NO20240365A1 - Deep learning inversion - Google Patents

Description

DEEP LEARNING INVERSION

Technical Field

There is provided a method of training an artificial intelligence (AI) algorithm, such as a neural network. In particular, there is provided a method training an AI algorithm (e.g. neural network) to perform applied geophysics for exploration, prospecting and / or engineering. There is also provided a method of performing a method of applied geophysics using an AI algorithm (e.g. neural network) trained in accordance with the training method and a computer program product.

Background

Applied geophysics refers to the practical application of geophysical methods and principles to investigate the Earth's sub-surface to solve real-world problems in fields such as exploration, prospecting and / or engineering. Often, applied geophysics involves forming a model of the Earth’s sub-surface. This model may be a model of one or more physical properties of Earth’s sub-surface. Such a model may be referred to herein as an “Earth model” or simply a model. The physical properties may include porosity, permeability, fluid saturation, an elastic property, P wave velocity, S wave velocity, density, Young’s modulus, Poisson’s ratio, mineral composition, grain size, texture, acoustic impedance, and elastic impedance.

Applied geophysics involves using one or more models of said physical properties to characterise a sub-surface region to solve a real-world problem. These problems include determining the suitability of at least a portion of the sub-surface region as a site for: resource exploration, exploitation and / or extraction, storage, and / or engineering projects more generally. For example, the one or more models of said physical properties can be used to assess and / or identify candidates for drilling for natural resource development and management, for example of hydrocarbons (e.g. oil and gas). Other resource-based examples could include mineral exploitation including coal and metals (e.g. to identify candidates for mining), or even for identifying sources of geothermal power (e.g. for use as a heat source such as for heating water) or groundwater. Storage site examples include identifying sites suitable for storage of carbon dioxide or nuclear materials. Examples of subsurface disciplines involved in identifying candidates for drilling are reservoir geophysics, reservoir engineering, reservoir characterization, resource assessment (volume estimations), drilling and production planning, reservoir management, risk assessment. Engineering examples of applied geophysics, in particular civil engineering, include optimizing the design of a structure or project, for instance its foundation and mooring, to accommodate sub-surface characteristics.

Engineering examples also include determining the suitability of a site for a project at all e.g. based on risk management or the conditions for mooring an onshore or offshore windmill based on subsurface conditions and geological hazards. Other problems could include analysing fault systems in a candidate sub-surface region and / or modelling the behaviour or likelihood of seismic activity in the candidate sub-surface region.

An important problem faced in applied geophysics is achieving the quantity and quality of data required to build accurate Earth models. Publicly available and / or proprietary data from wellbores do exist. These contain high resolution data. This data can include well log measurements of sonic, density, and neutron logs, which can be used to derive physical properties of the sub-surface such as porosity.

However, wells can be drilled far apart and well log measurements, core sampling and other data collection is only done for selected intervals in the wells. It is difficult to draw conclusions about a full area of interest based on the well data on its own. It is not practical or cost effective to measure sub-surface properties by drilling multiple well bores to collect the data for the model directly.

The use of geophysical surveys such as seismic surveys in applied geophysics can address the issue of lack of quantity of data (e.g. from well logs) because it is possible to cover a full (sub-surface) region (e.g. area or volume) of interest in a survey such as a seismic survey. A seismic survey may involve propagating seismic waves into Earth's interior from a plurality (e.g. an array) of seismic sources at the surface of the Earth or the sea. The seismic wave will propagate through the subsurface and reflect and refract at structural interfaces formed by geologic processes such as sedimentation, faulting and fracturing, and travel back to the surface where receivers such as geophones or hydrophones are located. In this way, a large area or volume of sub-surface space can be surveyed in a cost-effective and unobtrusive (non-invasive) manner. However, seismic data has low vertical resolution due to containing limited high-frequency information.

Geophysical inversion such as seismic inversion aims to estimate an earth model based on geophysical data such as seismic, and other available data and assumptions. It can address the issues of seismic data having low resolution and well (log) data having low quantity or coverage by combining the two sets of data. Seismic inversion is concerned with inferring the subsurface structure, stratigraphy, and properties, denoted as m, which includes but is not limited to elastic and petroelastic properties such as acoustic impedance, porosity, fluid content, and other geological and physical characteristics.

Methods of seismic inversion include convolutional methods applicable to poststack and pre-stack seismic, and solutions based on the wave-equation or eikonal equation for Velocity Model Building (VMB). One example of seismic inversion is the method of inverting for reservoir properties. The traditional approach first estimates elastic properties, such as acoustic impedance, and then leverages the elastic property estimates to infer reservoir properties such as porosity, lithology, and fluid saturation.

A seismic inverse problem can be mathematically described through the minimization of an objective function which often includes a least-squares data misfit and a regularization term. The elastic property inference is achieved by reverse-engineering the observed data dobs, which encompasses seismic data, welllog data (when available), and measurement errors e.

The relationship between the observed data, dobs, and the target subsurface properties, m, is established through a nonlinear forward modelling algorithm, F. This relationship can be expressed as:

(Equation 1)

Assuming e is small, we can derive an approximate, m∗, of the subsurface properties m by applying the inverse of the forward modelling algorithm F−1 to dobs:

(Equation 2)

The relationship between dobs and m is nonlinear and ill-posed. Consequently, one employs iterative or optimization methods that minimize an objective function J(m) to estimate subsurface properties.

(Equation 3)

Here, ∥ · ∥22 denotes the squared L2 norm. The first term represents the difference between observed data dobs and modelled data F(m), while the second term serves as a regularization component, controlled by the regularization coefficient µ, to guide the solutions toward physically plausible results. The reference model mref provides an initial estimate of subsurface properties. The aim of the traditional seismic inversion is to determine the optimal subsurface model m that minimizes the objective function J(m), with the reference model mref serving as a baseline or initial estimate of subsurface properties.

There are a number of limitations associated with the traditional seismic inversion, in which iterative or optimizations methods are used to estimate sub-surface properties. These include struggling to generalise leading to non-uniqueness and instability issues. Regularisation techniques, which integrate geological and geophysical data as prior information, can mitigate these issues to some extent. However, accuracy is still a problem. The more accurate an estimated model of a physical property can be, the higher the success rate when using that model to solve real-world problems (e.g. to make drilling decisions based on that model).

Inaccuracies can waste time, money and / or other resources. For example, each unsuccessful drilling attempt leads to significant wasted time and cost. Over the last two decades, the drilling success rate for exploration wells on the Norwegian Continental Shelf has been about 50%, which is considered high by international standards. Drilling a well on average costs about 75 million dollars. More accurate models would increase the success rate in oil and gas and other industries.

Furthermore, traditional seismic inversion techniques involve significant time and computation cost.

Above is a discussion of seismic inversion in which seismic and other data is converted to a model of one or more physical properties (e.g. acoustic impedance). It should be understood that similar inversion techniques are available for geophysical data that is not seismic data. This includes data from a magnetic, electrical, electromagnetic, gravity, or gradiometry survey. It should also be understood that inversion may be used to model parameters other than acoustic impedance. Inversion models and methods using different sources of data and for modelling different parameters have similar limitations to seismic inversion.

More recently, machine learning inversion, and specifically deep learning inversion (DLI) has emerged as an alternative to conventional inversion techniques. Unlike conventional inversion methods based on physical assumptions, deep-learning methods optimize a neural network to estimate the (physical) relationship between the input and output data.

In deep learning inversion (DLI) the inversion algorithm is a deep learning algorithm. An artificial intelligence (AI) algorithm, such as a neural network, is optimised to estimate a relationship between the input data (which may comprise one or more independent variables/predictors/features such as seismic data – e.g. one or more “seismic traces”) and one or more dependant variables (targets), typically physical parameters, to infer a model of such parameters (Earth model). DLI consists of an optimization phase called the training phase and a prediction phase.

In the training phase, the DLI algorithm optimizes one or more relationships between the features and the target(s). The target(s) are measured or modelled physical parameters. In particular, during training, the DLI algorithm receives the feature (e.g. seismic trace) and estimates a model of the physical parameter (e.g. estimates an “Earth model trace”). Training then comprises comparing the estimated model to a ground truth and optimising the algorithm (e.g. optimising weights of the algorithm) until the model of the physical parameter or Earth model estimated by the (optimised) algorithm is acceptable. The ground truth may be real measured data, such as measured data recorded in well logs.

The features may be inferred from available data such as geophysical surveys, well data, and other prior information and knowledge. Typical features for seismic inversion are seismic data and an initial model containing prior information. The neural network may receive the feature through an input channel. The feature may also be used as an output channel, for regularization purposes.

In the subsequent prediction phase, the optimized algorithm inverts from a new input (e.g. new seismic trace) to an earth model trace estimate. The DLI algorithm can in theory be optimized to predict any earth model property.

DLI has the advantage of addressing the problems of nonlinearity, non-uniqueness, and processing speed. Another benefit is that certain properties can be predicted directly from the input data without the need for the interim steps of traditional inversion. For example, the AI algorithm can predict porosity directly from seismic data without first estimating impedance.

Further improvements / optimisations of DLI techniques are needed for these techniques to be used accurately to solve the real-world problems described above, for example to reduce the risk of unsuccessful drilling attempts. Current models particularly suffer from inaccuracy when the data available for training is sparse. As described above, a typical source of training data might be from well logs.

However, well log data is generally sparse. There is a need to provide DLI techniques that are able to output accurate and reliable models even when well log data is sparse.

Summary

In view of the above, an object of the present invention is to overcome or at least mitigate drawbacks of prior art inversion-based applied geophysics methods and prior art machine learning methods.

In general terms, there is provided a method of training an artificial intelligence (AI) algorithm, such as a machine learning algorithm such as a neural network, to perform applied geophysics for solving real-world problems such as exploration, prospecting, and / or engineering. The method of the present disclosure comprises training the AI algorithm (e.g. neural network) using one or more discrete stratigraphic units. Stratigraphic units may be defined based on their depth in the well data. In other words, the stratigraphic units may advantageously be in the same domain as the ground truth. The method may be referred to herein as “stratigraphyguided deep learning” or SGDL. There is also provided a method of using an AI algorithm such as a neural network trained in accordance with said training method to perform applied geophysics to solve real-world problems such as exploration, prospecting, and / or engineering.

The inventors have found that basing stratigraphic units on well-data is advantageous because the stratigraphic units are therefore determined in the same domain as the ground truth (e.g. the depth domain). There is, therefore, no need to convert or correlate between different domains (time and depth) as part of the training process. The match between the time and the depth domain is given by the time-depth relationship which converts between the two domains. There is normally considerable uncertainty to the domain conversion and its accuracy. This can be seen, for example, when well data and a seismic survey are displayed together either in the travel-time or depth domain. Quantitative measures of the time-depth relationship for survey data such as seismic data and well data rarely show a correlation above 90%. By using well data to determine the stratigraphic unit feature, the model may advantageously learn from data that accurately represents the actual geological condition. If this were not the case, there may be discrepancies arising from a time-depth conversion and the algorithm might optimize for (learn) incorrect relationships.

Furthermore, the inventors have found that it is advantageous to train the AI algorithm with a feature which provides information about the discrete nature of stratigraphic units. In particular, the inventors have recognised that AI algorithms such as neural networks, for example convolutional neural networks (CNNs), typically excel at predicting continuous relationships but that sub-surface regions of the Earth consist of strata where each successive layer may have significantly differing characteristics. There may be abrupt changes in the parameter being modelled at the boundaries between strata which current DLI techniques struggle to model, particularly when the training dataset is small. The inventors have recognised that providing a variable of discrete stratigraphic units during training means that the network benefits from guidance related to abrupt changes of the target variable at the boundaries between strata. This helps the algorithm or network to optimize for (learn) what properties to expect in each stratigraphic unit. This significantly improves the accuracy of property predictions made by the network even when a relatively small dataset is used to train the AI algorithm or neural network.

The inventors have also found that the training and prediction are better when a plurality of stratigraphic units are input to one algorithm or neural network, in contrast to an approach of training a separate algorithm or network for each stratigraphic unit and then combining.

In some embodiments, the training method comprises identifying the discrete stratigraphic units from well data such as well tops. This is beneficial because well tops are picked based on the well log, i.e. in the same domain as the ground truth of the target variable. Furthermore, well data typically has a relatively high resolution (e.g. several samples per metre). This may be significantly higher than most geophysical data as measured from Earth’s surface (e.g. seismic data) which may have a resolution of one sample per 10 metres or less. Therefore, well data-based stratigraphic units may advantageously be precisely defined. As the skilled reader will appreciate, precise training data is crucial for training quality, and so for accurate predictions to be made by the trained algorithm.

In a first aspect, there is provided a method of training an artificial intelligence (AI) algorithm, such as a neural network, to perform applied geophysics for exploration, prospecting, and / or engineering. The method comprises receiving well data for a plurality of training wells. The method comprises, for each training well, identifying one or more discrete stratigraphic units based on the respective well data. The method comprises training a neural network with the or each discrete stratigraphic unit. The (trained) AI algorithm or neural network is arranged to perform applied geophysics by extracting a predicted sub-surface model or Earth Model of at least one physical property of a sub-surface region from one or more discrete stratigraphic units, for said sub-surface region. The advantages of using stratigraphic units in the training and / or of basing these stratigraphic units on well data has been discussed above. For the sake of completeness, the trained AI algorithm (e.g. neural network) inverts (i.e. infers) from (at least) the one or more stratigraphic units, to an Earth Model of at least one physical property.

As used herein, “applied geophysics” may refer to the study of Earth (in particular, sub-surface regions of the Earth such as the crust and / or near surface) to characterise a (sub-surface) region in order to solve a real-world problem, often with an economic aim in mind. The sub-surface region may be located onshore or offshore. Applied geophysics may be used for exploration, prospecting (or “geophysical prospecting”), and / or engineering problems related to a particular (sub-surface) region. This may mean that the predicted (sub-surface) model output by the trained AI algorithm or neural network can be used for these things. For example, applied geophysics (in particular, the predicted sub-surface model) can be used to assess and / or identify candidates for drilling for natural resource development and management of hydrocarbons. Other resource-based examples could include mineral exploitation including coal and metals (e.g. to identify candidate regions suitable for mining), or even for identifying sources of geothermal power (e.g. for use as a heat source such as for heating water) or groundwater. Exploration could also include identifying storage site examples. This might include identifying sites suitable for storage of carbon dioxide or nuclear materials as well as storage of other resources such as hydrocarbons (oil or gas). Engineering examples include optimizing the design of a structure or project to accommodate sub-surface properties. Engineering examples also include determining the suitability of a site for a project at all e.g. based on risk management. These problems could include analysing fault systems in a candidate sub-surface region and / or modelling the behaviour or likelihood of seismic activity in the candidate sub-surface region.

As used herein, “geophysical prospecting” may refer to the systematic process of determining and analysing the physical properties of the Earth’s sub-surface to identify and locate suitable sites for utilisation for an industrial purpose. In other words, geophysical prospecting may comprise identifying natural resources such as oil, gas, minerals or groundwater. Thus, geophysical prospecting may involve determining or otherwise identifying potential drilling candidates, for example. Alternatively, or additionally, geophysical prospecting may involve determining sites suitable for carbon capture and storage. For example, geophysical prospecting may involve determining or otherwise identifying candidate sites suitable as storage sites or reservoirs; injection well candidates; or requisition sites. In other words, geophysical prospecting may comprise identifying a candidate site for industrial utilisation or exploitation, in particular a drilling candidate and / or a storage candidate.

In some embodiments, the method is for training an AI algorithm (such as a neural network) to perform explorative geophysics or engineering geophysics. The field of applied geophysics may encompass exploration geophysics and / or engineering geophysics.

As used herein, a “stratigraphic unit” is a portion or region of the Earth that is distinguishable from the portions around it. A single discrete stratigraphic typically shares common characteristics in terms of lithostratigraphy (rock type), biostratigraphy (fossil content), chrono-stratigraphy (age of rock layers), seismic stratigraphy (stratigraphic and depositional facies), and / or sequence stratigraphy (depositional sequences). For example, a portion of Earth typically comprises rock layers or strata. Stratigraphy is the study of strata. Strata are classified into discrete stratigraphic units. While a single stratigraphic unit may be homogenous, adjacent stratigraphic units may be radically different from each other.

To formalize the mapping of stratigraphic units, practitioners use formal names for stratigraphic units that are mappable over large areas. The formal names are decided by the International Commission on Stratigraphy or by regional or national authorities, e.g. the Norwegian Offshore Directorate. There is a classification system of stratigraphic unit naming, where there are larger scale units and smaller scale units that form part of the larger scale units. For example, at the basin level, different units form part of a large-scale structural framework that might share the same era of deposition and that are named by the International Commission on Stratigraphy. On a smaller scale, e.g. at the reservoir level, units can be classified into fine-scale zones that share reservoir properties such as fluid flow parameters. Such fine-scale zones are named by the companies or practitioners working with the zones.

As used herein, “well data” refers to data associated with a well or borehole. In some examples, the well data used in the training is real measured field data from well logs. The well data may provide information of at least one of lithology (rock type), stratigraphy (layering), mineral composition, porosity, or fluid content. The well data may include depth information. The well log may comprise well top information. This well top information may be particularly useful for defining or identifying stratigraphic units. As described above, well data may advantageously have a relatively high resolution compared to other sources of geophysical data such as seismic surveys.

As used herein, “survey data” refers to data gathered as part of a survey such as a geophysical survey such as a seismic survey.

As used herein, “well tops” refer to specific depth markers in the well data. Well tops typically correspond to notable changes in lithology or other geological properties. In turn, these notable changes typically correspond to the boundaries of stratigraphic units. Thus, well tops can be used to form or identify discrete stratigraphic units for training the AI algorithm (e.g. neural network). Well tops may also be referred to as formation tops or marker beds.

Well logs (or borehole logs) may be based on real measurements made in a borehole. For example, well logging may be performed in boreholes drilled for hydrocarbon exploration or management (e.g. the oil and gas industry), groundwater and mineral and geothermal exploration. In some examples, the well data used in the training is synthetic data from one or more pseudo-wells. Synthetic training data can be generated using mathematical models, for example. For example, generating synthetic training data may comprise obtaining sample sub-surface models, conducting geophysical simulation on said models and generating the synthetic data. The synthetic data can be generated based on an acoustic wave equation, an elastic wave equation, coupled acoustic-elastic wave equations, Maxwell's equations, or potential-field equations, and the appropriate boundary conditions. In some examples, the algorithm is trained with only “real” measured data from well logs. In some examples, the algorithm is trained with only “synthetic” data. In some examples, the algorithm is trained with a combination of both real and synthetic data.

As used herein, a stratigraphic unit may have an upper boundary and a lower boundary. In some examples, the lower boundary of a first stratigraphic unit may also be the upper boundary of a second stratigraphic unit. An upper boundary of the first stratigraphic may also be the lower boundary of a second (or third) stratigraphic unit.

When the stratigraphic unit is identified based on well data, as is the case during training of the model, the upper boundary may be defined by a first feature from the well data, such as a first well top. Similarly, the lower boundary may be defined by a second feature from the well data, such as a second well top.

The boundaries between stratigraphic units may also be evident in survey data e.g. from a geophysical survey. For example, a high amplitude reflection in seismic data may be indicative of a boundary between stratigraphic units. Within each stratigraphic unit, the seismic data may show lower amplitudes and consistency in terms of frequency and other parameters. Stratigraphic boundaries identified in geophysical data may be referred to as horizons. As used herein, a horizon in geophysical data may be a feature identified or interpreted in geophysical data as being related to a boundary of a sub-surface rock formation. For example, as the skilled reader will understand, certain features in the geophysical data may be indicative of a sub-surface boundary of a rock formation. If the data is seismic data, then the horizon may be referred to as a seismic horizon. In other words, a seismic horizon may be a feature identified or interpreted in seismic data.

While the boundaries between stratigraphic units may be evident in geophysical survey data (e.g. as horizons), the inventors have found that it is preferable to use stratigraphic units based on well data, to train the AI algorithm (e.g. neural network). As above, this is because the feature is defined in the same domain as the ground truth target variable, overcoming the time-depth issue, and of the relatively higher resolution of well data compared to most geophysical data such as seismic data. Geophysical survey-based stratigraphic units may be used in the predictive phase, however. This is described in the second aspect below.

In some embodiments, the step of identifying the one or more discrete stratigraphic units comprises analysing the well data. If the well data comprises well top data, this may comprise identifying a plurality of individual well tops and / or assessing the suitability of the or each well tops for forming boundaries of one or more stratigraphic units. In some embodiments, these steps may be performed by a user. In other embodiments, these steps are performed as part of a computer-implemented method (e.g. are performed by a controller or processor). In such embodiments, the computer may receive the well data, process the well data, and output the stratigraphic units.

In some embodiments, the method comprises identifying a sequence of discrete stratigraphic units. In other words, the method may comprise identifying the specific order of the stratigraphic units (when there are a plurality of stratigraphic units). Different stratigraphic units may have a depth associated therewith. The order or sequence may refer to the relative order of the stratigraphic units by depth or laterally.

In some embodiments, the AI algorithm, such as neural network, receives the stratigraphic unit feature as an output channel for regularization purposes.

As the skilled reader will appreciate, a neural network may take multiple inputs. Thus, in some embodiments, the neural network is trained with further data (in addition to the discrete stratigraphic units).

In some embodiments, the method comprises receiving geophysical data for one or more of the training wells. The method may comprise training the neural network with the geophysical data (in addition to the well data). The geophysical data may comprise at least one of: seismic, time-lapse seismic, magnetic, electrical, electromagnetic, gravity, or gradiometry data. Types of seismic may be reflection or refraction seismic, onshore or offshore seismic, and higher and lower frequency seismic for shallow and deeper imaging. The seismic may also be post-stack, prestack, gathers, amplitude versus offset (AVO), amplitude variation with angle (AVA) or the unprocessed reflections received at the receivers. In such embodiments, the trained neural network may be arranged to extract the predicted sub-surface model or Earth model from the geophysical data and the discrete stratigraphic units. In other words, the trained neural network may take the geophysical data and the stratigraphic data as inputs and may output the predicted sub-surface model based on those inputs.

In some embodiments, the method comprises receiving time and / or depth data for one or more of the training wells. The method may comprise training the AI algorithm (e.g. neural network) with the time and / or depth data. In other words, the trained AI algorithm or neural network may take the time and / or depth data and the stratigraphic data as inputs and may output the predicted sub-surface model based on those inputs.

The inventors have found that the combination of geophysical data and / or time and / or depth data with the stratigraphic unit data is advantageous for optimizing the AI algorithm (e.g. neural network) more quickly and / or to arrive at a better trained / optimized algorithm or network.

In some embodiments, the well data (such as well top data) comprises data from one or more well logs. At least one of the well logs may have a resolution of at least one sample per meter (in depth), optionally at least 2 or 3 samples per meter.

In some embodiments, the method comprises receiving ground truth data of a physical property. This ground truth data may be taken from or comprise data from well logs (e.g. may be measured data) and / or may comprise data from pseudo wells or other a synthetic or simulated models or data). The ground truth property may also be referred to as the “target” for training or validating the algorithm. The ground truth data, and the location of the data that they are extracted from, may be located geographically close or distant from the candidate sub-surface region.

In some embodiments, training the AI algorithm (e.g. neural network) comprises reducing a difference between a sub-surface model of the physical property predicted by the algorithm or network and the ground truth. This may comprise reducing a loss function representing that difference.

In some embodiments, the trained AI algorithm (e.g. neural network) is arranged to extract a predicted sub-surface model of at least one physical property for determining or identifying a candidate drilling, resource storage or construction site.

In some embodiments, the trained AI algorithm (e.g. neural network) is (or may be called) a stratigraphy-guided deep learning inversion algorithm.

In some embodiments the physical property (of the predicted sub-surface model) is at least one of: an elastic property, P wave velocity, S wave velocity, density, Young’s modulus, Poisson’s ratio, mineral composition, grain size, texture, acoustic impedance, and elastic impedance, porosity, permeability, and fluid saturation.

In some embodiments, the AI algorithm (e.g. neural network) is arranged such that the predicted sub-surface model or Earth model comprises a plurality of physical properties of the sub-surface region. For example, the predicted sub-surface model may comprise a model of porosity and elastic impedance.

In some embodiments, the AI algorithm is a machine learning algorithm. An example of a machine learning algorithm is a neural network. The inventors have found that a convolutional neural network (CNN), in particular a temporal convolution neural network (TCN) may be particularly advantageously. The AI algorithm may comprise other AI architectures such as encoder-decoder networks, autoencoder networks, physical-guided neural networks, generative networks, and neural operators including physics-informed neural operators and Fourier neural operators.

Preferably, the neural network is a temporal neural network (TCN). The inventors have found that a TCN may be preferable because such networks excel in handling long data sequences and the data sequences involved in SGDL may be long (e.g. a seismic trace may comprise a large number of data points). In particular, a TCN may have a relatively large receptive field wherein the receptive field is the number of samples from the input that the network considers when predicting an output. Compared to classical CNNs, a TCN can achieve a relatively larger receptive field without having to be as deep. Other benefits of TCN is that it requires a low memory for the inference, parallelization, and stable gradient flow.

In some embodiments, the training may be described as supervised learning. In some embodiments the training may be described as unsupervised learning, hybrid learning, or transfer learning, or a combination of the above.

In some embodiments, the well data comprises data from 20 training wells or fewer, optionally 15 training wells or fewer, optionally 10 training wells or fewer, optionally 1 training well. This may be the case when the well data comprises or consists of measured data from real wellbores. As described previously, such measured well data may be relatively sparsely available. The sparsity of data has been a problem for current deep learning inversion methods. However, the inventors have recognised that, when using discrete stratigraphic units , the algorithm can be trained to determine accurate predicted sub-surface models or Earth models even with sparse availability of data (such as when the well data comprises fewer than 20 training wells etc.).

Of course, in some examples, the well data may comprise a larger data set. This may be because there is a large amount of measured well data available (e.g. from well logs) and / or because the well data comprises synthetic data from one or more pseudo-wells. In some embodiments, the well data comprises data from 200 training wells or more, optionally 2000 training wells or more, optionally 2,000,000 training wells or more.

According to a second aspect, there is provided a method of applied geophysics for exploration, prospecting, and / or engineering using an AI algorithm (e.g. neural network) trained in accordance with the method of the first aspect. The method of the second aspect comprises receiving one or more discrete stratigraphic units for a candidate sub-surface region. The method of the second aspect further comprises extracting, using the AI algorithm (e.g. neural network, a predicted sub-surface model of a physical property for the candidate sub-surface region for exploration, prospecting, and / or engineering.

In some embodiments, the method comprises receiving a sequence of stratigraphic units prior to extracting the predicted sub-surface model using the AI algorithm (e.g. neural network).

In some embodiments, the method may comprise receiving geophysical data, and / or time data and / or depth data, at the trained AI algorithm (e.g. neural network). This may be the case when the AI algorithm (e.g. neutral network) has been trained with additional features beyond just the discrete stratigraphic units – such as with geophysical data in the time or depth domain as described in relation to the first aspect. In such embodiments, the method may comprise correlating or otherwise tying the well-based stratigraphic units with the additional data (e.g. geophysical data in time or depth).

In embodiments in which geophysical data is received, the method may further comprise identifying the or each discrete stratigraphic unit based on that geophysical data. This may comprise identifying one or more horizons in the geophysical data. If the geophysical data comprises or consists of seismic data, then the step of identifying the one or more horizons in the geophysical data may comprise identifying one or more seismic horizons in the geophysical data.

In some embodiments, the step of identifying the one or more horizons comprises analysing the geophysical data. The method may comprise identifying points or areas in the geophysical data that represent presumed changes in geological layers. For example, this could comprise identifying a high amplitude reflection in a seismic image determining the presence of a stratigraphic boundary. In some embodiments, the identification of horizons may be performed by a user. In such examples, geological discontinuities may be identified by visual inspection of the geophysical data (perhaps when plotted as a depth vs geophysical property graph). In other embodiments, these steps are performed as part of a computer-implemented method (e.g. are performed by a controller or processor). In such embodiments, the computer may receive the geophysical data, process the geophysical data and output the horizons.

In some embodiments, the method comprises receiving (specific) horizon data such as seismic horizon data for the candidate sub-surface region. In some embodiments, identifying the or each discrete stratigraphic unit is based on the horizon data. In some embodiments there may be no need to include a step of identifying horizons in geophysical data.

In some embodiments, the step of identifying the horizon data may be performed by a user or by a computer as part of a computer-implemented method.

As described above, the inventors have found that stratigraphic units based on horizons (e.g. in geophysics data) may not be acceptable for training the neural network because of lack of accuracy. However, such stratigraphic units are acceptable in the prediction phase. In particular, using such stratigraphic units in the prediction phase means that predictions and models can be made for sub-surface regions outside of wells (i.e. regions for which well data is not available).

In some embodiments, the method comprises using the predicted subsurface model for exploration, prospecting, and / or engineering.

In some embodiments, the method comprises determining the suitability of the candidate sub-surface region as a drilling or storage site using the predicted subsurface model.

In some embodiments, the method comprises identifying one or more resources for exploitation such as identifying hydrocarbons, minerals including coal and metals, and / or groundwater.

In some embodiments, the method comprises changing or optimizing the design of a structure or engineering project to accommodate sub-surface properties.

In some embodiments, the method comprises determining the suitability of a site for an engineering project such as a construction project e.g. based on risk management. This may include identifying or analysing geotechnical properties, e.g electrical resistivity which can indicate the presence of stable or unstable clay, in a candidate sub-surface region and / or modelling the behaviour or likelihood of seismic activity in the candidate sub-surface region.

In some embodiments, the method comprises using the predicted subsurface model for at least one of the following:

1. Extracting layer porosity, lithologies and fluid saturations. In such examples, the well data and / or additional data used to train the neural network may comprise data from multiple log measurements.

2. Interpreting gas, oil and water volumes. In such examples, the well data and / or additional data used to train the neural network may comprise data from production logs.

3. Inferring reservoir permeability and extent. In such examples, the well data and / or additional data used to train the neural network may comprise pressure-transient data.

4. Mapping fluid fronts. In such examples, the well data and / or additional data used to train the neural network may comprise data from cross-well electromagnetic measurements; or

5. Integrating electromagnetic and seismic measurements for improved delineation of subsalt sediments.

6. Monitoring the location, concentration, and movement of CO2 underground at storage sites.

7. Modelling soil composition, sediment layers, and geological features for wind turbine site characterization, mooring design, risk assessment, and installation planning.

8. Geothermal energy exploration and production, such as resource assessment, reservoir management, drilling optimization, and environmental impact assessment.

9. Groundwater exploration and extraction, such as resource assessment, well siting and design, aquifer management, contaminant transport modeling, and risk assessment and decision support.

In a third aspect, there is provided a computer program to perform applied geophysics for exploration, prospecting, and / or engineering, the computer program product comprising computer-readable instructions which, when executed by a computer, performs the method of the first aspect or the second aspect.

Although different embodiments may be disclosed separately in the detailed description which follows, any feature of any embodiment may be combined with any other feature or combination of features of any embodiment. That is, all possible combinations and permutations of features disclosed in the present disclosure are envisaged.

Brief description of the drawings

Specific embodiments are described by way of example only with reference to the following figures:

Figure 1 is a diagram representing a prior art deep learning inversion (DLI) method;

Figure 2 is a flow diagram representing a training phase 200 of the prior art DLI method;

Figure 3A shows well log data for a first well in a first location;

Figure 3B shows seismic data for the first well in the first location;

Figure 4 is a diagram representing a training phase of a first DLI method according to the present invention for performing seismic inversion;

Figure 5 is a diagram representing a training phase of a second DLI method according to the present invention for performing seismic inversion;

Figure 6 is a diagram representing a training phase of a third DLI method according to the present invention for performing seismic inversion;

Figure 7 is a diagram representing the use of an algorithm trained as described in relation to Figure 4 in a prediction phase;

Figure 8 is a diagram representing the use of an algorithm trained as described in relation to Figure 5 in a prediction phase;

Figure 9 is a diagram representing the use of an algorithm trained as described in relation to Figure 6 in a prediction phase;

Figure 10 is a simplified schematic diagram of an example of the DLI algorithm used in examples according to the present invention;

Figure 11 schematically illustrates an applied geophysics unit according to the present invention;

Figure 12 illustrates a sub-surface region; and

Figure 13 illustrates a seismic image of a part of the sub-surface region of Figure 12.

Detailed description

According to embodiments of the present invention as disclosed herein, the abovementioned disadvantages of solutions according to prior art are eliminated or at least mitigated.

Figure 1 is a diagram representing a prior art deep learning inversion (DLI) method. Figure 1 can be said to represent a “prediction phase” 100 in which seismic inversion is performed. In this example, the DLI of Figure 1 is a method of seismic inversion.

The method comprises receiving seismic data 102 for a portion of a sub-surface region 1200 as illustrated in Figure 12. The seismic data 102 may also be referred to herein as a seismic trace.

Figure 12 is an illustration of the sub-surface region 1200 associated with the seismic data 102 of Figure 1. It shows an offshore cross section with the sea surface 1201, seabed 1202, and two rock boundaries 1203 and 1210. It shows two vertical wells 1204 and 1205 and a windmill 1208. The well 1204 illustrates an oil and gas well which penetrates a reservoir. The oil and gas has accumulated in a porous and permeable rock 1206. 1203 marks the boundary between the reservoir rock and an above cap rock, which seals off the reservoir, trapping the oil and gas. The well 1205 illustrates a CO2 injection well which injects CO2 that flows up to the saline aquifer 1207. 1203 marks the boundary to the above cap rock which seals off the reservoir, trapping the CO2. The windmill 1208 has a mooring 1209 below the seabed 1202. Below the rock boundary 1210 there is a reservoir 1211 which may contain oil and gas or which may be suitable for CO2 storage.

Figure 13 is an illustration of a seismic image 1300 from a seismic survey which has been conducted over the sub-surface region 1200 around the well 1204. The top of the seismic image is the sea surface 1301. A seismic horizon 1302 marks the seabed, that has been picked at the most shallow high amplitude reflector. Another seismic horizon 1303 has been picked based on the seismic amplitudes and coincides with the rock boundary 1203. 1305 marks the closest seismic trace to the well 1204.

Turning back to Figure 1, in this example, the seismic trace is a 1D extraction of a seismic reflection image, as illustrated by Figure 13. A seismic reflection image 1300 is a processed 2D or 3D representation of the subsurface based on the received seismic reflections. In this example, the seismic trace 102 has amplitude on an xaxis and time (in particular, two-way travel time) on a y axis. The two-way travel time is measured in milliseconds and represents the time it takes for a seismic wave that is generated at the Earth’s surface to propagate through Earth, be reflected by boundaries between rock layers of the sub-surface and for the reflection to be received at a sensor (again, at Earth’s surface).

The method then comprises taking the received seismic data 102 as input for a deep learning inversion algorithm (DLI algorithm) 104 and using that DLI algorithm to extract a predicted sub-surface model or Earth model 106 of at least one physical property of the sub-surface region. In this example, the Earth model 106 comprises a model of porosity. However, it should be understood that the DLI algorithm could be trained, given available data, to extract from the seismic data 102 an Earth model of any physical parameter of the sub-surface region.

Figure 2 is a flow diagram representing a training phase 200 of the prior art DLI method. In this example, the training phase 200 comprises use of a supervised learning technique. In the training phase, the DLI algorithm optimizes a relationship between the seismic and the Earth model.

In more detail, step 202 of the training phase 200 comprises receiving a first seismic trace which is extracted along the first (vertical or nonvertical) wellbore path for a first training well. Step 204 of the training phase 200 comprises inputting the seismic trace 202 into the DLI algorithm to predict an Earth model for the first wellbore path. Step 206 of the training phase 200 comprises outputting the Earth model as an Earth model trace that represents how the respective physical parameter (e.g. porosity) varies with depth/time. Step 208 comprises receiving ground truth data for the physical parameter. In this example, the ground truth data is porosity data from the well log for the first training well. The training well, from which the ground truth data 208 is taken, may be the well 1204 from the sub-surface region 1200. In that case, the seismic data 202 is extracted along the well 1204, from the seismic trace 1305 of the seismic image 1300.

Step 210 of the method comprises comparing the predicted Earth model trace to the ground truth data. If the Earth model trace is acceptable, the method moves to step 212 which is to stop optimizing the DLI algorithm. The predicted Earth model trace may be deemed acceptable if a loss function, representing the differences between the predicted Earth model trace and the ground truth data, meet a criterion. If the Earth model is not acceptable, step 214 is instead performed in which the DLI algorithm is optimised (e.g. by optimising the weights of the DLI algorithm). The method then repeats steps 204 onwards of the method, this time using the optimised DLI algorithm. The method repeats either until an acceptable match is found between the predicted Earth model trace and the ground truth data, a predetermined number of iterations have occurred, negligible improvement in model performance is observed across successive iterations, or an increase in validation loss suggests overfitting.

The training method is repeated with seismic traces from a plurality of different wells. For example, in the context of Figure 12, this might include well 1205 from the subsurface region 1200 in addition to the well 1204. Each well used in the training may be referred to as a training well. Validation wells that are used in training may also be referred to as a training well. The region around a training well which is used directly or indirectly in the training may be referred to as a training sub-surface region.

In this example, the prior art DLI algorithm is a deep neural network such as a convolutional neural network.

Neural networks excel at predicting continuous relationships when the features and target variable are continuous. One of the problems faced when performing and developing DLI techniques for earth model prediction is that the earth properties change abruptly, as adjacent rock layers (strata) may have radically different characteristics, akin to discontinuities in the target variable. This can be hard for the DLI to learn or predict without the features providing more explicit information about such discontinuities.

Strata can be classified into discrete stratigraphic units. Stratigraphic units are bodies of rock that are distinguishable from the units around them. A single stratigraphic unit typically shares common characteristics in terms of lithostratigraphy (rock type), biostratigraphy (fossil content), chronostratigraphy (age of rock layers), seismic stratigraphy (stratigraphic and depositional facies), and / or sequence stratigraphy (depositional sequences).

While the properties of a single stratigraphic unit can be homogenous, adjacent stratigraphic units can be radically different from each other. This can result in discontinuities at the boundaries between stratigraphic units in said physical properties. The presence of discontinuities in the physical property can pose a problem for the DLI, making it hard to learn the relationship between the features and the target, and eventually predict the properties correctly. In particular, when all the features are continuous the network will struggle to learn the discrete nature of the stratigraphic units. Improved property prediction can be achieved by providing information that helps the DLI algorithm learn the discrete nature.

Figure 3A and B show well log data and seismic data for a first well in a first location (e.g. well 1204 in Figure 12). Figure 3A is the well log data 302 which is acoustic impedance in this example. Acoustic impedance (in kg / (cm2 -s)) is on x axis and two-way travel time (in milliseconds) is on the y axis wherein two-way travel time can be related to depth through the velocity of the seismic waves. Figure 3B is the seismic data or trace 304 along the wellbore path. Amplitude is on the xaxis and two-way travel time (in milliseconds) is again on the y axis.

The boundaries between adjacent strata, as interpreted from the well log, are represented by the horizontal lines 303 in Figure 3A. As is evident in Figure 3A, sudden changes (i.e. discontinuities) are apparent at many of these boundaries.

These discontinuities can make learning and prediction hard for current DLI techniques that use continuous features. In this example, the horizontal lines 303 correspond to well tops or formation tops recorded in a well log for the first well. In Figure 3B, a number of seismic horizons 380 are identifiable in the seismic trace and are marked by dotted horizontal lines. Even though the well log data 302 and the seismic trace 304 are in this case converted to the same domain, it is evident from line 390 of Figure 3B that the well tops 303 and the seismic horizons 380 do not perfectly match. As described earlier, the well tops are normally a better match with the ground truth and therefore makes for a better stratigraphic unit feature in the training phase.

The stratigraphic unit feature may be based on well data as illustrated in table 1. Column 1 shows the depth (e.g vertical depth from the seasurface). Column 2 shows the name of the stratigraphic unit at each depth. The depth of the units may be derived from interpretations of well log properties, one of which are shown in column 3. The interpretations have indicated a Viking well top at depth 501 and a Hugin well top at depth 503. Based on the stratigraphic units, the stratigraphic unit feature in column 4 have been assigned the value 1 for the top unit, the value 2 for the middle unit, and the value 3 for the deepest unit. This categorical stratigraphic unit feature adds valuable context when training the AI algorithm. When there is more than one training well, which is usually the case, the same value of the stratigraphic unit feature is assigned to the same stratigraphic unit located in a different well. For example, a first stratigraphic unit may have the value 2 also in a different training well that is to be used in training for this sub-surface prediction case. If the first stratigraphic unit is not available in a training well, while second and third stratigraphic units are available, then the second and third stratigraphic units will still have the values 1 and 3 respectively. In one example, the first, second and third stratigraphic units may be the Cretaceous, Jurassic and Hugin units, respectively.

Table 1

The stratigraphic unit feature may also be a non-categorical feature, while still being based on the discrete stratigraphic units in the well data. The example in table 2 contains the same data and columns as table 1 but the stratigraphic unit feature is determined in a different way. In the table 2 example, each stratigraphic unit boundary has been assigned a value, and the stratigraphic unit feature values are interpolated between the unit boundaries. The Viking top boundary has been assigned the value 50, and the Hugin top boundary has been assigned the value 100. Interpolating between the Viking top value of 50 at depth 501 and the subsequent Hugin top value of 100 at depth 503 gives a value of 75 at depth 502. The values of the stratigraphic unit feature at depths 500 and 504 have been determined based on interpolation to boundaries at depths that are shallower and deeper than the values in the table, respectively the Cretaceous top and the Hugin bottom. When there are more than one training well, the same value will be assigned to the same boundary when located in another well. For example, the stratigraphic unit feature will have the value 50 assigned to the Viking top boundary also in another training well that is to be used in training for this sub-surface prediction case. The interpolation is done separately for each well, depending on the presence and distance between the unit boundaries in the well.

Table 2

The non-categorical stratigraphic unit feature captures the information of the stratigraphic unit boundaries in a similar way to a categorical feature. It may also carry the added benefit of capturing the property variations within stratigraphic units. The way of defining the stratigraphic unit feature as shown in the table 2 example can be compared to the concept of relative geological time, but it differs from conventional methods by the fact that it is determined based on the depth of stratigraphic unit boundaries in the well log, instead of being based on boundaries in geophysical survey data such as seismic horizons. The values at each stratigraphic boundary can be based on the geological age of the stratigraphic units. For the sake of simplifying the exemplification of the method, the following text is based on the categorical feature example.

Figure 4 is a diagram representing a deep learning inversion (DLI) method according to the present invention for performing seismic inversion. An important difference between the prior art DLI of Figure 1 and the DLI according to the present invention of Figure 4 is that the data that is input into the DLI algorithm is different. In particular, the method comprises receiving discrete stratigraphic units 452 as categorical data in addition to seismic data 402 for a first sub-surface region. In this example, the DLI algorithm is a neural network. As the skilled reader will appreciate, a neural network can receive multiple direct inputs.

As in the example of Figure 1, the seismic trace 402 has amplitude on the x-axis and time (in particular, two-way travel time) on the y axis. The two-way travel time is measured in milliseconds and represents the time it takes for a seismic wave that is generated at the Earth’s surface to propagate through Earth, be reflected by boundaries between rock layers of the sub-surface and for the reflection to be received at a sensor (again, at Earth’s surface).

The method then comprises taking the received seismic data 402 and the discrete stratigraphic units 452 as input for a deep learning inversion algorithm (DLI algorithm) 404 and using that DLI algorithm to estimate a sub-surface model or Earth model 406 of at least one physical property of the first sub-surface region. In this example, the Earth model 406 comprises a model of porosity. However, it should be understood that the DLI algorithm could be trained to extract any predicted physical parameter of the sub-surface region as a predicted earth model.

Figure 4 represents the training of the DLI algorithm. The training method in relation to Figure 4 is very similar to the training method described in relation to Figure 2. However, an important difference is that at step 202 the method comprises receiving the seismic data 402 and the discrete stratigraphic units 452. Furthermore, step 204 comprises inputting the seismic data 402 and the discrete stratigraphic units 452 into the DLI algorithm 404 to predict an Earth model for the path along the training wellbore.

Figure 4 shows a ground truth 408 for the physical parameter at the first sub-surface region. This ground truth 408 is used in the comparison step (step 212) of the method of Figure 2.

In the example subsurface region 1200, there are three notable stratigraphic units in the subsurface: The first between the subsea and the rock boundary 1203, the second between the rock boundaries 1203 and 1210, and the third below the rock boundary 1210. The well 1204 penetrates each stratigraphic unit. The well log of the well 1204 may be the well log 408. From the measured property 408, the stratigraphic units may be observable, and these might be marked by well tops (as illustrated in Figure 3A). The stratigraphic unit feature 452 can then be determined based on these well tops.

Table 3 below shows examples of values from the method. The features include depth values, stratigraphic unit values, and seismic amplitude values. For example, Table 3 shows that stratigraphic unit 1 is at time 500 and the seismic amplitude at this depth is -2. Table 3 also shows that stratigraphic unit 2 spans depth 501 and 502 and stratigraphic unit 3 spans time or depth 503 and 504.

Column 4 of table 3 labelled “Predicted Earth Model” shows example predicted values for the Earth Model (i.e. values of a parameter being modelled) at each depth. Column 5 of table 3, labelled “Ground Truth” shows example ground truth values for the parameter at each depth which, in this example, are taken from a well log. In the comparison step (step 212) the values of column 4 are compared to the values of column 5 to determine a difference (or loss function) therebetween.

Optimising the DLI algorithm (step 216) aims to reduce the differences between the values of column 4 and column 5. Training the algorithm involves the DLI algorithm learning that stratigraphic unit 2 typically has a property value of about 10, for example.

Table 3

In the example described above, seismic data 402 and discrete stratigraphic units 452 are input into the DLI algorithm in the training and prediction phase. However, it should be understood that the DLI algorithm can be trained or optimised to take various features. Figure 5 represents the training of the DLI algorithm using seismic data 402, discrete stratigraphic units 452 and time or depth data 454. Figure 6 represents the training of the DLI algorithm using only discrete stratigraphic units 452 (i.e. no time / depth data 454 and no seismic data 452). In other examples, some data other than seismic data 402 is used (in addition to the stratigraphic units 452) such as geophysical data like background velocity, magnetic, electrical, electromagnetic, or gravity data. In all cases, the DLI algorithm can be trained to output an Earth model of a physical property of the first sub-surface region.

However, it is often the case that more information can be provided to the DLI algorithm (by inputting a variety of different data types in addition to the stratigraphic units 452), so that the DLI algorithm is quicker or more accurately trained and / or with a smaller training data set. The inventors have found that it is particularly preferable for an input to include (in addition to stratigraphic units 452) at least one of seismic data and depth / time data.

Different types of seismic data may be needed as input for different property prediction tasks. Post-stack seismic data might be sufficient for an acoustic impedance prediction task. For reservoir property prediction like lithology or fluid it is often necessary to use pre-stack seismic and amplitude versus offset (AVO) or amplitude variation with angle (AVA) data. Furthermore, rock physics data or relationships might be needed as input to predict certain properties.

As further features, the inventors have identified background velocity, and initial property models to be advantageous. The features provide important information both on their own and in combination with each other. The background velocities can provide information about elastic properties at different depths. The initial property model can be populated with prior information and knowledge about the region of interest. The neural network excels at learning linear and nonlinear relationships between the features and the target.

The most important feature in the context of the present invention is that the wellbased discrete stratigraphic units, exemplified by the categorical feature 452, are included in the input. The inventors have recognised that providing the DLI algorithm with discrete stratigraphic units 452 as an input benefits the network by providing guidance related to the likely positions of discontinuities in the output model. This improves the accuracy of the algorithm and means that it can be trained more accurately and / or with smaller training sets while achieving an accurate output model. Consequently, the algorithm can achieve a robust, understandable, and geologically consistent output model.

An important aspect of the stratigraphic unit feature used in the training phase is that the stratigraphic units feature 452 are derived from well log data of the training well. Well log data generally includes information identifying well tops (or formation tops) as marking the top of stratigraphic units in the well. The stratigraphic unit feature used in training can be derived using these well tops as boundaries. An advantage of the stratigraphic units 452 being based on well top data from well logs is that they ae determined in the same domain as the ground truth well logs. Also, well logs typically have relatively high resolution and so the vertical precision of the well tops is relatively high, e.g. in comparison with geophysical survey data such as seismic data. The high accuracy of the stratigraphic unit feature makes it easier for the network to learn the relationship between the feature and the target.

In this example, in the training phase, the ground truth 408 is also well log data of a training well. An advantage of this is that the relationship between the well tops and the ground truth (and so the stratigraphic units 452 and the ground truth) is well defined in the well log.

In other examples, in the training phase, the stratigraphic units 452 and the ground truth 408 are derived from synthetic training data from pseudo-wells.

The representation of the stratigraphic units 452 in Figure 4 shows how different stratigraphic units have different discrete values at different depths, with the x-axis representing the value and the y-value representing depth. This way, the discrete stratigraphic units 452 are encoded for input in the DLI algorithm as a categorical input. In one example, this encoding comprises label encoding in which the shallowest unit is given the value 1, the next unit the value 2, and so on. In another example, one-hot encoding is used which treats each stratigraphic unit as a binary feature which is either set to 1 where in the unit is present, or 0 where it is not present. Other encoding methods may also be used.

Figure 7 represents the use of an algorithm trained as described in relation to Figure 4 in the prediction phase. In particular, Figure 7 shows how the algorithm receives as inputs stratigraphic units 752 and seismic data 702 for a candidate sub-surface region. The trained DLI algorithm 704 is arranged to output or extract from the inputs an Earth Model 706 which in this example is porosity. The ultimate goal of the prediction may be to localise, develop or monitor a reservoir for oil and gas extraction or CO2 injection such as the reservoir 1211 of the subsurface region 1200, for example.

In the prediction phase, the boundaries of the stratigraphic units 752 are based on horizons, in this example seismic horizons, usually from survey data. Seismic horizons are the stratigraphic boundaries as interpreted in seismic data. Seismic horizons 380 are lines representing layers of interest based on the interpretation of the seismic amplitudes. For example, a horizon might mark a high amplitude reflected from a boundary between two rock layers. Thus, the seismic horizons 380 can be used to identify the stratigraphic units 752.

In some examples, seismic horizon data is received directly in the prediction phase (i.e. discrete data at specific depths identifying individual horizons is received). In some examples, seismic horizons are identified from the seismic data 702, e.g. by interpretation of the seismic survey data or the processed seismic image.

The stratigraphic units 752 used in the prediction phase are preferably identified and labelled consistently with the stratigraphic units used in the training phase. For example, the stratigraphic unit of the Cretaceous period, or the Hugin geological formation present on the Norwegian Continental Shelf, may be identified both in the training data and the candidate subsurface region, making them suitable units to include in the stratigraphic unit feature.

The naming of stratigraphic units may follow international, national, or regional naming conventions, a company’s or practitioner’s naming conventions, or naming that are suitable for the particular case, e.g. depending on the availability of data, to simplify correlation, classification, and consistency of stratigraphic units.

The geophysical survey data preferable match the training well data. This is achieved by calibrating stratigraphic boundaries identified based on the well top markers with geophysical survey data such as seismic horizons, by tying the well and the geophysical data, e.g. by performing seismic-to-well ties which determines the time-depth relationship, and by determining the geophysical wave velocities to convert the geophysical data from the travel-time domain to the depth domain. In some embodiments, the match between the geophysical data and the well data has been done prior to the method being applied.

The inventors have found that it is preferable not to use horizon-based stratigraphic units 752 for training and to instead use well log-based stratigraphic units. This is because the stratigraphic units defined in the well log, e.g. based on well tops, has a better match with the ground truth, and thus makes for a better feature. Well tops are defined based on the well log data, i.e. in the same domain as the ground truth data. Also, the well log has a higher resolution along the training well than geophysical survey data, e.g. seismic horizons. However, in the prediction phase, it is necessary to determine stratigraphic units based on the geophysical survey data, e.g. seismic horizons, and other data and knowledge which is available away from the training well. This enables the application of the stratigraphic unit feature in the candidate sub-surface region away from well-bore or well-log data (which is where predictive algorithms are needed).

Figure 8 represents the use of an algorithm trained as described in relation to Figure 5 in the prediction phase. In particular, Figure 8 shows how the algorithm receives as input stratigraphic units 852, seismic data 802 and time or depth data 854 for a candidate sub-surface region. The trained DLI algorithm 804 is arranged to output or extract from the inputs an Earth Model 806 which in this example is porosity.

Figure 9 represents the use of an algorithm trained as described in relation to Figure 6 in the prediction phase. In particular, Figure 9 shows how the algorithm receives as input stratigraphic units 952 only for a candidate sub-surface region. The trained DLI algorithm 904 is arranged to output or extract from the input an Earth Model 906 which in this example is porosity.

Figure 10 is a simplified schematic diagram of an example of the DLI algorithm used in examples according to the present invention. The DLI algorithm 1000 is a temporal convolution network (TCN). The TCN is selected for its proficiency in managing sequential data in regression tasks, making it suitable for processing geophysical and geological information and predicting properties. Its architecture excels in capturing short and long-range dependencies in sequential data. It requires a low memory for the inference, parallelization, and stable gradient flow. There are four input channels 1002 for the different types of data input in this example. This includes seismic data, travel time data, stratigraphic unit data, interval velocity and RMS stacking velocities. There are a plurality of TCN layers 1004, a dense layer 1005, and an output layer 1006. In this example, the output layer 1006 comprises a single channel for an (earth) model of a single parameter. However, in another example, the DLI algorithm 1000 is arranged to comprise a plurality of output channels and so is able to output (earth) models of a plurality of parameters. As described, in the training phase, the output is compared to a ground truth target 1008 using a loss function. TCN layers 1004 are then updated based on the loss.

The Earth Model output in the prediction phase can be used to solve an applied geophysics problem such as an exploration, prospecting, and / or engineering problem. The skilled reader will be familiar with various ways in which models of physical parameters (i.e. earth models) can be used to solve such problems. For example, the suitability of a site for drilling for hydrocarbons can be identified when the Earth Model accurately models a hydrocarbon reservoir. An accurate model ensures that hydrocarbon prospects are properly evaluated, minimizing the risk of unsuccessful wells and optimizing the placement of drilling operations to efficiently extract the resources. Important properties of the model include key geological features such as reservoir rocks, seal rocks, and trap structures, or physical properties relating to the presence and volume of hydrocarbons, or relating to the ease of extracting the hydrocarbons, such as porosity and permeability. For example, a high porosity means there are more space within the reservoir rock for hydrocarbons to accumulate, while a low porosity means there may be insufficient pore space to store significant amounts of hydrocarbons. A porosity of 20-30% or higher is generally considered favourable for extracting hydrocarbons, while a porosity of below 5% is considered unfavourable. For the application of CO2 storage, the suitability of a site for storage of CO2 can be identified when the Earth Model precisely models a storage formation. including its potential storage volume and risk of leakage. Such storage formation may be a saline aquifer or a depleted oil and gas field. Physical properties relating to the reservoir’s capacity to store CO2 over time includes cap rock integrity, porosity, and permeability. As an example, properties that may indicate cap rock integrity are elastic properties, resistivity, and seismic velocity. Higher seismic velocities can suggest denser, potentially more impermeable rocks.

As an example, the sub-surface region 1200 visualized in Figure 12 underscores how a detailed Earth Model characterization aids in critical real-world decisions. If the subsurface is accurately modeled, it enhances drilling operations by optimizing well locations, as with well 1204, by indicating prime spots for hydrocarbon extraction within reservoir rock 1206. Similarly, for CO2 storage, a wellcharacterized model aids in assessing potential storage sites, as illustrated by well 1205 aimed at the saline aquifer 1207, enclosed by cap rock 1203. Furthermore, the model’s depiction of windmill 1208, supported by mooring 1209, shows the importance of accurate subsurface models to evaluate subsurface stability for renewable energy infrastructure.

Turning now to Figure 11, a schematically illustrated applied geophysics unit 1100. The applied geophysics unit 1100 comprises an input/output circuitry 1102, at least one processor 1101 and a memory 1103. The memory 1103 contains instructions executable by the processor 1101, causing the applied geophysics unit 1100 to perform one or more of the methods of the training phase and / or the prediction phase described above, as represented by Figures 4 to 9.

The instructions that are executable by the processor 1101 may be software in the form of a computer program 1104. The computer program 1104 may be contained in or by a carrier 1105, which may provide the computer program 1101 to the memory 1103 and processor 1101. The carrier 1105 may be in any suitable form including an electronic signal, an optical signal, a radio signal or a computer readable storage medium.

As used herein, the term “computer readable medium” may be a universal serial bus (USB) memory, a digital versatile disc (DVD), a Blu-ray disc, a software module that is received as a stream of data, a Flash memory, a hard drive, a memory card, such as a MemoryStick, a multimedia card (MMC), secure digital (SD) card, etc. One or more of the aforementioned examples of computer readable medium may be provided as one or more computer program products

Claims (23)

1. A method of training an artificial intelligence (AI) algorithm, such as a neural network, to perform applied geophysics for exploration, prospecting, and / or engineering, the method comprising:

receiving well data for a plurality of training wells:

for each training well:

identifying a plurality of discrete stratigraphic units based on the respective well data; and

training an AI algorithm with the respective discrete stratigraphic units;

wherein the AI algorithm is arranged to perform applied geophysics by extracting a predicted sub-surface model of at least one physical property of a subsurface region from one or a sequence of discrete stratigraphic units for said subsurface region.

2. A method according to claim 1, wherein the well data comprises information identifying one or more well tops and wherein the discrete stratigraphic units are based on the one or more well tops.

3. A method according to claim 1 or 2, wherein the method is for training an AI algorithm to perform explorative geophysics or engineering geophysics.

4. A method according to any one the preceding claims, wherein the method comprises, for one or more of the training wells:

receiving geophysical survey data, such as seismic data or background velocity; and

training the neural network with the geophysical data.

5. A method according to any one of the preceding claims, wherein the method comprises, for one or more of the training wells:

receiving time and / or depth data; and

training the neural network with the time and / or depth data.

6. A method according to any one of the preceding claims, wherein the well data comprises data from one or more well logs.

7. A method according to claim 6, wherein at least one of the well logs has a resolution of at least one sample per meter, optionally at least 2 or 3 samples per meter.

8. A method according to any one of the preceding claims, wherein the method comprises identifying a sequence of discrete stratigraphic units.

9. A method according to any one of the preceding claims, comprising receiving ground truth data comprising a sub-surface model of the physical property for one or more of the training wells.

10. A method according to claim 9, wherein training the AI algorithm comprises reducing a difference between a sub-surface model of the physical property predicted by the neural network and the ground truth.

11. A method according to any one of claims 9 or 10, wherein the well data, including the well top data, comprises time and / or depth information and the method comprises, for each training well, correlating the time or depth of the ground truth data sequence with the time or depth of the geophysical survey data.

12. A method according to any one of the preceding claims, wherein the trained AI algorithm is arranged to extract a predicted sub-surface model of at least one physical property for determining a resource storage or construction site.

13. A method according to any one of the preceding claims, wherein the trained AI algorithm is a stratigraphy-guided deep learning inversion algorithm.

14. A method according to any one of the preceding claims, wherein the physical property is at least one of: an elastic property, P wave velocity, S wave velocity, density, Young’s modulus, Poisson’s ratio, mineral composition, grain size, texture, acoustic impedance, and elastic impedance, porosity, permeability, and fluid saturation.

15. A method according to any one of the preceding claims, wherein the AI algorithm is a temporal neural network (TCN).

16. A method according to any one of the preceding claims, wherein the well data comprises data from 20 training wells or fewer, optionally 15 training wells or fewer, optionally 10 training wells or fewer, optionally 1 training well.

17. A method according to any one of the preceding claims, wherein the well data comprises data from 200 training wells or more, optionally 2000 training wells or more, optionally 2 000 000 training wells or more.

18. A method of applied geophysics for exploration, prospecting, and / or engineering using an AI algorithm trained in accordance with the method as defined in any one of claim 1 to 17, the method comprising:

receiving a plurality of discrete stratigraphic units for a candidate sub-surface region; and

extracting, using the AI algorithm, a predicted sub-surface model of a physical property for the candidate sub-surface region for exploration, prospecting, and / or engineering.

19. A method according to claim 18, comprising receiving horizon data such as seismic horizon data for the candidate sub-surface region.

20. A method according to claim 19, comprising identifying one or more discrete stratigraphic units based on the horizon data.

21. A method according to claim 18, comprising receiving geophysical data and identifying one or more discrete stratigraphic units based on the geophysical data.

22. A method according to any one of claims 18 to 21, further comprising using the predicted sub-surface model to determine the suitability of the candidate subsurface region as a drilling or storage site.

23. A computer program product to perform applied geophysics for exploration, prospecting, and / or engineering, the computer program product comprising computer-readable instructions which, when the program is executed by a computer, carry out the steps of the method of any of claims 1 to 17 or 18 to 22.