b'Seismic window Seismic windowMachine Learning can be either supervisedthe 2nd and the 5th of September. This or unsupervised. Supervised learningconference is being jointly organised by occurs where the data being evaluatedthe ASEG, PESA and AIG and replaces has already been interpreted and faciesthe very popular and successful ASEG classification and seismic prediction of rockConference and Exhibition.properties such as porosity or shale contentI was recently sent a list of papers that are available. Unsupervised learning occurswill be presented and after looking when the data is examined and groupedthrough the titles I have come up with without preconceptions (this grouping issome statistics (Table 2) and my top five done in multi-dimensional spacein 2D itspapers for seismic interpreters.like fitting a straight line to a set of points and using this line to predict a missingTable 2.AEGC 2019 statistics.value). Examples of this are waveform classification and attribute clustering. Some of the applications of Machine LearningNumber of papers accepted: 253presented at the EAGE conference areNumber of papers with a seismic/ 97Michael Micenko shown in Table 1. As you can see therepetroleum content:Associate Editor for Petroleum is a diverse range of geological andNumber of petroleum papers email@example.com geophysical applications for ML. companies:Number of papers from oil26Machine learningTable 1.Some geological andcompanies:geophysical applications of MachinePapers from contractor / service14takesoff Learning. companies:Number of papers that are ofAll of themThe flavour of the day at this years1. Automated detection of micro-seismic events interest to interpreters:EAGE Conference in London was Machine2. Seismic inversion using a deep neural networkLearning. 3. Ground roll attenuation Top five papers (based on titles and in This is something new to me so I set4. Predicting gas content of shale gas reservoirs no particular order):off to find out what Machine Learning5. Efficient seismic data interpolation In the absence of a fully-fledged (ML) actually is. What I found was that6. Porosity prediction conference handbook my top five list almost all the vendors had rebadged7. Texture based classification of seismic patches of presentations could be handy for their existing software with the buzzdeciding where you need to be whilst at words Machine Learning. They may8. Assignment of biostratigraphic age ranges the conference (Note: I am excluding my have modified the code a bit but if youexcellent talk on the Cimatti Field to keep are familiar with neural networks orThe trained eyes of interpreters havethings as unbiased as possible).waveform classification you alreadybeen recognising textures in seismic data know something about ML, whichfor a long time, but can only describeRegional stratal slice imaging of the is a subset of Artificial Intelligencethe process with phrases like I know itNorthern Carnarvon Basin, WA. Tony (AI), a branch of computer science.when I see it but cant describe exactlyMarsh, Chevron.ML is growing in importance in thewhat Im seeing. This is how I would petroleum geosciences because wedescribe Deep Learning. The computerPyxis: a study in cost efficient near field need to extract information from vasthas a vast data library that it compares toexploration, discovery and appraisal. Peter data sets. So we let computers do thethe input data and assigns a probabilityThomas, Woodside.work because machines, apparently, canor weight to particular outcomes. WhileThe road to Dorado: Factors leading to a capture uncertainty and are consistentAI has been around since the 1950s,play opening hydrocarbon discovery. Dr but, more importantly, they are fast. Deep Learning is relatively new and hasFrederick Wehr, Wehr Advisory.Brian Russell (CGG) described ML asrelied on the increase of GPU power and nothing mysticalits actually just acomputational speed over the last 5 years. Eromanga oil traps a multi field post transform of the input data. In a similarmortem. Keith Martens, Martens There are many more applications andPetroleum Consulting.vein, Paul de Groot (OpendTect) describesseveral algorithms to choose from and the ML as finding complex relationships inkey is to find the best method for a specificDude, wheres my AVO? A case study data from a variety of sourcesusingtask. Ill finish with these words of wisdom,from the Browse Basin, Northwest Shelf, statistical techniques. ML givesagain from Brian Russell, if the processAustralia. Dr Said Arimibeshell, Discover computers the ability to learn from datadoesnt work on a simple case it sure wontGeoscience.without being specifically programmed.work on a complicated one.To do this the machine needs a largeOf course there are other good papers amount of data so it knows what to lookbeing presented so get along and for. One thing we never have is enoughHeads up on AEGC 2019 support the authors and the conference data to represent a complete range ofeven if its just for a day. And yes, there possibilities, but this shortfall can oftenYou should all know by now that AEGCis a session on Artificial Intelligence and be filled with synthetic or modelled data. 2019 is being held in Perth betweenMachine Learning.AUGUST 2019 PREVIEW 38'