b'AEGC 2023Short abstractssmall number of homogenous layers will, in many situations,is expected to be a robust tool to define lithology automatically. more closely resemble the real earth with its layered stratigraphy.In this work, we apply unsupervised ML techniques such as Depth and thickness estimates of geological layers can then beK-means, Fuzzy C-Mean, and Support Vector Machine for binary directly obtained from the model parameters, rather than vialithology clustering (coal layers and other lithology). Then the somehow thresholding a smooth conductivity function. other lithology is subdivided into detailed facies in the wells with a manual interpreted lithology log. Finally, the labelled data set A disadvantage of modelling stratigraphic layers directly is that,above will be used for supervised ML such as EtraTrees, Random if the number of layers in the model is incorrect, then the resultsForest, Extreme Gradient Boosting, Multilayer Perceptron, and may be biased or misleading. This paper attempts to determineConvolutional Neural Network techniques to predict the target, the most-probable number of layers represented by an AEMbanded coal seams, and other lithology on a large number of decay, via a Bayesian Occams razor. Increasing the number ofun-interpreted wells. The input well logs are Gamma-ray (NGAM), layers in the model will always yield a better fit to the data, soGammagamma (HRD), Density (DENS), and Resistivity (RES) data-fit cannot be used to select between models with differentlogs from the wells in Ha Lam coalfield, Vietnam.numbers of parameters. However, the Bayesian model-selection methodology, which can be used to choose between modelsIn a nutshell, our approach comprises two major stages: First, an with different numbers of layers, yields an automatic Occamunsupervised ML model is used to create labelled data. Second, factor, penalising models with more layers. The combinationa supervised ML model is trained from the above-labelled data of data-fit improving with number of layers, and Occam factorto predict lithology for the entire wells on the field.worsening, yields a natural and consistent assessment of the most-probable number of layers.LithoSurferA unique relational database platform for The technique is applied to AEM data from the 2010 Fromeexploration data.TEMPEST AEM survey. Some implementation details are discussed, especially relating to the integration requiredFabian Kohlmann 1, Wayne Noble1, Romain Beucher2, Moritz over all possible model parameter values for each modelTheile1, Alejandra Bedoya Mejia3, Malcolm McMillan4 and with a different number of layers. Examples are shown fromSamuel Boone4decays obtained close to boreholes, allowing the results to be1 Lithodat Pty Ltdcompared to known geology, and illustrating the usefulness of2 The Australian National University 3 The University of Adelaide 4 the method. The University of Melbourne Well managed, standardised data is vital for the exploration Automatically identifying lithofacies from wireline logs:industry but as most available geoscientific datasets are A case study in Cuu Long Basin, Vietnam. regionally bound and have bespoke implementations, it Duy Thong Kieu 1 , Hong Trang Pham1, Quang Man Ha2,is challenging to merge all data into a consistent global VietDung Bui3, Qui Ngoc Pham3 and Huy Hien Doan3 framework. Lithodats vision is to provide explorationists with global geoscientific databases and analytics to decrease the 1 Institute of Geological Sciences (IGS), Academy of Science &time taken to gain professional exploration insights to simplify Technology (VAST), Viet Namand de-risk resource discoveries.2 PetroVietnam Exploration Production Corporation - PVEP, VietnamTo achieve this, Lithodat has developed LithoSurfer, a secure 3 Vietnam Petroleum Institute, Hanoi, Vietnam online data platform for viewing, analysing and extracting data in a geological and spatial context. LithoSurfer provides Lithofacies is important for reservoir evaluation. In this work, wequick access to a wealth of information (analytical metadata, present a workflow to define the lithofacies automatically frommachine parameters, lab information, persons, references etc.) wireline logs. Our workflow includes three phases: in the first phaseacross multiple analytical techniques and localities. LithoSurfer the boundaries are automatical defined from wireline logs byharvests its power from being built on a relational database using the recurrence technique; in the second phase, we extractwith a modular architecture of very detailed data models for the data set from wireline logs within the boundaries and put it ineach analytical technique. All publicly available analytical a modified fuzzy c-means clustering process. The second phasedata is validated, standardised and integrated into our cloud-results are the input of a machine learning process to identify thehosted database. Proprietary data can be securely integrated facies. We apply our workflow to a data set in Cuu Long basin,and normalised into LithoSurfer and analysed together with all Vietnam. The results are comparable with expert analysing results. public data. This consolidation opens up the full potential that spatial geoscience data has to offer and is a vast improvement Automate lithology prediction process from well logon storing data in separate spreadsheets and folders as often happens within teams and exploration projects. LithoSurfer data using Machine Learning and Deep Learning: Amakes dispersed and complicated geoscience datasets case study from Ha Lam coalfield, Vietnam. understandable and easily accessible to any explorationist.Duy Thong Kieu1, Nguyen Binh Kieu2, Ngoc Cuong Phi2 andWith LithoSurfer geoscientists can now visualise, combine Duy Phuc Do2 and export data from areas of interest including diagrams, 1 Institute of Geological Sciences (IGS), Academy of Science &graphs and auto generated reports on the fly. Having data Technology (VAST), Viet Nam and analytical tools at your fingertips for systems such as 2 Vinancomin-Mining Geology Join Stock Company geochronology, thermochronology, geochemistry, VR and thermal histories helps increasing exploration insights and Manual interpretation of massive well log data is time-consumingworkflow efficiency. All data can be extracted in multiple and prone to human bias. Machine Learning (ML) predictionformats to take full advantage of new techniques such as 111 PREVIEW FEBRUARY 2023'