b'Seismic window Seismic windowCSIRO not alone with ML characteristics of a particular facies with depth that bamboozles machine The Preview Editor sent me an articlelearning attempts to classify facies submitted by Beloborodov et al becausebased on well logs alone. To solve this she thinks that I am an expert in Machineproblem the researchers have come Learning, which is a far cry from the truth,up with a hybrid approach with the but I vaguely know how it works. machines doing their thing based input The article recognises that the wellfrom rock physics in the form of depth log characteristics of a facies or rocktrends.type vary with depth and each faciesWell, I dont want to steal the limelight may have a different depth dependentso here is the full article describing the Michael Micenkorelationship. It is this variation in theCSIRO approach.Associate Editor for Petroleummicenko@bigpond.comNot by machine learning alone: Automatic facies classification for seismic inversionR. Beloborodov, M. Pervukhina,rescue by providing understandingThe free parameters of the best-J.Gunning, J. Hauser, I. Emelyanova,of depth-dependent rock properties.fitting rock physics model, such as M.B.Clennell (CSIRO) Numerous theoretical and empirical rockcrucially important cement fraction or Marina.Pervukhina@csiro.au physics models exist for several specificcoordination number, are estimated at rock types. Only some of these modelsthe same time. The algorithm is most The CSIRO has developed a new approachtake depth-dependency into accountsuccessful when theoretical rock physics to rock facies classification that fusesexplicitly. These models are generallymodels fit nicely the petrophysical data machine learning and rock physics. Pre- developed for rock types that exhibitin a well but a user can instead choose processing of well log data required forstrong variation of elastic propertiesto fit empirical trends from local or quantitative interpretation of seismic datawith stress, such as unconsolidatedglobal datasets. However, the physical now takes minutes instead of weeks. sandstone. Overburden stress, as a proxyparameters of practical importance such With the advent of the machine learningfor depth, is explicitly incorporated intoas cement fraction or sorting cannot be (ML) era, many time-consuming andtheir equations. Other models, such asestimated in this case.tedious manual processing tasksthose for cemented sandstones, shales have been automated, drasticallyand carbonates, do not explicitly deriveIn addition to identifying the most reducing time from survey to discovery.rock properties as functions of stressprobable rock type the EM algorithm also Unfortunately, petrophysical faciesbut may be indirectly related to depthestimates the probabilities for alternative classification, a pre-processing stepvia changes in density, porosity orrock types. This uncertainty estimation required for quantitative interpretation,cementation. can be propagated forward into seismic inversion and reservoirquantitative interpretation, for example, modelling, is not one of those tasks.It is well established that different rockwhen using a probabilistic approach for Previous attempts at developingtypes exhibit distinctive compactionseismic inversion.automatic facies classification algorithmstrends. For this reason, the compaction have failed owing to a fundamentaltrends that would generally hinder rockThe challenges around rock typing, problem, namely, that petrophysicalfacies classification can be transformedsuch as selecting the set of rock physics properties of rocks change withinto positive characteristics. Data pointsmodels that best describes the true compaction, a ubiquitous geologicalcan be assigned to the right classeslithologies, are not unique to energy process. This means that the propertiesusing both their physical propertiesresources. As mineral exploration moves of a rock vary significantly at differentcombination at each depth and alsointo more complex geological settings depths even within the same faciesthe variations in these propertiesunder sediment cover, rock physics having the same composition andwith depth. The ML algorithm calledmodels are becoming increasingly depositional history. These burial trendsExpectation-Maximization (EM) isrelevant for the robust delineation and of petrophysical properties obscure classadvantageous to accomplish this.characterisation of mineral resources. boundaries and hoodwink the best MLStarting from a random guess onStrategies that have proven to be algorithms. class memberships, the EM algorithmsuccessful in the context of energy iteratively fits rock physics modelsresources are likely to also be successful However, when ML algorithms fail onand then updates memberships tofor the more challenging setting around their own, rock physics comes to themaximise the likelihood of the system.mineral resources.33 PREVIEW OCTOBER 2020'