In this webinar, David Markus (Head of AI, dGB Earth Sciences) presents the latest advances in machine learning for seismic interpretation within OpendTect workflows.
The session covers how AI-driven methods enhance interpretation across the full subsurface workflow — from seismic preprocessing and denoising, to fault and horizon interpretation, inversion, reservoir characterization, and uncertainty quantification.
Topics include:
- Deep learning for denoising, super-resolution, and artifact removal
- AI-assisted fault detection and co-pilot workflows
- Automated seismic-to-well ties and wavelet estimation
- Machine learning for facies classification and attribute selection
- Deep learning-based inversion and rock property prediction
- Bayesian methods for uncertainty and scenario modeling
- Foundation models and agentic AI workflows in geoscience
The key message: AI is a powerful assistant — not a replacement for experienced interpreters. Human expertise remains central to geological plausibility, QC, and final decision-making.
#SeismicInterpretation #MachineLearningGeoscience #AIInGeophysics #OpendTect #SeismicInversion #ReservoirCharacterization #FaultDetectionSeismic #UncertaintyQuantification
Duration: 27:39
--- Transcript ---
Good morning, good afternoon, depending on where everybody is from. My name is David Markus, I'm the head of AI for dGB Earth Sciences,. And today I'm going to be presenting a webinar entitled Machine Learning Advances in Seismic Interpretation.
The idea of this webinar is to present exactly that, machine learning advances in seismic interpretation, a lot of which are already included in our software. And some of which are not, but we are actively working on, and it will be clear from the presentation which are which. Moving on, let me introduce who is dGB.
We are a privately owned geoscience and software services company founded in 1995. We've provided innovative seismic interpretation solutions to the oil and gas industry. Most people are probably familiar with us as being the company behind OpendTect, which is the world's only open source seismic interpretation system,.
And we also have many leading edge commercial plugins. It's an international company with offices in the Netherlands, USA, India, Malaysia,. And Vietnam, and represented by agents around the globe.
We have commercial users at national oil companies, majores and independents, research and development institutes,. And consultants around the world, and with that I bring us to the main topic of today's webinar, which I've mentioned before, which is machine learning advances in seismic interpretation. And in the fall of 2025, I presented this with our CEO, Nanne Hemstra, in Malaysia.
And so for all of the people who were unable to attend that webinar, sorry, that live presentation, I will present today's webinar. So AI and machine learning are really doing three main things in our industry, and across the whole workflow, they really are boiling down to these three items, which are automating repetitive tasks, such as fault picking, horizon tracking, attribute selection,. And so forth, learning complex non-linear mappings, going from seismic to impedance, elastic properties or directly to reservoir parameters,.
And making uncertainty explicit using Bayesian and probabilistic methods where we attach uncertainty to AI predictions. And this is really how interpreters like to think about risk anyway. But in almost all the recent work, the message is the same.
AI is a powerful assistant, not an autonomous interpreter, and us humans are still in charge of geological plausibility, QC,. And final decisions. Let's remind ourselves what the goals of seismic interpretation studies are.
One is to define basin architecture, improve the definition of targets, developing play concept analogs, advancing those play concepts in deep areas with little or no well control, validating rock physics as a predictive. And diagnostic tool, reduce risk in drilling program. Of course, this list is not exhaustive, but those are major goals of seismic interpretation studies,.
And they're all accomplished by addressing problems with seismic to well resolution and understanding better the amplitude rock properties relationships. So the typical seismic interpretation process that allows us to do this is, as we all know, starting from data intake. And QC through processing, well ties, wavelet extraction, time-depth relationships, building a structural framework, mapping faults.
And main horizons, extracting attributes, doing a stratigraphic interpretation, doing rock properties, quantitative interpretations, prospect reservoir characterization onto model building. And simulation links. Finally, looking at uncertainty and alternative scenarios.
And then there are future directions that we can also look at using new technology. So these workflows, of course, differ by context. And, you know, just looking at a few quick contrasts, exploration versus development.
Exploration has more emphasis on risking, regional frameworks, QI, where there's less well control. Development has much more dense well data, focus on well placement, problems like thin beds, Looking at environments, plastics versus carbonates, plastics have a channel lobe imaging, utilizing amplitude stratigraphy, carbonates would be looking at subtle impedance contrasts, things like harsh fracture buildups. And fractures and buildups and utilizing attributes and curvature for those sorts of things.
And then conventional versus unconventional. Unconventional focuses on natural fracture networks, brittleness, looking for sweet spots for landing zones and this sort of thing. So depending on what we're doing, where we're doing it and why we're doing it, we choose different workflows.
So looking at processing, including pseudo 3D in this, we can use machine learning to transform legacy 2D data into 3D post stack images. Using machine learning models and widely available post stack data, we aim for fidelity close to real 3D seismic. And we want to provide a detailed subsurface model that enhances structural clarity, improves targeting, reduces uncertainty.
And allows for efficient 3D survey planning. And can help us cost effectively influence economic decision making. Also in the initial processing, especially when we're going into a seismic interpretation system, we're looking at data intake in QC.
So things like using machine learning for denoising, footprint removal and resolution enhancement, we can use models like convolutional neural networks. UNETs are very popular in this regard and we train those to remove noise and potentially even super resolve seismic volumes. And recent work shows that these networks can sharpen thin beds and small faults while preserving amplitude fidelity with a clear uplift in interpretability.
And this AJAX CNN model was trained to do a series of post processing steps in order to enhance the visual quality. And interpretability of a seismic volume. This includes frequency dependent, structurally consistent noise attenuation and bandwidth enhancement.
And as the previous model was, it was developed in conjunction with Lundin Energy Norway's Geolab that probably should have been on the previous slide as well. But we have other models such as a horizontal multiple removal model that we can use to remove seabed multiples or other targeted multiples after flattening on a target horizon. We have a de-smile model as well developed in conjunction with Lundin and that is the removal of smiles that occur on seismic cross lines due to sparse sampling in the cross line direction.
And then these models are trained to remove these artifacts and other dipping noise. And so if I go back and forth between these pictures, we can see the difference between the multiple. And de-smile.
We can also use GANs, Generative Adversarial Networks, for super sampling. So improving deep learning fault interpretation by generating realistic synthetic seismic data using this generator discriminator model to produce high fidelity images. And then sharpening those.
And this increases the accuracy of our fault prediction in complex data sets. OK, now moving on to band of extension or super resolution. For seismic data, we can predict high frequency band of extended images from lower frequency, lower resolution images.
Of course, we want to prevent things like hallucination, which are high res artifacts are unrelated to the true seismic, the true geology. And these hallucinations can especially happen if you try to use models that were pre-trained in another region. Let's say you pre-train something in the North Sea and try to apply it directly to something in the Malay Basin.
You run the risk of hallucinating geology that doesn't exist. That's not to say that you couldn't use those models, pre-trained models as initial starting points. And then use the new data to retrain them and improve the weights so that you got better results faster.
But that's something you need to test, whether you start from scratch building a model, depending on the data that you have. And it all depends on the amount of data that you have going forwards. And so, you know, at the end of the day, what we want to do is we want to learn a mapping that predicts this bandwidth extended patch from patches anywhere in the volume.
Then when we move on to well ties, wavelets and time depth relationships, there are workflows in machine learning that allow for automatic wavelet estimation. We do not have those currently inside of OpendTect, but it's an active area of interest. And development.
And so various authors have used neural nets, multilayer perceptrons to train and estimate zero-phase wavelets that can. Then be used to generate synthetics and match seismic better than conventional well logs or well-based estimates. There are completely automated seismic well ties.
The recent work by Li et al used cascaded machine learning based matching strategies to automate this seismic to well tie step, including pre-processing of logs, computing reflectivity, generating synthetics,. And then using AI to optimize the time shift and alignment between synthetic and seismic trace. This significantly reduces manual tweaking and gives more consistent ties across many wells.
Potential next steps would be to utilize learn time depth trends. With enough wells, models can learn local time depth trends and help flag wells whose check shot VSP is inconsistent with the seismic response, act as a QC layer for the velocity model, or even identify ranges of fluid sensitivity. And AVO reflectivity.
And then, of course, we can augment that with AI driven wavelet estimation and auto matching. So again, we're automating the classic wavelet, synthetic and tie loop, and we even interpret a focus on rejecting. And accepting ties rather than dialing every parameter by hand.
We can also use machine learning for structural framework interpretation, faults and horizons, fault detection. And segmentation, again, using convolutional neural networks to come up with faster and more consistent fault extractions than manual or attribute paste picking. And this changes the structural workflow in such a way that we can pre-compute fault volumes by deep learning models.
We can auto track horizons and confidence maps and then again, interpreters spend more time validating, merging. And editing machine learning output rather than hunting faults from scratch. In that regard, we can describe a workflow that we're currently working on in OpendTect, which we call a copilot fault predictor.
And this is a machine learning model that's trained on synthetic seismic data. The train model can be applied to 2D and 3D data to generate fault probabilities, and. Then you can tune that to a data set used in transfer learning using interpret examples with data augmentation.
The whole idea here is that we want to pick like a human does and rather than have one machine learning model that extracts tens of thousands of potential faults. What we want to do is we want to pick some major faults and then have the machine learning model learn that these are the faults that we want to pick. And it will then help us pick similar faults on this line and potentially on the next line.
And then help us join those together. And so this workflow is a copilot assisted workflow in that we're working together to come up with a solution on just the faults that we're interested in rather than every particular fault or discontinuity that might exist in the seismic,. But actually doesn't add to our information.
Using machine learning in seismic attributes and stratigraphic interpretation, attribute selection driven by machine learning, of course, it's been around for a very long time. It's probably one of the earliest uses of machine learning is still as valid today as it was. Then.
And so instead of manually testing amplitude, frequency, and curvature, algorithms can evaluate which attributes matter for facies classification,. And this can significantly improve facies classification performance. And it has practical impacts on stratigraphic interpretation in that we can propose these candidates via clustering or segmentation.
We can provide probability match for certain facies and then we can let interpreters interactively adjust labels. And retrain these lightweight models on the fly. An example of something like this kind of quick look model would be a chimney cube where we have a specialized 3D seismic data attribute generated model.
And we can highlight anomalies, in this case chimneys, which are linked to gas filled zones. And fluid migration, which could appear as chaotic or pipe like in standard seismic views. And the idea is to map these pathways of fluid migration from deep source rocks to shallow reservoirs.
And it's a critical tool for petroleexploration and de-risking prospects, especially in the presence of hydrocarbons. And looking at things like charge and seal analysis and also identifying potential shallow gas accumulations, which can pose drilling hazards. Moving on to rock properties and quantitative interpretation.
Simply, bad wells equals bad results. So we want to use machine learning to do a lot of things with our input data, cleaning. And flagging spikes, depth shifts, identifying segments, using machine learning based outlier anomaly detection, training these models on good intervals.
And in hoping that we can predict the things in bad intervals. We want to impute data where we use the train model to predict missing sections and entire missing curves in new wells. We do want to make sure that we respect the geology and so that we still follow real vertical trends, facies breaks,.
And we want to be able to quantify trust in these models. Taking these things forward in terms of rock properties and quantitative interpretation, we can do a deep learning based seismic inversion where we can learn direct mappings from seismic to rock properties. And things like acoustic and elastic impedance.
And it turns out that these things can actually outperform classical deterministic inversion under noise and non-linearity. And we want to always make sure that we're constraining everything by physics. And so there are new AVO inversion approaches that embed the Zoeppritz linearized AVO equations directly into the neural networks.
These improve stability and adhere to physical laws and they yield better elastic parameter estimates for reservoir prediction. Here's an example of some kinds of results for prospect reservoir characterization using just such kinds of workflows. And these deep learning based workflows tend to bypass classical inversion pitfalls.
They learn direct mappings from the seismic and well logs, reservoir properties and facies. That is with the bad caveat that there is enough training data and careful validation. We can then also have AI derived property volumes like probability net sand, probability brittle shale,.
And we use these along traditional net to gross porosity maps when screening prospects and ranking drill locations. This brings us, of course, to model building and simulation. And all of these AI derived structural and property volumes are key priors / inputs to static.
And dynamic reservoir models. So everything that we've done so far is culminating in this model of the reservoir. And what we will then provide as input simulation.
And so we want to have implicit continuous models. We want to use Bayesian AI inversion, crystochastic modeling, and uncertainty quantification. We want to be able to do AI upscaling and property transfer.
And the bottom line is, is that workflows are moving from a single best impedance cube to ensemble based uncertainty of where reservoir models. And machine learning very much enables us to be able to do these things. That brings us, of course, to uncertainty and alternative scenarios.
And with all of these things, all kinds of scenario models, looking at simulations, varying all of these things, these help us get a handle on uncertainty. And we can use various machine learning methodologies to help us in these workflows using things like Bayesian deep learning. And variational inference.
Uncertainty fields for interpretation tasks where we can take all of the data that we have at input, for example, horizon fault networks, produce confidence cubes, machine learning based porosity, facies models, output probabilities rather than the hard labels. We combine this with classic or Bayesian facies classification and stochastic seismic conversion, and then we can support multi-scenario structural. And facies models.
And so AI machine learning has a deep impact on the workflow where this uncertainty now can be attached to every AI artifact, whether it's horizons, faults, impedance, facies, et cetera. And this makes it easier to build low, mid high cases rigorously instead of by ad hoc manual adjustment. And of course, I use the term easier in quotes.
All of these methods require our strict care and understanding of all of the data that went in, the limitations of the machine learning methods that we have applied. And factoring that in to our uncertainty and alternative scenarios. Now, I'll discuss future directions from a few minutes now, and these are kind of a new breed of AI machine learning models.
And so the three big ones are foundation models, which are big pre-trained seismic geologic models that you can quickly adapt tasks for fault, facies, inversion, these kinds of things. Large language models as orchestrators that help us plan workflows and call right models and tools for us,. And then knowledge capture agents, which turn our reports, wells, and projects into searchable living playbooks.
And these things exist today in our industry in various guises at various companies and various pieces of software that there's no real unification of all of these things at the present time. But this is again, a future direction and it is the goal that all of these things will come together. And aid our subsurface interpretation.
And so why would we want to bother with foundation models versus regular machine learning models? Because they're efficient. We can pre-train on unlabeled seismic and then we can fine tune on tens and hundreds of labels rather than thousands of labels.
They tend to generalize better across basin. And then you can have one backbone, one model that provides a lot of different tasks. So say, faulting, horizons, facies, even inversion can all be heads of the same model, simplifying deployment.
And maintenance. And these are the gateways then to multimodal earth models where we can aim to unify seismic logs, remote sensing, climate. And texual knowledge, all in one framework.
And of course, this is the dream and we have to be careful, of course, as I mentioned before, of things like hallucinations. And where knowledge that was learned in one place, which may not be applicable in another place, if you haven't fine-tuned well enough on your specific location, you may introduce issues from another location. So these models are very powerful, but with that power, you must exercise also caution.
And you must understand what you're putting in is going to affect what you're going to get out. And that's true of just about any science result. Looking at orchestration and knowledge, large language models.
These large language model-style foundation models for geoscience sit on top of everything. You can think of them as a front-end brain in natural language on top of the back-end seismic brain. So we can have interpretation co-pilots, which specialize on geoscience via domain pre-training, retrieval, augmented generation,.
And they can help us answer questions like, what does this AVO anomaly mean in this play time? Suggest workflows, parameter ranges, QC checks, and provide kind of a glue across formats where they can help to script processing flows, generate notebooks or reports, summarize interpretation results for decision makers,. And they can be valuable assistants that streamline these workflows and automate decision making.
There's just an example. I don't suggest anybody does this for their real work, but I simply fed the map on the right into chat GPT 5. 1.
I guess now we're under 5. 3 Already. And I asked it simply, what do the colors in this image represent?
And it came up with this description, which is deceivingly thorough and at the same time doesn't say much at all. And that's what you have to be very careful of in utilizing all of these kinds of models. They're designed to please you.
They will come up with a lot of information that could be useful. It will sound great. It will sound convincing.
But as the interpreter, we must be able to look through this chaff and find the real targets. Moving on to agentic, AI powered workflows, and this is simply we're now using machine learning models, as I mentioned before, as co-pilots or as smart junior geoscientists that can help us do our workflows. And they can help us plan, call up the right models, loop until we have a useful result while we stay in charge.
That's the idea. And so instead of a single black box model, we have a reasoning orchestration layer that understands the goals, such as QC, this new survey, map channels around well A. And so forth, breaks those into steps and then calls tools in the order that we need to call them to do what's next.
And of course, we always need guardrails and human checkpoints, so it never silently makes field decisions. And this is really an interesting new way of working. And it's probably the way that machine learning, AI, is really going to change the way we work at the office on a daily basis.
Finally, we can use all of these things, especially in terms of these agentic AI workflows, to capture knowledge. And that's probably the most important future aspect of any of these things, is to capture. And codify best practices, get this on-demand workflow guidance, build a living base.
And knowledge base that we can interrogate. And we can have this basin where decision support in real time, where it can help us surface relevant analogs. And typical failure modes while we're interpreting.
And so it can find information that's in the company knowledge base for this basin that maybe lost time, maybe that the person that was working on this field has retired. And with them the knowledge went out the door. And we have new staff now looking at new plays in the same basin, it could be 10s, 15s, 20 years later.
And we have this knowledge base that we can interrogate using these AI models to help us surface this information that will help us interpret better today. And while we're doing that, it's continuously updating this institutional memory. And so new projects and wells and interpretations automatically refine these workflows and base understandings.
And that's kind of really the big goal of these things moving forward. So the big question for us is can we trust this new breed of models? And the short answer is that they're powerful, but they can be very easy to over-trust, very hard to validate.
And very expensive to keep on the rails. Foundation models and agentic AI can massively accelerate the standardized interpretation and knowledge sharing, only if you treat them as fallible colleagues with strict quality assurance, governance. And physics-based checks.
And they should not be used as infallible replacements for experienced geoscientists. So to summarize how AI machine learning transforms seismic interpretation, I think we've seen through all of the different workflow bits that machine learning has accelerated that we have faster. And more scalable interpretation, bringing us interpreters to focus on geology scenarios and decisions, richer subsurface insights.
So we can learn these complex patterns between seismic logs and properties. Hopefully we have more consistent and objective results, standardizing workflows across assets and teams, better handling of uncertainty, making these mid-low, high cases more data-driven. And transparent.
And I think the main point that I've been trying to make is that human plus AI machine learning, first thing I don't really like the term AI, I use it interchangeably here,. But human plus AI is the best value where AI acts as a co-pilot and not an automated autonomous interpreter,. And we as geoscientists remain responsibility for geological plausibility, integration and final calls.
And with that, I will thank you. Thank you, SEAPEX, where I gave this presentation last time, and distinguished attendees,. And I thank all the people of this webinar who sat through it, and I will ask you to give questions,.
And we will try to provide thoughtful responses. And with that, I thank you very much, and I thank everybody at dGB for providing all of the work that went into making these models possible that we have inside of OpendTect,. And I wish you a good day.
Thank you so much.