerates interpretation without hiding the evidence: interpreters can QC each probability volume against the underlying attributes and amplitudes, and they can update labels where the model fails. The workflow becomes an iterative interpretation loop rather than a one-time “black box” prediction.

Figure 2. Offshore New Zealand example (stratal slice). (a) Conventional seismic amplitude slice. (b) Semblance/coherence slice computed over the same interval. Discontinuities are abundant in semblance, but their geologic meaning is ambiguous: some correspond to faults, some to channel margins, and some to volcanic boundaries or other features. (c) A cross section showing the horizon on which the stratal slice is made.

Figure 3. Multi-class interpretation learned from discontinuity attributes on the same stratal slice shown in Figure 2 (fault probability, channel probability, volcanic probability).

Example 2: RGB spectral decomposition illuminates channels; AI learns channel probability volumes

Spectral decomposition attributes are among the most powerful tools for stratigraphic interpretation because they convert a broadband seismic signal into frequency components, that usually respond to bed thickness and interference patterns (Partyka et al., 1999). In practice, interpreters often compute a set of magnitude volumes at selected frequencies (or spectral bands) and then blend three of them into an RGB display. When the chosen frequencies are appropriate for the target interval, the resulting RGB blend can illuminate channels, bars, lobes, and other geomorphic elements with clarity, that is difficult to achieve from conventional amplitude slices.

Figure 4 provides an example of an RGB spectral decomposition blend through a channelized interval. The channel fairways are expressed as coherent, sinuous color patterns that track changes in thickness and tuning across the system. In a manual workflow, interpreters commonly map channels directly on the RGB image because it maximizes visual contrast.

However, the same issue arises as with discontinuity attributes: the attribute highlights patterns but does not guarantee geologic meaning. Lineaments in an RGB blend may be channel margins, but they may also be faults, fracture corridors, or acquisition/processing artifacts that have their own frequency signatures. In some cases, faults cut channels and produce linear spectral edges that can be misinterpreted as channel boundaries. In other cases, spectral anomalies may be caused by wavelet variations or localized noise, especially near survey edges or under poor illumination, and acquisition footprint can imprint lineaments that are unrelated to geology (Marfurt et al., 1998b).

A supervised AI model can be trained to leverage the strength of spectral decomposition while avoiding these pitfalls. The interpreter labels channels on a small number of RGB slices (or on the underlying spectral components), explicitly distinguishing channel geobodies from non-channel lineaments such as faults. Because the RGB blend already concentrates the relevant stratigraphic information, the model often requires fewer labels than a model trained directly on full-stack amplitude. In effect, the attribute reduces the dimensionality of the problem to a representation where channels are more separable (Waldeland et al., 2018; Kaur et al., 2023).

The output is a channel probability volume (Figure 5) that can be generated quickly for the full survey. Compared with interpreting channels directly from the original seismic, the probability volume is easier to threshold, extract, and integrate with geomorphologic measurements (width, sinuosity, orientation) and with reservoir modeling workflows. The probability volume also supports rigorous QC: uncertain areas can be targeted for additional manual interpretation and label updates, improving the model iteratively.

This example illustrates a general principle: spectral decomposition attributes are not optional “pretty pictures”. They are effective information compressors that make stratigraphic patterns explicit, enabling both faster manual interpretation and more efficient AI training. In the age of AI, that makes them more useful, not less.

Figure 4. RGB spectral decomposition blend through a channelized interval. Three spectral magnitude volumes at selected frequencies are mapped to red, green, and blue channels to enhance geomorphic patterns. Sinuous channel fairways are emphasized, but some linear lineaments may correspond to faults or non-channel features.

Figure 5. Channel probability volume generated by training an AI model on a small set of interpreter-labeled RGB spectral decomposition slices. The probability output highlights channels consistently across the survey and suppresses non-channel lineaments that appear in the RGB blend (e.g., faults).

Example 3: Predicting pseudo-log volumes from wells and attributes

Many interpretation problems ultimately require property estimation, in addition to geometric mapping. Wells provide high-resolution measurements (gamma ray, density, porosity, VP/VS, lithofacies), but they sample the subsurface sparsely. Seismic provides continuously denser coverage, but the relationship between amplitudes and rock properties is indirect and often biased, uncertain and non-unique—data is band-limited with much lower vertical resolution, and is affected by source wavelet, tuning, and processing pitfalls. Machine learning provides a practical bridge between well and seismic domains: train a model on well-log targets using seismic attributes sampled at the well locations, then predict a “pseudo-log” volume across the full seismic grid (Hampson et al., 2001).

This workflow is conceptually simple (Figure 6), but its success depends on how closely the attributes capture the relevant feature. A gamma-ray log, for example, is not a direct elastic property. It is often a proxy for shale content and depositional facies, which may correlate with impedance, amplitude texture, spectral response, and stratigraphic architecture. In some settings, a small set of attributes can capture those relationships surprisingly well; in others, the mapping is weak unless additional conditioning or stratigraphic constraints are applied.

A practical pseudo-log workflow contains several steps that must be executed carefully:

1) Well-to-seismic tie and stratigraphic consistency: Log targets should be tied to seismic time (or depth) with appropriate wavelets, and the prediction interval should be defined stratigraphically. Mis-ties or mixing of multiple depositional regimes within one training window can dominate the error budget.

2) Attribute extraction at wells: Attributes are sampled at the well locations within the target interval. Choices such as window length, dip steering, and whether attributes are computed on conditioned seismic matter because they control repeatability and sensitivity to noise.

3) Model selection and validation: Regression models range from linear and partial least squares approaches to nonlinear methods such as random forests, gradient boosting, and neural networks. Regardless of model choice, cross-validation and blind wells are essential. A model that fits training wells perfectly may still fail away from wells if it is overfitting noise or acquisition footprint.

4) Prediction and uncertainty: A single predicted volume is rarely sufficient for decision making. Interpreters want to know where the model is reliable and where it is over or under-estimating Uncertainty can be assessed using blind-well residuals, ensemble methods, or spatial measures of similarity to the training data.

In this context, “attribute engineering” is not a buzzword; it is often the difference between a trusted property volume and an unusable one. Many of the most reliable workflows use a small set of interpretable attributes (e.g., impedance, spectral measures, continuity/texture measures) combined with careful QC against wells and geologic expectations. In some cases, attributes can be generated or optimized using machine learning itself — feature selection, dimensionality reduction, or learned representations—but the end goal remains the same: construct predictors that are geologically meaningful and stable enough that interpreters can trust the resulting pseudo-log volumes.

The value proposition is significant. A well-trained pseudo-log volume can guide facies mapping between sparse wells, highlight lateral changes in reservoir quality, and provide screening volumes for prospect ranking or development planning. But that value only exists when the attribute set and the training workflow are designed for interpretability and robustness, not just numerical accuracy on a training set.

Figure 6. Pseudo-log prediction workflow and example results. (a) Training data preparation: tie wells to seismic and extract attribute samples in a stratigraphic window. (b) Model training and blind-well validation (e.g., measured vs predicted gamma-ray crossplot). (c) Examples of predicted pseudo-gamma volume and log facies volume.

Practical guidelines for robust attribute-plus-AI workflows

The three examples above share a common theme: attributes and AI are complementary. Attributes make the signal more diagnostic; AI propagates interpretation consistently. The following practices have repeatedly proven important when building attribute-plus-AI workflows, whether the goal is structural mapping, stratigraphic geomorphology, or property prediction.

1) Start with seismic conditioning and QC: Attributes amplify what is already present in the data. If acquisition footprint, residual statics, or processing artifacts dominate the input, they will often dominate the attribute volumes and the AI results. Footprint suppression and consistent conditioning have long been recognized as prerequisites for reliable attribute mapping (Marfurt et al., 1998b). Structure-oriented filtering and consistent amplitude/phase conditioning should be considered early, and attribute anomalies should be cross-checked against raw amplitudes.

2) Choose attributes that are physically and geologically motivated: For faults, discontinuity and curvature families are often the most informative (Bahorich and Farmer, 1995; Chopra and Marfurt, 2007). For channels, spectral decomposition and amplitude texture attributes frequently outperform generic discontinuity measures (Partyka et al., 1999). For properties, incorporate elastic-sensitive attributes (impedance, AVO/AVA terms, inversion products) where the physics supports it (Shuey, 1985; Castagna and Swan, 1997). Avoid “kitchen-sink” attribute sets that add redundancy and encourage spurious correlations.

3) Respect scale: Window size and frequency selection determine what geologic scale an attribute responds to. A coherency window that is too large can smear small faults; a spectral blend built from inappropriate frequencies can hide thin-bed responses. When possible, compute attributes at multiple scales and let the interpreter select which scale matches the target feature.

4) Build labels that teach meaning, not just edges: The most efficient labels are not always the most detailed. For supervised segmentation, sparse but representative labels across different parts of the survey often outperform a single densely labeled area. When a single attribute expression can correspond to multiple geologic classes (e.g., discontinuities from faults, channels, and volcanics), label those classes explicitly so that the model can learn to separate them.

5) Validate with blind tests and geology: Probability volumes should be validated not only by visual plausibility but also by blind slices, blind wells, and geologic consistency checks. For example, channel probability should honor stratigraphic framework and known depositional trends; fault probability should be consistent with regional stress regime and structural style. For pseudo-log volumes, blind-well crossplots and residual maps are non-negotiable.

6) Treat AI outputs as interpretation products with uncertainty: A probability volume is not a deterministic map. Thresholds should be tuned to the decision context, and uncertainty should guide where to invest additional interpretation effort. Iterative cycles—label, train, predict, QC, refine—are often more productive than seeking a “perfect” one-shot model.

7) Use attributes to improve trust and clarity: One reason interpreters trust attributes is that their behavior is familiar: coherence highlights discontinuities; curvature emphasizes flexures; spectral decomposition responds to tuning. When AI outputs are compared against these familiar cues, it becomes easier to understand where the model is correct, where it is missing features, and where it is being fooled by artifacts. This transparency is critical for adoption in high-stakes interpretation projects.

Integration matters: connecting attribute generation and AI iteration

Often attribute computation and AI model building are still separate workflows performed in different software environments. That separation slows iteration: interpreters compute attributes in one tool, export volumes, build training sets in another tool, and then struggle to keep track of versions, parameters, and QC results. The result is often a gap between the people who understand the geology and the people who build the models.

Integrated workflows aim to remove that friction by keeping attribute generation, visualization, training data creation, model training, and prediction in one environment. The practical benefit is not only speed, but also reproducibility. When the attribute parameters, training labels, and model versions are captured together, teams can audit how a probability volume was produced and can update it when seismic processing changes or new wells are added.

Conclusions

AI is transforming seismic interpretation, but it is not making seismic attributes obsolete. On the contrary, attributes remain the most practical way to enhance the parts of the seismic response that matter for specific interpretation objectives—discontinuities for faults, spectral response for stratigraphy, curvature for subtle structure, and elastic attributes for reservoir properties.

The central message is simple: attributes highlight candidate features; interpretation assigns meaning. Whether that interpretation is performed directly by a human interpreter or learned by an AI model, the workflow still depends on representations that make geologic patterns separable and interpretable. Semblance does not know what a fault is; an RGB blend does not know what a channel is. Human labels teach an AI to distinguish those classes, and well-designed attributes make that teaching efficient and reliable.

For interpreters and teams adopting AI, the opportunity is to treat attributes as a strategic advantage. Compute fewer but better attributes, build labels that encode geologic meaning, validate rigorously, and use probability volumes as interpretation products with uncertainty. When done well, the combination of attributes and AI reduces cycle time, improves repeatability, and strengthens QC—allowing teams to move from seismic signals to geologic decisions with greater confidence.

Acknowledgments


I thank Nilanjan Ganguly, Chief Editor of the CSEG Recorder, for the encouragement to complete this article and for inviting the expansion of a technical blog into a full-length manuscript. The seismic attribute calculation and AI-based seismic interpretation capabilities described here build on R&D work completed by my colleagues at SubsurfaceAI Inc., including Dr. Donghai Zhuang, Dr. Yong Shu, and Dr. Rongfeng Zhang. The reviewers’ comments and suggestions are also greatly appreciated and have helped improve the clarity and quality of the article.

References

Bahorich, M.S., and S.L. Farmer, 1995, 3-D seismic discontinuity for faults and stratigraphic features: The coherence cube: The Leading Edge, 14(10), 1053-1058, doi:10.1190/1.1437077.

Barnes, A.E., 1993, Instantaneous spectral bandwidth and dominant frequency with applications to seismic reflection data: Geophysics, 58(3), 419–428, doi:10.1190/1.1443425.

Bednar, J.B., 1998, Least squares dip and coherency attributes: The Leading Edge, 17(6), 775–778, doi:10.1190/1.1438051.

Castagna, J.P., and H.W. Swan, 1997, Principles of AVO crossplotting: The Leading Edge, 16(4), 337–344, doi:10.1190/1.1437626.

Chopra, S., and K.J. Marfurt, 2005, Seismic attributes – A historical perspective: Geophysics, 70(5), 3SO–28SO, doi:10.1190/1.2098670.

Chopra, S., and K.J. Marfurt, 2007, Curvature attribute applications to 3D surface seismic data: The Leading Edge, 26(4), 404–414, doi:10.1190/1.2723201.

Gersztenkorn, A., and K.J. Marfurt, 1999, Eigenstructure-based coherence computations as an aid to 3-D structural and stratigraphic mapping: Geophysics, 64(5), 1468–1479, doi:10.1190/1.1444651.

Hampson, D.P., J.S. Schuelke, and J.A. Quirein, 2001, Use of multiattribute transforms to predict log properties from seismic data: Geophysics, 66(1), 220–236, doi:10.1190/1.1444899.

Kaur, H., N. Pham, S. Fomel, Z. Geng, L. Decker, B. Gremillion, M. Jervis, R. Abma, and S. Gao, 2023, A deep learning framework for seismic facies classification: Interpretation, 11(1), T107–T116, doi:10.1190/INT-2022-0048.1.

Marfurt, K.J., R.L. Kirlin, S.L. Farmer, and M.S. Bahorich, 1998a, 3-D seismic attributes using a semblance-based coherency algorithm: Geophysics, 63(4), 1150–1165, doi:10.1190/1.1444415.

Marfurt, K.J., R.M. Scheet, J.A. Sharp, and M.G. Harper, 1998b, Suppression of the acquisition footprint for seismic sequence attribute mapping: Geophysics, 63(3), 1024–1035, doi:10.1190/1.1444380.

Partyka, G., J. Gridley, and J. Lopez, 1999, Interpretational applications of spectral decomposition in reservoir characterization: The Leading Edge, 18(3), 353–360, doi:10.1190/1.1438295.

Shuey, R.T., 1985, A simplification of the Zoeppritz equations: Geophysics, 50(4), 609–614, doi:10.1190/1.1441936.

Taner, M.T., F. Koehler, and R.E. Sheriff, 1979, Complex seismic trace analysis: Geophysics, 44(6), 1041–1063, doi:10.1190/1.1440994.

Waldeland, A.U., A.C. Jensen, L.-J. Gelius, and A.H.S. Solberg, 2018, Convolutional neural networks for automated seismic interpretation: The Leading Edge, 37(7), 529–537, doi:10.1190/tle37070529.1.

Wu, X., Y. Shi, S. Fomel, and L. Liang, 2018, Convolutional neural networks for fault interpretation in seismic images: SEG Technical Program Expanded Abstracts 2018, 1946–1950, doi:10.1190/segam2018-2995341.1.

About the Author

Dr. Renjun Wen is the founder and CEO of subsurfaceAI Inc. (formerly Geomodeling Technology Corp.), an international geoscience software company headquartered in Calgary, Canada.

He completed his Ph.D. in Petroleum Geology from the Norwegian University of Science and Technology (NTNU) in 1995. Following his doctoral studies, Dr. Wen worked as a contract researcher at the Statoil Research Center. In 1996, he founded Geomodeling Technology Corp. in Calgary, building on his research experience at NTNU and Statoil.

Dr. Wen has published extensively in the fields of reservoir modeling, sub-seismic fault modeling, geostatistical image processing, seismic attribute analysis, and, more recently, AI applications in geoscience. He played a leading role in the development of innovative reservoir modeling technologies through the SBED Joint Industry Project (2000–2010), a collaborative effort involving nine major international oil companies.

He currently serves on the editorial board of Petroleum Geoscience (EAGE) and is a registered Professional Geologist with APEGA (Association of Professional Engineers and Geoscientists of Alberta).

An interview about his professional career was published by the RECORDER in September 2021.