Organization responsible: SNN Adaptive Intelligence, SHELL

People involved:

• Bert Kappen (project leader, researcher, SNN)
• Wim Wiegerinck (researcher, SNN)
• Martijn Leisink (researcher, SNN)
• Willem Burgers (researcher, SNN)
• Mirano Spalburg (researcher, SHELL E&P)

Project description:

We will build a petrophysical decision support system to interpret measurements acquired in boreholes and to recommend further steps. This system will be based on a Bayesian network, modelled with expert knowledge from Shell EP.

The exploration for oil and gas requires real-time petrophysical expertise to interpret measurements acquired in boreholes and to recommend further steps. High time pressure and the far reaching nature of these decisions, as well as the limited opportunity to gain in depth petrophysical experience suggest that a decision support system that can aid the petrophysicist will be very useful. In this project, we aim to build in collaboration with experts at Shell EP, a Bayesian network model the petrophysical domain. Similar models may be useful in other ICIS domains, e.g. a decision support to interpret and recommend measurements for situation awareness in contaminated hazardous environments, chemical pollution etc.

The main goal is to demonstrate the feasibility of the required system. In the way to this goal, we encounter questions from real-world modelling. How to model and infer a Bayesian network in a real-world domain? What are the essential model assumptions? How to do inference with continuous variables with noisy nonlinear relations?

We define a model based on the available Shell knowledge. The model is a Bayesian network in the continuous domain. It has noisy nonlinear relations between the variables. Computation is intractable, and therefore we resort to approximate inference, in particular sampling due to its general applicability.

Initially, we tried an off-the-shelve sampler (BUGS), which turned out to be insufficient. Therefore we choose to build our own software. We build the model in Matlab, we developed several model representations and implemented several samplers, from which we concluded that an exponential representation combined with an MCMC (Markov Chain Monte Carlo) performed sufficiently well (consistent and reproducible results within reasonable time - half hour/hour). Next, with this result, we implemented the software in C++, in order to obtain a speed up the software, and to make a demo with a graphical user interface. The result is a speed-up to a few minutes.

Publications: