• convolutional neura network
• faults detection
• geotermal eploration
• machine learning
• seismic acquisition
• Value of information

The objective of this thesis is to present a value of information (VOI) analysis of seismic data that uses a machine learning technique for deriving the required Bayesian statistics. The VOI methodology evaluates the benefit of acquiring additional information, by means of data collection activities, before making a decision with uncertain outcome.
VOI analysis is studied here in the decision context of a geothermal drilling project in which the decision considered is whether to drill the production well or leave the field. Considered the significant uncertainty associated to the exploration of geothermal resources and the high cost of a geothermal drilling operation, it is fundamental to carry out extensive investigations to get a measure of the availability of the resource and the economic potential of the project. One possible way to assess the value of data collection activities is value of information analysis.
VOI calculations must be performed before collecting the data, so that the decision maker evaluates if the information to be gathered could have a significant impact on the decision (drilling a well in this case) and if it’s worth to invest in the survey by comparing the VOI result with the cost of the acquisition.
For the VOI example proposed in this thesis, a seismic reflection survey is simulated to assess the impact of acquisition geometry on the reliability of seismic imaging and, in particular, of the imaging of major faults within the reservoir. Having knowledge of the fault locations in the subsurface when exploring for geothermal resource is crucial since fault systems highly affect the production and recharge of geothermal fluids. The VOI quantitatively evaluates the potential value that the information gathered by seismic data could add to the decision context.
The first task was to build synthetic P-wave velocity, S-wave velocity and density models based on data, reported in the literature, related to the geothermal field at Larderello (Italy). Seismic velocity and density values across the model are needed to generate the synthetic data through a two-dimensional elastic forward modelling. From the simulated measurements, 2D seismic images were produced through reverse time migration (RTM).
To investigate how well faults can be resolved in the migrated images, a convolutional neural network (CNN) is applied to detect faults within the images and to provide the reliability measure needed for deriving VOI. The reliability is expressed by the posterior probability that quantifies how frequently the predicted faults within the migrated image will match the actual presence of faults and vice versa. The CNN applied to perform the fault binary segmentation task is the FaultSeg3D model, developed by Wu et al. (2019). The network is fed with 3D seismic images and it outputs a fault probability map (probabilities between 0 and 1) which has the same size of the input seismic images.
The FaultSeg3D model has been re-trained after creating a new synthetic training dataset composed of 251 3D seismic images and their corresponding fault labels with the size of 128x128x128 and a validation dataset of 26 seismic and fault labelled volumes of the same size. Although the CNN was trained by using synthetic images created with a different workflow, the training leads to satisfactory prediction performance of fault locations when applied to migrated images.
The quantities of correct/incorrect interpretations are used to derive likelihoods and posterior probabilities for VOI calculations. The value of information analysis not only takes into account the reliability of the information (interpreted faults), but also the prior uncertainty about the faults presence in the investigated subsurface and the economic possible outcomes of the decision in case of a successful/dry well.
The VOI was analytically calculated by assuming a discrete prior probability and Net Present Value (NPV) of the geothermal project. Then, the calculation is done in the cases of different scenarios for the following parameters: reliability of the information, prior probability and economic impact of the decision (NPV) to evaluate how they could affect the results.

Wu, Xinming, et al. "FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation." Geophysics 84.3 (2019): IM35-IM45.