b'Education matters The top of D region characterisesof 10 km from synthetic experimenttechniques are a logical addition to the a velocity discontinuity - the Ddemonstrate the feasibility and effectivitynumerous possible approaches that can discontinuity, which likely originates fromof applying grid search method tobe applied to the search for ore deposits.the phase transition from bridgmaniteinvestigate the 1-D shear wave velocity (Bm) to post-perovskite (pPv) at thestructure in the D region. The three core research chapters in pressure-temperature conditions of theIn order to probe the global existence andthis thesis develop the application of lowermost mantle. This discontinuityorigin of the D discontinuity, we applymachine learning in the context of has been widely observed in a fewthe grid search method to the lowermostmineral exploration. Emphasis is placed regions, while much of the Earth ismantle beneath Alaska, northern Pacificon the Random Forests algorithm still unmapped. This poses a challengeand Central America. Our detailed seismicfor mapping lithology in a range of in deciphering the origin of the Dimaging studies allow for speculationsettings and at a variety of stages in discontinuity and resolving the debatethat D layer is caused by both thermalthe exploration process. Information whether it is a global feature. and chemical effects. Beneath Centralentropy is used to assist both in assessing America, the D layer is dominantlyand communicating any complex This thesis presents a compilation ofcaused by thermal effects associated withcombinations, and potential inaccuracy, published, accepted and submittedFarallon slab remnant, but also affectedof classification results. Through the work. The main objective is to provideby varying chemical compositions inthesis, methods are employed with new insights on the origin of the Dthe lowermost mantle. By contrast, infuture practical usage in mind, such region and its relationship with mantleregions such as northern Pacific andthat machine learning may be used by dynamics. There are three chapterseastern Alaska, subduction-relatedthe geologist (as domain expert) in an of the thesis. Chapter I contributeschemical heterogeneities are likely toobjective manner.to probing the global existence ofplay a dominant role to explain the sharpThe first of these core studies uses the D discontinuity by adding newvelocity change in adjacent areas. Thisthe Random Forests algorithm to re-observations in unexplored areas ondifference in the D origin could be tightlyclassify the solid geology lithology map the Earth. Chapter II focuses on thelinked to the slab history in differentof the Heron South project, located development of a new method to fillregions. This thesis shed a light on ain the Eastern Goldfields of Western in the gaps of the existing methods.better understanding of the D origin andAustralia. This study uses geophysical Chapter III is dedicated to investigatingits associated mantle dynamics. and remote sensing data, in the the potential origin of the D region by applying this new method to the DSteve Kuhn, University of Tasmania:absence of geochemical samples and region beneath Alaska, Northern PacificMachine learning for mineral exploration:geological ground truthing with most and Central America. Prediction and quantified uncertainty atof the project under transported cover. multiple exploration stages. This is characteristic of an early stage, To explore the distribution of the Dreconnaissance exploration project. A discontinuity in the globe, we first addsparse training sample of 1.6 percent of a new observation of D discontinuitythe total area, is taken as training data, beneath the Central Atlantic withallowing much of the areas geology the new data that greatly improved priorfreedom to be reclassified. This study poor data coverage. Through forwarddemonstrates that Random Forests, waveform modelling, we observe awith proper consideration given to D discontinuity with varying depthsampling and training data selection, and sharpness along a high velocitycan be used effectively to produce or corridor from north to south beneath theimprove geological mapping in little Central Atlantic. Our results also provideexplored areas. Information entropy is evidence on the potential origin of theshown to be valuable in predicting where D discontinuity, which is attributed toclassification was likely to be inaccurate a Bm-pPv phase transition within bothor a region highly complex.pyrolitic mantle and slab debris beneath Central Atlantic. The second core study uses Random Forests to produce a solid geology map To address some of limitations ofof the Kliyul porphyry prospect of British the current methods on imaging theColumbia, Canada, using a fusion of detailed seismic structures in theavailable geophysical and geochemical D region, we develop a grid-searchdata, typical of a greenfields stage scheme to constrain the detailed 1-Dexploration project. Soil and rock chip shear-wave velocity structure in the Dsample sites were taken as training region with quantitative assessmentMachine learning describes an array ofdata, used to classify the remainder of the uncertainty of 1D models. Thiscomputational and nested statisticalof the project area. Assessment of the new method addresses problems ofmethods whereby a computer can learnprobability distributions produced using non-uniqueness and subjectivity in theand subsequently make predictionsthe Random Forests algorithm enabled conventional trial-and-error waveformor identify patterns in data. With theregions with an elevated probability of modelling method which evaluatesincreasing volume and variety ofintrusions (a key indicator lithology) to the fit between data and syntheticsnumerical data in the geosciences, andbe mapped, even where not observed qualitatively. A good recovery of thewidespread availability of the neededin training data. The results of this study input model with a vertical resolutioncomputing power, machine learninghighlight the value of a soft, ensemble 37 PREVIEW DECEMBER 2021'