Generating realistic reservoir models that match observed well data is important for understanding and predicting subsurface behaviour. By combining static and dynamic data from wells with prior geological knowledge and fluid flow models, geoscientists and reservoir engineers can generate representative subsurface reservoir models.

Data assimilation integrates information from models and observations to generate improved estimates of systems for prediction and uncertainty analysis. In subsurface modelling, data assimilation has been applied for history matching reservoir models by assimilating spatio-temporal data such as downhole measurements of pressure, temperature, and flow rate. Integrating static and dynamic data from wells in a data assimilation workflow for history matching can aid the development of reliable reservoir models for forecasting reservoir performance and maximising fluid recovery. However, for large-scale, highly nonlinear dynamic systems, such as the geothermal reservoir models considered in this research, conventional data assimilation techniques can be computationally expensive, as the number of parameters to be estimated is orders of magnitude higher than the observed measurements.

This research develops models of geothermal reservoirs by applying artificial neural networks for data assimilation and predictive modelling. Using deep neural networks for well conditioning, history matching, and reduced order modelling, representative models of geothermal reservoirs that honour observed well data are generated and can be used to forecast reservoir behaviour at a fraction of the time required for direct numerical simulations. The approach presented here has many advantages over conventional methods: 1) the use of grid-free surface-based geological models (SBGMs) that preserve geological realism and model resolution, 2) neural networks for data assimilation and well conditioning to accelerate the history matching process, 3) space-filling curves for representing geothermal reservoir simulation data on unstructured meshes to aid data compression, and 4) transformer-based models for temporal evolution and replacing the geothermal reservoir simulator for rapid reservoir forecasting.