Background
Log data, such as cores & seismic, gathered during exploration & appraisal of an oilfield comes with a lot of uncertainty that can have implications on accurately modelling the reservoir. Compounded with issues associated with the differing scales of the presented data, are inconsistencies in well logs either derived from incorrect measurement or accuracy limitations of the tools of the day. Whilst data uncertainty can result in reduced understanding of the reservoir and recoverable reserves, it can also negatively impact accurate well placement and ultimately production & field life.
Reservoir engineers have traditionally looked towards computationally demanding physical models to compensate for inaccurate or inadequate field data. Whist this solution provides relatively reliable results for operators, the simulation techniques used, require significant user input & data manipulation and are time and computing intensive.
Data & AI solutions
Differently scaled, and highly heterogeneous field data including seismic, well logs & production data were used to train the model. Well log information included gamma ray, deep resistivity, bulk density, neutron porosity and total porosity measurements. Data was relatively incomplete and required significant enhancement before use in the application.
Two artificial neural network (ANN) models were used to 1) produce synthetic well data 2) predict production based on both synthetic & actual logs. Multilayer ANNs were used to capture correlations between seismic, wells logs, production and completion data. The synthetic logs produced could then be used for reservoir and production estimation using a separate performance network that predicted production of each reservoir location. These were then verified against conventional simulation techniques for accuracy and run-time comparison.
Multi-layered feedforward ANNs were used throughout this work with the number of neurons in each layer optimised between 100 and 300.
Businesss case & beneifts
The development of an intermediate approach to well characterisation & placement which complements conventional simulations could benefit the business by reducing workloads. An AI solution that can produce synthetic well logs, of greater accuracy & resolution, would greatly benefit reservoir characterisation workflows and ultimately the accuracy of reservoir estimations and future production.
The well placement design method was significantly enhanced with the AI first-pass optimisation tool. Reservoir engineers were able to reduce design uncertainty by focusing their attention on AI optimised wells instead of running detailed and costly computational simulations across all options.
