Man, Machine and Data: A Mineral Exploration Perspective Eun-Jung Holden, Geodata Algorithms Team The Centre for Exploration Targeting, School of Earth Sciences The University of Western Australia Team members: Daniel Wedge, Jason Wong, Peter Kovesi, Yathunanthan Vasuki, Tom Horrocks, David Nathan
Frodeman (1995) on Geological Reasoning We are seldom in possession of all the data we would like for making a decision, and it is not always clear that the data we possess are unbiased or objective. We are forced to fill in the gaps in our knowledge with interpretation and reasonable assumptions that we hope will be subsequently confirmed. Geological reasoning; geology as an interpretive and historical science. R. Frodeman1995 Geological Society of America Bulletin v. 107, no. 8, p. 960-968. Human interpretation outcomes are highly variable BUT smart decisions can be made based on observation, common sense, experiential learning and INTUITION
Machine Analysis Pros Fast Objective Reproducible results Cons False positives Black box Lack of confidence in the uptake Example: Structural interpretation of magnetic data Structures are associated with discontinuities within data Automated first pass analysis can minimise human biases Geological interpretation o Different types of geological features o Chronological order of structures CET Grid Analysis Extension for Geosoft Oasis Montaj Challenging to model implicit/tacit knowledge that connects the dots
Integrating machine and human intelligence to maximise geological knowledge gain and minimise human bias Equip geoscientists with tools for data analytics (exploring facts to answer a specific question) and data mining (recognition of hidden patterns for new knowledge discovery) Ensure that the algorithms are interpretable and useable by geoscientists for their decisions Involving geoscientists as much as it can to allow controlling the algorithmic process Providing adequate post processing capabilities to control the quality of the outputs to support their decision making Visualisation, human computer interaction and workflows play an important role
Research & Development Improving confidence in geologists decision making for: 1. Structural interpretation of regional scale potential field data Geological Survey of Western Australia, Australian Research Council (2014-2017) 2. Structural interpretation of televiewer downhole images Rio Tinto (2010-2013) 3. Validation of geologists logging of drill hole data using geochem assays Rio Tinto (2015-2017) 4. Multi-element geochem data analysis for drill hole data (applied to Kevitsa mine data) First Quantum Minerals (2015-on-going)
1. Structural Interpretation of Potential Field Data For structural interpretation, interpreter seeks discontinuities (edges, ridges, valleys) - highly variable interpretation outcomes Automated image analysis & interactive visualisation can be used to provide interpretation spell check CET Grid Analysis Extension (for Geosoft Oasis Montaj) http://www.geosoft.com/pinfo/partners/cetgridanalysis.asp Holden et al. 2016, Improving assessment of geological structure interpretation of magnetic data: An advanced data analytics approach, Computers & Geosciences
Integrated Exploration Platform (IEP) Software toolset for ArcGIS & MapInfo with a novel computer-assisted and interpreterdriven approach to integrated multidata interpretation for regional scale exploration data Interactive visualisation using image blending, new image enhancement and structural and lithology interpretation support tools The first public version of the IEP is available from http://www.waexplorationplatform.com (Geolocked to datasets within WA) Sponsored by the Geological Survey of WA (GSWA) in the Exploration Incentive Scheme (EIS), and Australian Research Council (ARC) linkage grant (LP140100267).
2. Structural Interpretation of Televiewer Images Funded by Rio Tinto Iron Ore (2010-2013) Commercialised by Advance Logic Technology,2015 Televiewer images are used for structure analysis for structural geos and geotechnical engineers. Hundreds of kilometres of televiewer data to analyse a year creating a bottleneck in industry workflow. Develop a software platform to provide automated structure analysis & workflow Malcolm Rider, The Geological Interpretation of Well Logs.
Drill and log Flow / block model Zonation Structures Analyse structure confidence Assisted rapid detection Represe nt picks Stability analysis
Commercialisation Televiewer Image Analysis Integrated into Image & Structure Interpretation Module for WellCAD (v5.1) by Advanced Logic Technology (ALT), 2015
3. Validation of Geologists Logging Rio Tinto Iron Ore (2015-2017) Automated logging Validation Assistance (AVA) RC chips from exploration boreholes routinely logged Holes logged in intervals as compositions of material types Mineralogy and texture (e.g. BIF20 GOE50 GOL20 SHL10) Material Type Classification Scheme (MTCS) (Box et al., 2002) allows for each material type theoretical assay, hardness etc Validation process corrects logged composition against laboratory assays Essential for accurate grade estimation and managing risk in blending strategy Not just a mathematical optimization problem as it requires a geologically acceptable solution
Geologically Informed Solution Machine learning from expert validators & legacy data: o Learning the process - common material type swap patterns depending on logging errors (including commonly confused materials in which geological/geochemical scenarios) o Commonly occurring combinations of material types Geologically informed material swaps and substitutions combined with mathematical optimization Propose validated compositions with confidence value geologist makes the decision! Machine learning of the patterns of material type swap from expert validators
Auto-Validation Assistant Logged Composition Trained Material Type Swaps depending on the assay error Material Type Modification based on current assay error and stratigraphy Iterate till stabilise (no change in best composition) Optimisation of Material Type Percentages minimising assay error, change in lump%, hardness Intermediate Penalty Calculation based on stratigraphy, conflicting material types, texture etc Trained Association Rules for material type combinations Final Penalty Calculation based on chip shape, colour, association rules Proposed Validated Compositions with confidence value
Evaluation AVA used to validate 1996 intervals - 74.3% accepted without change Deposit scale validation (14,600 intervals) o Compared the errors for logged composition, manually validated composition and top-ranked AVA composition o Results compared in terms of Error tolerance factor which is the proportion of allowed error for that assay element (i.e. error tolerance factor of 1 means it has the maximum allowable error for that assay element) DET: Detritals intervals MB: Mineralised bedded intervals SH: Shale intervals ALL: All intervals
4. Kevitsa Mine Data Analysis Sponsored by First Quantum Pty Ltd, PhD project of Tom Horrocks (2015-on-going) Kevitsa Ni-Cu-PGE deposit is a data-rich natural laboratory: o 3D Geophysical Inversion models: density, mag susc., conductivity, velocity o Wireline logging; core measurements o Geochemical assays of drill core o Geological logs Analysis techniques for integration: 3D boundary detection for petrophysically distinct rock units (above) Dimensionality reduction of geochemical dataset for domaining and characterising alteration (next)
Alteration Embedding A dimensionality reduction technique, t-distributed stochastic neighbour embedding (t-sne), creates excellent 2D representations of Kevitsa s geochemical assays (n = 16 165). Embedding (above), which uses 11 elements relevant to alteration, chemically distinguishes between altered and unaltered samples. Can interpret geochemical process of host rock hydration (next slide)
Alteration Embedding Hydration Chemical transition of olivine pyroxenite (bottomright of the main cluster) to metaperidotite (top-left of the main cluster) associated with hydration process can be observed such as decreasing magnesium & manganese
Summary Data science solutions for geological interpretation can and should utilise geoscientists knowledge and intuition if appropriate Innovative algorithms are integrated into a workflow and interactive visualisation is important for practical decision support interpretable data science For industry, identify bottlenecks in the data workflow and find solution in data science (building blocks) a gradual & sure adaptation of data science Acknowledgements: Geological Survey of Western Australia; Australian Research Council (LP140100267); Rio Tinto Iron Ore; First Quantum Minerals