Seismic imaging is one of the most important tools for exploring the Earth’s subsurface in geophysics. It’s how geoscientists create pictures of underground rock layers using sound waves (seismic waves). A critical piece of this process is the velocity model – a model that describes how fast seismic waves travel through different rocks underground. An accurate velocity model is essential for making clear seismic images; in fact, velocity models are the key to high-resolution techniques like seismic migration (similar to focusing a blurry picture). Without a good velocity model, the subsurface image can be distorted or unclear, much like a poor lens distorts a photograph. This article will explain what velocity models are and why they matter, how geophysicists traditionally build these models from raw shot gathers (the initial seismic recordings), and how machine learning (ML) is revolutionizing the process. We will cover key ML techniques, the advantages and challenges of using artificial intelligence in this field, and some real-world examples. The goal is to explain these concepts in simple terms for a general audience.
What Is a Velocity Model and Why Is It Important?
A velocity model in geophysics is essentially a map of the subsurface that tells us the speed at which seismic waves travel through each location underground. Different rock types transmit seismic waves at different speeds (for example, waves travel faster in solid granite than in loose sand). When geoscientists perform a seismic survey, they send sound waves into the ground (by using a controlled source like explosives or a vibrating truck) and record the echoes that bounce back from rock layers. The velocity model helps translate the time it takes for those echoes to return into actual depths and positions of the rock layers. In other words, if we know how fast the waves were going, we can figure out how far down a reflecting layer is, much like knowing the speed of sound in air helps a radar measure how far away an object is.
The importance of a good velocity model cannot be overstated. It directly affects the quality of seismic imaging. If the velocity model is wrong, the resulting image of the subsurface will be unfocused or misplaced. Think of it like needing the correct prescription in a pair of glasses to see clearly – the velocity model provides the correct “focus” for seismic data. Accurate velocity models are prerequisites for advanced imaging methods such as reverse-time migration, which produce high-resolution images of subsurface structures. In oil and gas exploration, for example, a reliable velocity model can make the difference between accurately locating a potential reservoir or missing it entirely. It’s also vital in other applications like earthquake seismology and underground construction, where knowing how waves propagate through the Earth can inform safety and design.
Raw Shot Gathers and Seismic Data Acquisition
Before we get into building velocity models, let’s clarify what raw shot gathers are. When a seismic survey is conducted, a seismic source (like dynamite or a specialized vibroseis truck) generates vibrations (seismic waves) that travel into the ground. These waves reflect and refract off different layers of rock and are picked up by sensors on the surface called geophones (on land) or hydrophones (in marine surveys). Each sensor records a time series of the ground shaking – essentially a little squiggly line called a seismogram. A shot gather is the collection of all these seismograms recorded from a single “shot” (one activation of the seismic source). In a shot gather, the horizontal axis might represent the distance of each sensor from the source, and the vertical axis is time (showing when echoes arrive at each sensor). The recorded signals include various events: the initial direct wave that goes along the surface, reflections from subsurface layers, refractions, and possibly some random noise.
In their raw form, shot gathers look like a series of wiggling traces (one trace per sensor) aligned by the shot time. This is the raw data – unprocessed, full of information about the subsurface. However, it’s not immediately interpretable as an image. Geophysicists apply many processing steps to turn these raw wiggles into a meaningful picture of underground structures. One of the crucial steps in processing is using the right velocity model to correct for the time it takes events to travel. For instance, if a reflector is deep, the echo comes back later; if the velocity is higher, the same travel time would mean a deeper reflector. So, the velocity model is used to convert these time measurements into depth and to line up reflections properly when creating an image.
To give an intuitive example, imagine shouting in a canyon and listening for the echo. If you know the speed of sound in air, you can estimate how far away the cliff is by timing the echo. If your assumed speed of sound is wrong, your distance estimate will be wrong. Similarly, in seismic imaging, the velocity model provides the “speed of sound” in each rock layer so that we can correctly map echo times to distances (depths). A raw shot gather is essentially the recorded echoes from one shout (seismic shot), and the velocity model is needed to interpret those echoes correctly.
Traditional Methods of Velocity Model Building
Building a velocity model from seismic data has traditionally been a complex, multi-step process. For many decades, geophysicists have relied on a combination of techniques and a lot of human expertise to derive velocity models from raw data. Here are some traditional methods and their characteristics:
- Manual Velocity Analysis (Normal Moveout Analysis): In earlier days, a common approach involved something called normal moveout (NMO) velocity analysis. Geophysicists would sort seismic data into gathers where signals reflect off the same point (common midpoint gathers) and then adjust a trial velocity value to “flatten” the reflection events. Essentially, they’d apply different velocity corrections to see which one makes the reflected signals line up straight in time. The best-fitting velocity becomes part of the model. This was often done by looking at a display called a velocity spectrum or semblance, which shows how coherent the reflections are for different velocity choices. Traditional velocity analysis required manually picking the best velocity peaks on these spectra. An expert would examine plots and choose velocities that make the data look most consistent. While effective, this manual picking is labor-intensive and somewhat subjective. As datasets grew in size (modern surveys can record thousands of shots and millions of traces), manual picking became a huge burden. It’s like trying to focus thousands of slightly blurry photos by hand – feasible on a few, but daunting at scale.
- Tomographic Inversion (Seismic Tomography): To draw an analogy, seismic tomography is similar to medical CT scans, but for the Earth. In travel-time tomography, one uses the arrival times of seismic waves (especially refracted waves or first arrivals and reflected arrivals) to infer the velocity structure. Geophysicists pick the travel times of certain waves (like the first arrival of a seismic wave to each sensor) and use algorithms to adjust the velocity model so that the predicted travel times match the observed ones. This often involves solving many equations and is an iterative process. Tomography can handle large amounts of data and can yield a smooth velocity model of the subsurface. However, it typically assumes that changes from an initial model are small (a linearized approach) and may simplify geology (for example, it might have trouble with very complex structures like sharp contrasts). Picking the right travel times (and the correct matching of which layer caused which arrival) is tricky and usually requires human quality control. Tomography is also time-consuming and computationally intensive, and still requires a decent initial guess to start. If the initial model is too far off, tomography might converge to an incorrect solution.
- Full-Waveform Inversion (FWI): This is a more recent (and advanced) technique which is often described as a game-changer – when it works. Full-waveform inversion uses the entire shape of the seismic waves (not just picked travel times) to update the velocity model. Essentially, the computer simulates seismic wave propagation through an initial velocity model and compares the simulated data to the actual recorded data (the raw shot gathers). It then tweaks the velocity model to minimize the difference between simulated and real data, repeating this over and over until ideally the synthetic data matches the field data. FWI can, in theory, produce very high-resolution velocity models that capture fine details (it uses all the wiggles, not just a few picked points). However, FWI is notoriously computationally expensive and sensitive to initial conditions. It requires massive computing power (especially for 3D surveys) and a reasonably good starting velocity model to avoid converging on a wrong solution (a problem known as “cycle skipping” in FWI). Like tomography, FWI might struggle in areas where the data lacks very low-frequency information or where there’s complex geology like salt bodies – it can get trapped in a wrong answer if not carefully guided. Running FWI for a large 3D survey could take days or weeks on a supercomputer. Because of these challenges, FWI is often used in combination with other methods (for example, using tomography first to get a coarse model, then FWI for fine details).
Each of these traditional methods has limitations. Manual velocity picking is slow and subjective, and it doesn’t capture complex velocity variations well because it usually yields a layered (piecewise) model adequate for stacking data, not detailed complexities. Tomography and FWI are powerful but time-consuming and computationally expensive, and they rely heavily on human expertise and quality control steps. Moreover, traditional methods often involve simplifying assumptions (e.g., considering the earth in layers or starting with a smooth model). This means they might miss or misrepresent complex geological features like abrupt velocity changes or irregular bodies (imagine a pocket of gas or a salt dome). For example, seismic waves might bend around a high-speed salt body, and a linear tomography approach could struggle to place that salt body correctly. In summary, building a velocity model the old way can be a bit like assembling a puzzle by hand: doable with skill and time, but very challenging if the puzzle is huge and complex.
How Machine Learning Is Revolutionizing the Process
The advent of machine learning (a branch of artificial intelligence) has introduced a completely new paradigm in velocity model building. Instead of laboriously picking velocities or running heavy physics-based simulations for every new survey, what if a computer could learn the relationship between the raw seismic data and the velocity model? Machine learning, especially deep learning, excels at finding patterns in large datasets without explicit human instructions. In recent years, researchers have begun training deep learning models to interpret seismic data in much the same way that these models have been taught to recognize images or understand language.
Imagine you have many examples of seismic shot gathers (the input) and the correct velocity models that produced those gathers (the output). A machine learning algorithm can be trained on this pairing: it will adjust itself (its internal parameters) until its predictions of a velocity model from a given shot gather match the known correct models in the training set. Essentially, the ML model learns to mimic the traditional inversion process by example. Once trained, you can feed it new raw shot gathers it hasn’t seen before, and it will predict a velocity model for that new data all on its own.
One pioneering approach used a deep fully convolutional neural network – a type of deep learning model commonly used for image analysis – to directly map from raw seismic data (even unprocessed shot gathers) to the subsurface velocity model. This is a drastic departure from conventional methods: it treats velocity model building as a pattern recognition problem rather than a classical physics problem. The neural network doesn’t rely on simplifying assumptions about geology; instead, it relies on the relationships it learned from data. During training, the network effectively “observes” many scenarios of wave propagation and builds an internal representation of how seismic signals correspond to velocity structures. In the prediction stage, that trained network can produce a velocity model in a matter of seconds, which is astonishing considering that FWI might have taken weeks on the same task.
To put it simply, machine learning is changing velocity model building from a hand-crafted, iterative physics experiment into a data-driven, automated prediction task. This revolution means that a process which used to require significant manual effort and computation can be accelerated. Some researchers have described this as moving from relying on “prior-knowledge assumptions” (in physics methods) to relying on “big-data training” (in ML methods). The machine finds its own way to explain the data, often discovering subtle features that a human might overlook. Another way to think of it: instead of a human expert spending hours per survey to tweak velocities, we spend that time up front teaching a computer using many examples, and then the computer can apply that knowledge much faster on new data.
It’s worth noting that the machine learning approach doesn’t eliminate physics – the training data often comes from simulations based on physics, and the patterns it learns are rooted in the physical behavior of seismic waves. However, ML can capture highly non-linear relationships inherently, whereas traditional methods might linearize or simplify the problem. For instance, a deep network might learn the tell-tale sign in the data that indicates a hidden low-velocity pocket, something a simplistic tomography might smear out. During training, the network automatically extracts multi-layered features from the seismic data without human guidance, and it doesn’t require an initial velocity guess to start the process. This means it can potentially handle situations where starting from scratch would stump conventional methods.
Machine learning is also revolutionizing the workflow by potentially enabling near real-time updates. There are scenarios like geophysical monitoring (e.g., checking CO₂ injection in a reservoir or detecting changes in a volcano) where being able to quickly update a velocity model from new shot data is extremely valuable. Once an ML model is trained, producing a velocity model for new data is so fast that it opens the door to real-time inversion that was previously impractical.
Key Machine Learning Techniques Used in Velocity Model Building
Several machine learning techniques and strategies are being explored to build velocity models from seismic data:
- Deep Convolutional Neural Networks (CNNs): CNNs are a class of neural networks especially good at processing grid-like data such as images. Seismic shot gathers can be viewed like an image (with axes of time and receiver position, and amplitude values like pixel intensity). CNNs slide small filters over the data to detect patterns (for example, a certain curving shape in the gather that might indicate a certain layer). In velocity model building, researchers have used fully convolutional networks (FCNs) and U-Net architectures (a type of CNN that works well for translating one image to another) to go from the seismic record directly to a velocity model image. These networks have many layers that progressively extract higher-level features from the raw seismic traces. Early layers might detect simple things like a wave arrival, and deeper layers might detect complex patterns like the interference of multiple reflections, all contributing to understanding the subsurface velocities.
- Supervised Learning with Synthetic Data: A common approach is to use supervised learning, meaning the network learns from examples where the “correct answer” is known. Since for real Earth we rarely know the exact velocity everywhere (unless we have well logs or previous studies), a lot of training is done on synthetic data. Scientists generate synthetic shot gathers using computers for random but realistic earth models (essentially solving the physics forward problem to create training pairs). Each synthetic gather comes with a perfectly known velocity model (because we generated it). These serve as training pairs for the ML model. The hope is that the network will generalize from these synthetic examples to real data. However, differences between synthetic and real data (noise, geology complexity, etc.) can be an issue – more on that in the challenges section.
- Transfer Learning and Pre-trained Models: In some cases, networks trained on one type of geology or region can be fine-tuned to work in another. This is analogous to how an image recognition model trained on millions of photos can be adapted to recognize specific kinds of objects with a bit more training. For seismic velocity models, researchers have tried to pre-train networks on large sets of synthetic models and then apply transfer learning to adjust the network with a smaller set of real examples from the target area. This approach helps to bridge the gap between synthetic and field data by giving the model a “baseline” understanding of seismic physics, then teaching it the specifics of the new region.
- Physics-Informed Neural Networks (PINNs) and Hybrid Methods: Another technique is to incorporate physics knowledge directly into the ML model. For example, a network might be designed to not only predict a velocity model but also to ensure that if you forward-model seismic waves through that predicted model, they match the observed data. This way, the ML model is guided by physical laws (the wave equation) as well as data. These are sometimes called physics-informed neural networks. Hybrid methods are emerging where machine learning does part of the job and physics-based inversion does the rest. One idea is using ML to come up with a good initial velocity model that can then be refined by a traditional method like FWI – essentially combining the speed of ML with the proven accuracy of physics-based refinement.
- Ensemble Models: Given the complexity of the problem, some researchers have even tried using an ensemble of neural networks – multiple networks that each learn a slightly different aspect, whose results are combined. This can sometimes yield more robust predictions (reducing the chance that one weird mistake by a single model will throw off the result). An ensemble might, for instance, include networks trained on different frequency bands of the data or different scales of velocity features, which together produce a final model.
To make these techniques concrete, let’s imagine an example: We have raw shot gather data from a survey in a geologically complex area. We feed this data into a trained CNN. The first layers might detect simple events like the first arrival (direct wave) and primary reflections. Deeper in the network, neurons might start detecting patterns corresponding to a certain type of layer sequence (say, a high-velocity layer over a low-velocity layer, which creates a characteristic pattern of refracted and reflected energy). The network’s final layers then assemble all these detected features and output a velocity value for each subsurface location, constructing a model. If the network was trained well, the output could look very much like what an expert would produce after a long manual inversion – but it appears with a single click. In practice, such a network can scan through huge volumes of seismic data and suggest velocity models that would have taken an army of geophysicists a lot of time to do by hand.
Advantages of Using Machine Learning in Velocity Model Building
Machine learning offers several compelling advantages over traditional velocity model building methods:
- Speed and Automation: Perhaps the most dramatic advantage is speed. Once a machine learning model (like a deep neural network) is trained, applying it to new seismic data is extremely fast – often a matter of seconds or minutes to produce a velocity model, compared to weeks of work using manual analysis or intensive computation with FWI. This speed enables quick turnaround for processing data and even the possibility of real-time seismic imaging in the field. Automation also means we can process many surveys or very large 3D datasets with minimal human intervention, freeing experts to focus on interpretation rather than number-crunching.
- Handles Complex Patterns: ML algorithms, especially deep learning models, are very good at capturing complex, non-linear relationships. Traditional methods might oversimplify (for example, tomography might assume waves travel in mostly straight lines or that the Earth is slowly varying). A trained ML model can, in theory, learn the subtle signatures of complex structures (like intricate fault networks or irregularly shaped bodies) directly from data. It does not impose a simple mathematical model upfront; instead it figures out the model that best explains the data. This means machine learning can potentially produce more accurate and higher-resolution velocity models in geologically complex areas than some conventional methods, which might miss those complexities. For instance, if there’s a tricky salt dome, an ML model might recognize the reflection/refraction patterns around it and adjust accordingly, whereas a human might need multiple iterations of trial and error to get it right.
- Reduced Human Bias and Labor: Because the process can be automated, the subjective element of human picking is reduced. Different interpreters might pick slightly different velocities in manual analysis, but an ML model will produce a consistent result given the same input. This consistency is valuable for large projects. Moreover, tasks that were tedious for people (like picking thousands of velocity points) are handled by the machine, which can tirelessly process data without fatigue. The human experts are still crucial (especially in training and validating the models), but their effort is shifted to supervising the ML rather than doing everything by hand. Over time, this could significantly cut down the cost and time of seismic data processing projects.
- Learns from “Big Data”: By training on large datasets (possibly compiling knowledge from many surveys or extensive simulations), machine learning can uncover statistical patterns and best practices that an individual human might not notice. For example, it might learn a sort of “average behavior” of seismic velocities in certain rock types by seeing them repeatedly in training data. This could make it more robust to noise: since it has seen so many examples, a bit of random noise in one gather won’t fool it easily; it will focus on the essential features that matter. In a sense, an ML model can carry the collective learning from a vast amount of data – something a single geophysicist would accumulate only over a long career of experience.
- Integration of Multi-Source Data: Although our focus is on raw shot gathers, it’s worth noting that ML models can be designed to take in various types of information together. For instance, one could feed both seismic data and well log data (direct measurements of velocity at certain points) into a network, and it could learn to integrate them. Traditional methods often did this in separate steps (e.g., using well data to constrain tomography manually). An ML model can blend them naturally if designed appropriately. This flexibility extends to combining different seismic attributes or even different geophysical methods (like gravity or magnetic data) to help build the velocity model. Machine learning provides a framework to fuse these data sources seamlessly by learning the correlations.
In summary, machine learning has the potential to make velocity model building faster, more detailed, and less dependent on tedious manual work. It’s bringing a level of efficiency and pattern recognition to seismic processing that mirrors what’s been seen in other fields like medical imaging and computer vision.
Challenges and Considerations of Using AI in this Field
While the promise of machine learning in velocity model building is exciting, there are important challenges and caveats to be aware of:
- Need for Quality Training Data: Machine learning models are only as good as the data used to train them. In seismic velocity modeling, getting quality training data is a big challenge. We usually need pairs of seismic data and corresponding correct velocity models. For real Earth data, the true velocity model is never perfectly known (we have some info from well logs or previous studies, but not everywhere). Thus, a lot of training relies on synthetic data (simulated examples). If the synthetic scenarios don’t cover the true complexity of the geology of interest, the ML model might not generalize well. Earth’s geology is immensely varied – a network trained mostly on, say, gentle sedimentary layers might struggle if given data from an area full of volcanic rocks. As one industry analysis put it, “due to the complexity of the earth, and the geological uniqueness of any one location, determining the appropriate training data can be challenging.”. Bridging the gap between synthetic training data and real field data is an active area of research. Techniques like data augmentation (adding noise, varying parameters) and transfer learning are used to make models more robust, but it remains a key hurdle.
- Generalization and Model Validity: Following from the data challenge, there is the issue of generalization. An ML model might perform amazingly on examples similar to its training data but fail on a new case that’s outside its learned experience. This is analogous to a student who memorized answers to specific questions but is stumped by a slightly different question on the exam. If an ML model is trained on relatively simple geology or on synthetic data that, for example, lacks certain real-world effects, it might give inaccurate results for a complex real survey. There have been cases where models trained on synthetic data gave blurry or incorrect velocity models on field data. Researchers found that training on some real data (if available) produced more plausible results. This means that for critical projects, one cannot blindly trust an out-of-the-box AI – some local training or calibration might be needed, and results should be verified.
- Interpretability and Trust: Machine learning models, especially deep neural networks, are often described as “black boxes.” They might have millions of parameters and it’s not straightforward to understand how they made a particular decision. In fields like geophysics, where decisions can have big financial and safety implications, practitioners are cautious. An ML-predicted velocity model might look reasonable, but if it disagrees with an established method, geophysicists will want to know why. Is it revealing a real geological feature that was missed, or is it hallucinating because of a weird quirk in the data? This lack of interpretability means that ML-derived models often undergo scrutiny and are sometimes used in tandem with traditional methods rather than outright replacing them. Building trust in AI results is an ongoing process. Techniques to interpret neural networks (like seeing which part of the seismic data influenced a given velocity prediction) are being developed, but it’s still not as transparent as, say, seeing that “I picked that travel time, which clearly corresponds to this layer.”
- Computational Resources for Training: While using a trained ML model is fast, training that model can be very resource-intensive. It might require generating thousands of synthetic training examples and running on specialized hardware (GPUs or TPUs) for days or weeks to learn the patterns. This is somewhat analogous to FWI in upfront cost – you invest the computation in training rather than in iterative inversion for each survey. If the ML model needs to be retrained for each new geological setting, the computational savings might diminish. Researchers are exploring ways to make training more efficient, but it’s a factor to consider: not every company or team has the massive computational resources that some cutting-edge projects use.
- Model Limitations and Accuracy: No method is perfect. ML models can sometimes output velocity models that are smooth averages and lack sharp detail if they haven’t learned to reproduce sharp boundaries (some early ML models produced overly smooth models compared to physics-based inversion which can sometimes get sharper contrasts). Conversely, a network might predict a sharp feature that is actually an artifact. The accuracy of ML methods still needs to match the reliability of well-established physics-based approaches in all cases. Currently, many ML approaches are validated on synthetic benchmarks or small case studies. Achieving the same level of universal trust as, say, travel-time tomography (which has decades of success) will take more validation. Some in industry believe that directly building a full 3D velocity model by machine learning still has some way to go. It might work in one scenario but need tweaking in another. So, ML in this field is powerful but not yet a push-button magic solution for all problems.
- Integration into Workflows: Introducing ML into traditional seismic processing workflows requires changes in software, training geophysicists to use and understand these tools, and ensuring compatibility with other steps. For example, if an ML velocity model is produced, it might need to be fed into existing migration software. Ensuring the ML output is in the right format and can be refined or adjusted by geophysicists (if needed) is important. There’s also the consideration of uncertainty estimation – traditional methods often allow experts to gauge uncertainty (like error bars from tomography). ML models don’t natively provide uncertainty, but researchers are looking at ways (like Monte Carlo dropout or ensemble models) to estimate confidence in the predictions. Knowing how sure or uncertain the model is can guide how it’s used (e.g., if the ML model is uncertain in one area, one might collect more data there or double-check with conventional methods).
In summary, while AI and machine learning are powerful new tools for velocity model building, they come with their own set of challenges. Successfully using ML in this field requires careful handling of training data, validation against known results, and often a combination of human expertise and AI – leveraging the strengths of both. It’s a learning process (quite literally) for the geophysics community as they integrate these technologies into practice.
Real-World Applications and Case Studies
The use of machine learning for velocity model building is relatively new, but it is rapidly gaining attention in both academia and industry. There are already some promising case studies and applications that showcase how this technology can be applied:
- Synthetic Benchmark Studies: Many early demonstrations have been on synthetic models that mimic real geology. For example, researchers often test their ML algorithms on famous synthetic models like the “Marmousi model” (a challenging, detailed model of subsurface geology used as a benchmark). In these tests, ML-predicted velocity models have been compared to those obtained by traditional inversion. Some studies showed that a well-trained neural network could recover velocity structures nearly as well as full-waveform inversion, but in a fraction of the computation time. These synthetic tests are important proofs-of-concept – they show the potential and help fine-tune the methods.
- Field Data Example – Gulf of Mexico: One published study applied deep learning to field seismic data from a geologically complex region in the Gulf of Mexico. The goal was to see if a network trained on a mix of field and synthetic data could recover the velocity model in an area with salt bodies and other complexities. The results were encouraging: the ML-derived velocity model made geological sense (for example, it identified the shape of a salt dome reasonably well) and, when used for imaging, produced migration images with clear features. Interestingly, this study found that training the model on some real data from the area (not just synthetics) was key to success – it improved the accuracy and geologic realism of the predictions.. This serves as a case study that ML can work on real seismic surveys, not just on toy problems, provided we carefully prepare the training process.
- Oil & Gas Industry Trials: Major energy companies and geophysical service firms (like Shell, BP, TGS, and others) have been investing in research on machine learning for seismic processing. For instance, TGS (a geoscience data company) reported on methods to automatically build velocity models by combining machine learning with traditional techniques. The idea in such industry trials is often to use ML to assist or accelerate parts of the workflow. One application is using ML for first-arrival picking (identifying the first seismic wave on each trace, which is a step needed for tomography) – a task that historically required tedious manual effort can be automated with neural networks. Another application is generating an initial velocity model that can be input to FWI, thereby shortening the FWI process. Companies have presented case studies in technical conferences (like SEG – Society of Exploration Geophysicists annual meeting) where they show an ML model built from one survey and then applied to another nearby survey to quickly produce a starting velocity model, which is then refined by conventional means. These real-world trials are demonstrating significant time savings. For example, instead of spending months in a velocity modeling cycle for a new seismic line, an ML approach might deliver a decent model in days, which can then be fine-tuned.
- Academic Research and Competitions: Academic groups are actively publishing research on new algorithms and even holding competitions. One example is an open competition where participants were given seismic data and tasked with predicting velocity models using AI, to spur innovation and compare approaches. Such efforts have led to new ideas like probabilistic neural networks that output not just one velocity model but a range of possible models with uncertainty estimates. In one complex case study from Colombia, researchers used a machine learning method to generate a prior velocity model for FWI. This prior helped the FWI converge better than it would have from a simplistic starting model, indicating that even when ML isn’t used alone, it can enhance traditional methods.
- Integration in Software Platforms: Sensing the trend, some geophysical software providers are integrating machine learning modules into their seismic processing software. This means that in the near future, a geophysicist interpreting data might have a button to “Run ML Velocity Analysis” which provides a quick velocity model suggestion. Early versions of these tools might be used in a advisory capacity – e.g. “here’s what the AI thinks the velocity model is; does that match your expectations?” – and the human can then adjust or accept parts of it. Over time, as confidence grows, the ML output might be used more directly.
It’s important to note that while there’s a lot of excitement, most real-world applications to date are hybrid – combining machine learning with human expertise and physics-based checks. For example, a company might use ML to get a first draft of the velocity model, then run a quick simulation to verify that the model explains the data, and finally perhaps do a limited fine-tuning with conventional inversion. This approach leverages the best of both worlds: ML for speed and pattern recognition, and physics for final accuracy verification. The case studies so far generally report that ML can dramatically reduce the time to reach a reasonable model, and in regions with prior training data, it can even improve the detail in the model. As more case studies emerge, especially from different geological settings (deserts, deepwater, fold belts, etc.), we will better understand how universally applicable these techniques are.
Conclusion
Machine learning is poised to significantly impact the field of seismic velocity model building, turning what used to be a laborious and time-consuming task into a faster, more automated, and potentially more accurate process. Velocity models are the unsung heroes of seismic imaging – without them, we can’t correctly interpret the echoes that reveal subsurface structures. Traditional methods of building these models, while effective over the past decades, have struggled with the increasing data volumes and geological complexity, often requiring heavy human involvement and computational resources.
The introduction of AI, and specifically machine learning, offers a new path. By learning directly from data, ML algorithms can infer velocity structures from raw shot gathers in ways that complement or even replace some of the classical approaches. Early results show that machine learning can drastically cut down processing time and help illuminate complex geological features that might be tricky for conventional methods. The use of deep learning techniques like convolutional neural networks allows the processing of seismic data in a holistic way, identifying patterns and relationships that would be hard to program manually.
However, the journey is just beginning. The success of machine learning in velocity model building depends on our ability to provide good training data, validate results, and integrate these tools into the geophysical workflow. Challenges like generalizing to new geology and maintaining trust in the results mean that, at least for now, machine learning is often used in tandem with traditional methods rather than as a complete replacement. We’re likely to see a phase where machine learning accelerates and enhances the process (doing the heavy lifting of data crunching and initial estimates), while human experts and physics-based models ensure the results make sense in the real world.
In conclusion, machine learning has begun to transform velocity model building from raw shot gathers, bringing us into an era where much of the seismic processing can be assisted by intelligent algorithms. This revolution is making it possible to go from field data to subsurface images faster and possibly with more detail than ever before. As the technology matures, we can expect more case studies demonstrating success and gradually more acceptance of AI-driven workflows in geophysics. The impact of machine learning on velocity model building is a prime example of how AI can take a complex scientific task and not only speed it up but also push its boundaries further, ultimately helping geoscientists better understand what lies beneath our feet. The future of seismic imaging will likely be a synergy of human insight and machine efficiency – a true collaboration between geophysicists and algorithms working together to explore the Earth.
