Den 14. aug. 2013 kl. 09:03 skrev Alf Birger Rustad: Dear developers,
We have done a terrific effort improving the OPM codes the last year, we all deserve a tap on the shoulder. On the reservoir simulation part it seems like we are ready to take on the next big challenge. Namely, demonstrating that we can replace an industry reservoir simulator on one full field model effectively. The Norne case seems most relevant for many reasons. The most prominent being: it is available to academia, several of us are already familiar with it, it converges relatively easily and simulation time is low, it is about as standard as it gets (i.e., few perks and hacks). I believe we should have the following design goals (please add if you miss some): -All physics present in the full field model should be accounted for (e.g., fully implicit black oil and well models with relevant input accounted for in the simulations) -We need to be able to parse the data and include files, and feed them to the simulator without user intervention -We need to produce relevant grid and summary output for 3D visualization and 2D production data without user intervention -We need to be within striking distance performance wise. I believe three times slower than relevant industry simulator is acceptable here. You will find pointers to the Norne model here: http://www.ipt.ntnu.no/~norne/wiki/doku.php?id=start What do you guys think, is this something we can achieve by Christmas? I think this may be doable, and agree that it makes for a good target case, due to its openness. To your second point, I must caution that I do not know what tricks and special features are used in the data-as-given that could give us trouble. In order to support this point, the opm-parser module will be needed in a usable state. We can and should start working on the case before that is available, but probably with a somewhat modified input deck. Moving forward, there are a few points I'd like to focus on: 1. I think that the current effort in testing and validating the sim_fibo_bo.cpp simulator program using simpler cases is essential, to ensure that we can trust the simulator when moving to more complex cases, such as Norne. 2. The above mentioned simulator does not handle switching of well targets, which is needed for a real case. We have done this in other (operator splitting) simulators, and for fully implicit simulation we have done it in MRST. 3. The simulator and autodiff library has not been optimized. Although the autodiff library has been written in such a way that it should be possible to get good performance, very little profiling has been done (preliminary results indicate that quite some time is spent in the Eigen::SparseMatrix matrix multiply operation). I think we should be able to make the linear solver the bottleneck. 4. The Norne case is just large enough to be unfeasible with a direct solver (at least if we want some performance). So work needs to be done on preconditioning strategy etc. Again, much experimentation has been done in MRST, so it should be doable. Atgeirr
_______________________________________________ Opm mailing list [email protected] http://www.opm-project.org/mailman/listinfo/opm
