Building Reservoir Models With Seismic And Well-log Data Via Deep Learning

Horizon can be picked from seismic images by consistently following seismic reflectors,which is a fundamental and crucial step for seismic interpretation but remains a time-consuming task.Although various automatic methods have been developed to extract horizons in seismic images,most of them may fail to pick horizons across discontinuities such as faults and noise.To obtain more accurate horizons,we propose a dynamic programming algorithm to efficiently refine manually or automatically extracted horizons so that these horizons can more accurately track reflectors across discontinuities,follow consistent phases,and reveal more geologic details.In this method,we first compute an initial horizon that may not be accurate and only needs to follow the general trend of the target horizon.Therefore,such an initial horizon can be obtained with an automatic method,manual picking,or interpolation with several control points.Then,we extract a sub-volume of amplitudes centered at the initial horizon and meanwhile flatten the sub-volume according to the initial horizon.We finally use the dynamic programming to efficiently pick the globally optimal path that passes through global maximum or minimum amplitudes in the sub-volume.By doing this,we are able to refine the initial horizon to a more accurate horizon that follows consistent amplitude peaks,troughs,or zero-crossings.As our method does not strictly depend on the initial horizon,we prefer to directly interpolate an initial horizon from a limited number of control points,which is computationally more efficient than automatically or manually picking an initial horizon.In addition,our method is convenient to be interactively implemented to update the horizon while editing or moving the control points.More importantly,these control points are not required to be exactly placed on the target horizon,which makes the human interaction highly convenient and efficient.We demonstrate our method with multiple 2D and 3D field examples that are complicated by noise,faults,and salt bodies.Deep learning methods have been popularly used in geophysical problems,such as model building,fault interpration,reservoir prediction,and so on.Many researchers also have introduced the deep learning to predict acoustic impedance from seismic data which is typically considered as an ill-posed problem for traditional inversion methods.Most of the deep learning methods,however,are based on a 1D neural network,which is straightforward to implement but often yields laterally unreasonable discontinuities in predicting a multi-dimensional impedance model trace-by-trace.We improve the deep learning-based method to predict impedance by implementing it with a 2D convolutional neural network(CNN)and introduce the constraints of an initial impedance model into the network.We first calculate the initial impedance model with interpolation from the impedance logs with the guidance of seismic structures and then the initial model is input to the network to provide a low-frequency trend control.The results show that it is helpful for both the 1D and 2D CNNs to yield stable impedance predictions.The architecture of the proposed 2D CNN is quite simple but training the CNN is not straightforward because a full 2D impedance label is not available.To prepare a 2D training dataset,we first define a random path that passes through multiple well logs.We then follow the path to extract a 2D seismic profile and an initial impedance profile which together form an input to the 2D CNN.The set of well logs(traversed by the path)serves as a partially labeled target.With the randomly extracted 2D training datasets,we train the CNN with weak supervision by using an adaptive loss where the output 2D impedance model is adaptively evaluated at only the well logs in the partially labeled target.As the 2D training datasets are randomly chosen from the original 3D survey in all directions,the trained 2D CNN can predict a consistent 3D impedance model section-by-section in either inline or crossline direction.Synthetic and field examples show that the proposed 2D CNN is more robust to noise,recovers thin layers better,and yields a laterally more consistent impedance model than a 1D CNN with the same network architecture and the same training logs.Subsurface modeling plays an important role in hydrocarbon exploration but remains a challenging task that typically involves a full and reasonable integration of geophysical observations and geologic constraints.We present a workflow to fully utilize seismic amplitudes,well-log properties,and interpreted seismic structures to build geologically reasonable models.We take the Volve field data as an example and apply our workflow step by step as follows:First,we perform some preprocessing on the provided Volve seismic data,horizons,and well-logs to remove anomalous values and adjust seismic-well ties in depth domain.Second,we use a dynamic programming based method to fill the holes and refine the vertical positions of the provided horizons and efficiently pick more horizons.We further use the horizon surfaces to interpolate a relative geologic time(RGT)volume which can be considered as an implicit structural model representing seismic structural and stratigraphic features.Third,we integrate the provided well-logs and the computed RGT volume to compute a subsurface model that conforms to both well-log properties and seismic structural and stratigraphic features.Finally,we predict a final model with deep learning using seismic and well-log data and meanwhile introduce the initial model as low-frequency constraints into the network.The results show that our workflow is able to produce geologically reasonable subsurface models with high lateral continuity and vertical resolution.