With all the enthusiasm surrounding big data and analytics it would be reasonable to presume that the Petroleum data driven analytics session at the Society of Petroleum Engineers Annual Technical Conference and exhibition was were the real action was. The PDDA session was quite well attended with nearly 100 present, perhaps not quite enough to support the current big data hype!
Schlumberger’s Vikas Jain reported on the fruit of three years R&D on data-driven well log analysis. Today’s log data is ‘big,’ with large spatial, temporal and multi-well datasets, some with more than 100 dimensions. ‘Traditional’ interpretation workflows use lab and core data to train generic regression models. But this kind of sequential approach is sub optimal as we can only comprehend a small number of measures so we use data reduction. Also generic models may not be applicable everywhere. Ideally we should mine the whole dataset but this is a non-trivial exercise as a high level of domain expertise is required to mine data in a meaningful fashion and produce reproducible results. Enter ‘next generation’ petrophysics described as a ‘packaged idempotent* workflow.’ Data-derived classification and regression models are baked into the workflow. For instance, fluid distribution and facies can be obtained by classifying NMR signatures, analysis of LWD data gives a gas fingerprint and rock facies classification from spectroscopy, bringing petrophysical expertise to the end user–174735.
Ognjen Grujic (Stanford University) has trialed data mining on a large shale dataset provided by Anadarko to forecast production and ultimate recovery. This is a ‘tricky’ problem and the speed of development means there is not enough time for a geomodeling approach. A novel ‘functional principal component analysis’ approach was used to derive decline curves for various combinations of fracturing, geology, number of stages and so on. Geography and frac parameters were most influential, geology and log-derived TOC less so. The main tools used were the ‘R’ statistical packages DiceKriging and FDA–174849.
Luka Crnkovic-Friis gave a thinly-veiled commercial for Peltarion Energy’s Frac Forecast a ‘deep learning’ based approach to geology-driven shale estimated ultimate recovery (EUR) prediction. Deep learning is used in speech and image recognition, tasks which are hard to do with an equation. Input was an 800k point dataset of geology (thickness, poroperm and TOC), well data and production logs. The approach is claimed to halve the error in EUR, down from 66% to 33%. If geology changes prior knowledge can be added, possibly from pilot wells. Peltarion is seeking a partner to develop the technology–174799.
Today’s computers make it easier to run very large numbers of simulations to study the impact of geology on recovery factor. But as Satomi Suzuki (ExxonMobil ) reported, the conventional approach can still take too long – particularly to analyze the results. An alternative approach was suggested by the EU-backed Saigup project which used association rule mining (as used by Amazon and others to solve the market basket problem). The technique reduces the problem space. It was shown for instance that 80% of 5 spot/500m spaced wells have high EUR and/or high recovery was observed for down dip progradational geology and high aggradation angle. Further investigation compared environment of deposition with fault patterns–174774.
Rigorous conventional reservoir studies are time and labor intensive and rely on subjective assumptions and sparse data. Sarath Ketineni (Penn State) has used artificial neural nets (ANN) a.k.a. soft computing to minimize geological uncertainty. ANN expert systems map complex relationships in the data automatically. A seismic data cube, some well data for training are input to the net and the output is synthetic well logs at any location. One trial involved 13 3D attribute volumes, well logs, completion and production data. Matlab’s ANN functionality was used on part of data set for training. The results as presented were underwhelming although ‘very good agreement’ was claimed–174871.
Full physics modeling is a rigorous but computationally intensive approach to the evaluation of system response. Surrogate models provide an approximation of full physics models and can be used in uncertainty quantification. Srikanta Mishra (Battelle) compared experimental design (ED) and Monte Carlo (MC) simulation strategies for building such surrogate models. The Battelle study evaluated a reservoir’s potential for CO2 storage from nine inputs (reservoir thickness, caprock, poroperm, injection rate and so on). A full compositional simulation was performed with CMG’s GEM simulator. Various techniques (Box-Behnken, augmented pairs, latin hypercube, maximum entropy) were trialed as ways of choosing starting points for full physics simulation. For larger numbers of factors ‘you might as well do MC’. Kriging also ran. Maximum entropy won out over Box Benhken while augmented pairs was the worst performer. Screening data for ‘heavy hitters,’ data points that bias the results, is important–174095**.
Apache Corp.’s David Fulford argued that machine learning can now be considered to be reliable technology*** for evaluating unconventional well performance which today is poorly understood. Pressure transient propagation can take years. While engineers need to forecast, their tool of choice (decline curve analysis) has an implicit bias. Apache uses Bayesian machine learning to account for bias and ‘always converge on the true answer.’ Markov chain analysis with the Metropolis-Hastings algorithm was applied to data from the Bakken/Elm Coulee field. Hindcasts of cumulative production showed that IP30 was no predictor of EUR. Hi proppant loading increases initial production but EUR is not affected. Is ML reliable technology in the sense of the SEC? Fulford believes it is, ‘Our approach shows all of the necessary criteria for its consideration as reliable technology’–174784.
* a.k.a. blinding with science!
** See also our report (Vol 20 N° 7) on Battelle’s EluciData Hadoop/NoSQL big data engine.
*** Important (and contentious?) terminology in the context of reserves reporting.
© Oil IT Journal - all rights reserved.