The recent use of spatial coordinate features in multilayer perceptron (MLP) neural networks provides opportunities for novel applications in potential field geophysics. So-called coordinate MLP networks allow for learning a representative function of potential fields from their surveyed samples. We present a novel method for implicit neural representation of potential fields, demonstrate the quality of the learned implicit function by encoding synthetic and real airborne geophysical survey line data, and compare the result to grid data processed with traditional gridding methods. We further demonstrate the analytical calculation of gradients directly in the continuous domain of the neural network using automatic differentiation, with the same framework used to train the neural network representation. A regular grid created with the proposed method closely matches the ground truth reference synthetic forward model, with a root mean-square error of 10.3 nT, compared to 18.75 nT for minimum curvature. Horizontal gradients calculated with this method are accurate against numerically derived gradients, while the vertical gradient is poor for these case study data. The training process is rapid, and only requires recorded samples from a single survey extent.

Observations of geological structures are often made at different scales and often can cross multple orders of magnitude. This attribute of scale though is often not explicitly incorporated into the workflow of geological modeling and is usually treated as data preparation or sampling bias. The spectral properties of the discrete Laplacian operator, when applied to reconstructed surfaces from implicit modeling though offer a potential means of bridging this gap, when also combined with appropriate directional statistical anaysis. We present an example of how bedding orientation measurements from a 1:5000 scale surface map and drillhole bedding orientation picks from borehole televiewer images can be integrated using the manifold harmonics of the Laplacian operator and a mixture of von-Mises Fisher probability distributions. This provides automated insights for sampling for modeling and also possible kinematic and tectonics processes.

This paper presents the use of Paired Hierarchical Variational Autoencoders (PHVAEs) for image-to-image translation and cross reconstruction. We explore a novel method for constructing and applying HVAEs for these types of tasks, and compare this model against existing GAN, transformer, and diffusion based methods, including pix2pix, StegoGAN, ResViT, and BBDM. The method utilizes deep, hierarchical VAEs with paired optimization to take advantage of the bidirectional latent space. The proposed model outperforms all alternative models in PSNR on all tested tasks, and attains consistently competitive SSIM, demonstrating its effectiveness and benefit for prediction-focused translation tasks such as cross reconstruction, while remaining parameter efficient.

In iron ore exploration, mineral samples are obtained from a drill hole and their compositions logged as proportions of material types, which encapsulate mineralogical and textural attributes. This logging is used to inform the properties, i.e. mineralogical, grade, and handling, of the orebody during mining and processing. Logging has historically been recorded manually, and as such is subjective and time consuming.We have developed a system that uses objective datasets, namely Fourier Transform Infrared reflectance spectra, geochemical assays, and a priori geological constraints, within a machine learning system to objectively log these mineral samples. The system is trained on samples logged by expert geologists in a laboratory environment. The resultant objective logging optimally satisfies constraints including matching the theoretical chemistry (calculated from the logging) to the laboratory assays, and satisfying the sample’s known geological context.We present comparisons of the logged material type proportions recorded by expert geologists in a laboratory environment with the proportions produced by the objective logging system, and also against geologists’ field logging proportions of the same samples. These comparisons demonstrated that the objective logging system’s proportions closely matched the geochemically validated expert-logged proportions for important individual material types, and over larger mineralogical groupings, and groups of material types of different textural hardness. Additionally, the system outperformed field logging, relative to laboratory-based logging, according to r2 and RMSE metrics. These results demonstrate an improvement over geologists’ field logging, while removing the subjectivity inherent in the human logging process.

Fourier Transform infrared spectroscopy (FTIR) is an emerging cost effective and rapid mineralogical charac-terization technique being applied in the geosciences. Detecting anomalous FTIR spectra is especially relevant to the geoscience domain, as it may indicate abrupt changes in geology or mineralogical composition of the rock sample being examined. Given a large volume of data, detecting anomalies that exhibit significant and abrupt spatial and compositional variability is a time-consuming and challenging task. This paper explores the use of an unsupervised variational autoencoder (VAE) for determining anomalies that may exist within a set of FTIR spectra collected from reverse circulation (RC) drill chip samples spanning several iron ore deposits from the Pilbara region in Western Australia. Diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) were measured from 1,579 two-metre composite samples. Our results showed that the VAE was effective in separating anomalous spectra from spectra typical of unmineralized banded iron formation by leveraging the probabilistic latent representation of the spectra in as few as two latent dimensions. To validate our results, detected anomalous samples were compared with their respective geochemical assays to analyse their mineralogical differences, which may have led to the anomalous spectra. In the iron ore sample data used in this study, the observed spectral anomalies were shown to have elevated concentrations of Al2O3 and TiO2 wt.% while being several standard deviations below the mean Fe2O3 wt.% indicating mineralogies rich in shale as opposed to iron oxide rich mineralogies. While the paper demonstrates the efficacy of the VAE in anomaly detection, it can also be effective in assuring the quality of the FTIR data as a pre-processing step, which is critically important for machine learning applications.

Densely sampled geophysical surveys are a key driver for mineral exploration, but sample density, and therefore grid resolution, is limited by survey cost. Consequently, computational methods are resorted to for upsampling, or 'super-resolving', of gridded geophysical survey data. However, existing approaches such as interpolation filters do not leverage high-resolution detail from pre-existing geophysical surveys. Through the application of state-of-the-art deep learning super-resolution architectures, accurate high-resolution grids are predicted from corresponding low-resolution grid priors. Specifically, the RDN and ESRGAN+ neural network architectures, which were originally developed for enhancing images, are applied to enhance magnetic surveys and trained with high-resolution and low-resolution magnetic grids of the same extent. 80 m cell size magnetic grids are upsampled to 20 m cell size using this method, and predicted value and structural accuracy comparisons to the corresponding ground truth 20 m grids are presented over two test sites in the Eastern Goldfields Superterrane, Western Australia. The method based on RDN achieves 53 % lower error compared to Bicubic interpolation. The case study demonstrates that the deep learning approach can improve the resolution and frequency content of geophysical surveys without requiring additional sampling expense, while remaining accurate against known ground truth surveys. Super-resolution may assist in the interpretation of low-resolution survey grids, however these upsampled grids cannot perfectly recreate the accuracy of highly sampled surveys gridded at their optimal cell size.

Accurate basement delineation is crucial for mineral deposits where formation is associated with a paleobasement topography, such as in detrital iron ore deposits. In such environments seismic surveys provide useful data to assist the mapping of the basement interface. A key stage of the process is identification of the first arrival of elastic waves at a recording station, or so-called 'first break'. Although a variety of automated first break detection algorithms were proposed previously for use in different seismological environments, first break detection in hard rock environments remains challenging due to low amplitude arrivals and the presence of non stationary noise within the seismic traces. This study explores the use of three types of supervised neural networks for first break detection for a seismic refraction dataset of 15,000 traces, over a detrital iron ore deposit. A fully connected neural network (NN), a Convolutional Neural Network (CNN), and a Long-Short Term Memory recurrent neural network (LSTM) are evaluated and their predicted first breaks are compared to manually picked first breaks and predictions from the widely used Coppens' method. The results showed that the CNN and LSTM were particularly effective for the first break detection, achieving mean square errors of 0.051 ms (with a statistical bias of 2.2 ms) and 0.058 (with a statistical bias of 21.2 ms) respectively against manual interpretation, compared to -35 ms for Coppens' method and x ms for the NN-122 ms.

The interaction of geological structures at different scales often results in complex three-dimensional geometries. This can especially be the case when interpreting progressively deformed multi-layer fold systems which can result in complex non-cylindrical geometries of varying scale and congruity. Curvature analysis methods have lately been developed to assist this for analysing three-dimensional geological surfaces which do not need to assume cylindricity, but these methods are often applied to three-dimensional seismic horizons which have specific considerations that can affect scale analysis. We show how the application of manifold harmonic analysis together with implicit geological modelling can assist in the interpretation of fold structures. This is demonstrated by developing a three-dimensional model of the Padtherung Syncline, central Hamersley province, Western Australia, and applying our manifold harmonic analysis methods. Our results show how the analysis can be used as evidence to re-interpret the polyclinal structure as progressively deformed rather than a superposed structure.

3-D geochemical subsurface models, as constructed by spatial interpolation of drill-core assays, are valuable assets across multiple stages of the mineral industry's workflow. However, the accuracy of such models is limited by the spatial sparsity of the underlying drill-core, which samples only a small fraction of the subsurface. This limitation can be alleviated by integrating collocated 3-D models into the interpolation process, such as the 3-D rock property models produced by modern geophysical inversion procedures, provided that they are sufficiently resolved and correlated with the interpolation target. While standard machine learning algorithms are capable of predicting the target property given these data, incorporating spatial autocorrelation and anisotropy in these models is often not possible. We propose a Gaussian process regression model for 3-D geochemical interpolation, where custom kernels are introduced to integrate collocated 3-D rock property models while addressing the trade-off between the spatial proximity of drill-cores and the similarities in their collocated rock properties, as well as the relative degree to which each supporting 3-D model contributes to interpolation. The proposed model was evaluated for 3-D modelling of Mg content in the Kevitsa Ni-Cu-PGE deposit based on drill-core analyses and four 3-D geophysical inversion models. Incorporating the inversion models improved the regression model's likelihood (relative to a purely spatial Gaussian process regression model) when evaluated at held-out test holes, but only for moderate spatial scales (100 m).

The Darling Range in Western Australia is a major bauxite producing region. Clearing, excavation and rehabilitation activities related to bauxite mining have influenced land cover within this region since mining commenced in the 1960s. This paper presents a study that used machine learning and time series visualisation to analyse the land cover changes of the Darling Range using time-lapse multispectral images, with the aim of understanding the impact of the mining activities and land rehabilitation patterns of the region. Land cover changes were analysed using 14 Landsat Thematic Mapper (TM) images between 1988 and 2014. The spatial distribution of land cover was classified automatically using machine learning algorithms, and their temporal changes were visualized for analysis. Supervised classification was carried out based on the six spectral features of the images using three machine learning algorithms, namely support vector machine (SVM), random forest (RF) and Naive Bayes (NB). The results showed that the RF algorithm achieved overall accuracy, (ratio of correctly classified samples divided by the total number of samples), greater than 95% for all years. The temporal changes of land cover distribution over the study period were visualized using a change map. These changes were compared with land clearing and rehabilitation records of a mining company operating in that region. A close correlation was observed between the automated analysis outputs and the company's records. This work demonstrates the potential use of machine analysis in improving the accuracy of spatial data related to land clearing; and in monitoring vegetation recovery of closed and rehabilitated mines.

SummarySeismic data processing and analysis focuses on identifying the arrival of seismic waves or ‘first-breaks’. The identification of the arrival of first breaks is complicated by the variance of recording quality typically found across the dataset. In an exploration setting, models need to be developed and refined multiple times. Picking these first breaks then becomes time consuming, limiting the interpreter to processing their dataset rather than considering the implications of their model. Machine Learning as a field continues to respond to many data centric issues within geoscience. However, the field as a whole continues to grapple with balancing the power of these new techniques against operator expertise and skill.This paper presents a methodology to identify the first break in seismic refraction data using a Long-Short Term Memory (LSTM) network, which is a recurrent network architecture. I propose one way to delineate between different groups of traces that the operator would intuitively pick differently, by using dynamic time warping to generate a distance matrix of the seismic traces for clustering. This clustering of trace types allows for a more targeted selection of training samples. I conclude with a proposed framework for the integration of operator skill with machine learning speed and repeatability.

In minerals exploration, routine drilling is performed and the data logged from these drillholes, including lithological composition, assays, and downhole geophysical measurements such as natural gamma logs, are used to create geological interpretations of the strata within each drillhole. A 3D geological model can be created by identifying corresponding stratigraphic boundaries within multiple drillholes. These models can be used for understanding the formation and the mineral endowment of a deposit. We introduce a system for producing stratigraphic interpretations of iron ore exploration drillholes in the Pilbara region in Western Australia. The algorithm firstly classifies each data modality independently for each geological interval, for example 2 m, with classification results for each stratigraphic unit as output. These classifiers, for geological logging, assays, gamma logs, were trained on historical datasets over a wide range of strata in the Pilbara. The influence of each classifier can be adjusted according to the user's preference, and a novel optimisation algorithm incorporates known geological features such as dykes, faults and thicknesses of various stratigraphic units, to objectively create the best fit interpretation of the geology. A geologist can then adjust this interpretation to include local knowledge. Manual interpretations of 396 drillholes from a high-grade iron ore deposit are compared to interpretations of the same hole prepared by the algorithm. Analysis of interval-by-interval interpretations, and basement geology demonstrate that without any human input, similar interpretations are produced while reducing manual effort.

Dimensionality reduction provides a simple, two-dimensional representation of multi-element geochemical assays, which facilitates visualisation of complex data and enhances their interpretation. A recently proposed dimensionality reduction algorithm, namely t-distributed stochastic neighbour embedding (t-SNE), generates effective two-dimensional representations of a wide range of datasets based on pairwise statistical distances of the input. However, direct application to multi-element geochemical assays has been shown to produce representations which can fail to separate specimens by a desired geological property, such as state of hydration. Since t-SNE is a statistical distance-based method, these sub-optimal representations may be due to the presence of dimensions (i.e., elements) irrelevant to the desired property-an issue often termed the 'curse of dimensionality'. To address this shortcoming, t-SNE was applied to (i) 31 elements in a geochemical assay database covering 16 000 drill core intervals intersecting the Kevitsa mafic-ultramafic intrusion (Lapland, Finland); and (ii) a subset of 11 elements capable of discriminating between unaltered and altered host rock specimens, as determined by a Random Forest classifier within a recursive feature elimination framework. The resulting representation more effectively separates altered and unaltered specimens, and we demonstrate that it produces more favourable representations than alternative well-known methods (namely, a self-organising map and principal components analysis) applied to the same dataset. We also demonstrate that the proposed t-SNE representation is applicable for re-logging of the specimens' alteration state as logged by geologists, and in particular provides visual insight into the labels suggested by a black box statistical re-logging algorithm.

Records of past exploration in open-file mineral exploration reports are an important source of information for mineral explorers. These reports document existing geological knowledge that may be relevant to modelling ore forming processes in a particular area of interest. This paper presents the development of GeoDocA, a geological document analysis system, that applies automated text analysis techniques with the specific aim of assisting geologists in browsing of and searching for documents based on relevant geological contents within a large repository of documents. GeoDocA analysed 25,419 exploration reports and using a customised set of keywords pertaining to broad categories such as mineral occurrences, rock types, alteration types, and geological time. An interactive user interface was developed to facilitate visual analysis of exploration reports. For individual reports, it provides a summary of their content in graph form, a gallery of extracted figures and tables, and a list of similar reports based on shared geological keywords. In addition, it assists document search efforts through auto-generated keyword suggestions which are based on associations of keywords learnt by the system from all reports in the repository. While the text mining methods reported here is the foundation for further development to incorporate semantic analysis towards geological knowledge extraction, the outcomes of this study demonstrate the effectiveness of automated text analysis in supporting a fast analysis of a large number of reports to identify the targeted mineral systems and their associated geological environments.

SummaryIn minerals exploration, routine drilling is performed and the data logged from these drillholes, including lithological composition, assays, and downhole geophysical measurements such as natural gamma logs, are used to create geological interpretations of the strata within each drillhole. A 3D geological model can be created by identifying corresponding stratigraphic boundaries within multiple drillholes. These models can be used for understanding the formation and the mineral endowment of a deposit.We introduce a system for producing stratigraphic interpretations of iron ore exploration drillholes in the Pilbara region in Western Australia. The algorithm firstly classifies each data modality independently for each geological interval, for example 2m, with classification results for each stratigraphic unit as output. These classifiers, for geological logging, assays, gamma logs, were trained on historical datasets over a wide range of strata in the Pilbara. The influence of each classifier can be adjusted according to the user’s preference, and a novel optimisation algorithm incorporates known geological features such as dykes, faults and thicknesses of various stratigraphic units, to objectively create the best fit interpretation of the geology. A geologist can then adjust this interpretation to include local knowledge.Manual interpretations of 396 drillholes from a high-grade iron ore deposit are compared to interpretations of the same hole prepared by the algorithm. An interval-byinterval comparison of these interpretations demonstrates that without any human input, similar interpretations are produced while reducing manual effort.

Geophysical inversion can produce 3D models of the subsurface's physical properties. The smoothness of property variations in these models makes it challenging to automatically find boundaries of homogeneous regions, where these boundaries may have implications for petrophysical transition and are significant for geologic interpretation. We have developed a new boundary detection technique that nonparametrically identifies and subtracts homogeneous regions from the 3D model, leaving exposed edges. The method is based on kernel density estimation of local property variations, in which the number of modes in the kernel density estimates (i.e., the local mode cardinality [MC]) is used to identify edge voxels within the model. Two edge detection operators were developed: one using local values exclusively and the other incorporating the spatial distribution of local values into the local MC, which is more sensitive to local variations. To assist in the geologic interpretation, continuous boundary surfaces are generated from the identified edge voxels using a gradient-based 3D image morphological operation. The technique was evaluated on synthetic rock property models and effectively identified the lithologic boundaries, even with non-Gaussian noise. The proposed operators were also applied to visualize edges in density, conductivity, magnetic susceptibility, and seismic tomography models of the Kevitsa Ni-Cu-PGE deposit (Lapland, Finland) generated by geophysical inversion. The boundaries detected via the proposed technique can be used for visualization and may be useful in further geostatistical computation due to their statistical foundation.

Logging of exploration drillholes is a routine practice and its accuracy is essential for resource evaluation and planning in the minerals industry. Logged compositions record a set of material types with standardized mineralogy and texture characteristics. The material types logged may vary due to diversities in mineralization and geology, but also due to subjective biases and human error, leading to significant challenges for the industry. Thus, there is a need to validate the field logging whereby the material types and their percentages are adjusted to reconcile with laboratory assay values, while retaining the physical characteristics and geological context. We introduce the Auto-Validation Assistant (AVA) algorithm that applies data mining methods to geologists’ validation patterns recorded in a training process over hundreds of intervals of iron ore exploration drillholes. The AVA modifies the material types selected in the logged composition and their percentages according to geological rules learned in the training process, and proposes to the geologist a number of validated compositions with optimized geochemistry and mineralogical hardness, while also considering visible properties such as chip shape and color. Using the confidence value provided with each validated composition, the geologist can make informed validation decisions and remains in control of the validation process, while harnessing computational power. Experiments were conducted to evaluate the auto-validated compositions generated by AVA: one to analyze the acceptance rate of the AVA generated compositions by geologists for 1,996 intervals in drillholes from different sites; and the other to compare manual and AVA-validated compositions using 14,600 drillholes from one entire deposit. The results showed the acceptance rate of AVA-validated compositions (without further change) of 74.3%, leading to significant time savings over tedious manual validation, while demonstrating that AVA provides comparable but more consistent results. The algorithm is fast and repeatable and can be adapted to different material types and training datasets, with potential applications beyond iron ore exploration.

Downhole acoustic and optical televiewer images, and formation microimager (FMI) logs are important datasets for structural and geotechnical analyses for the mineral and petroleum industries. Within these data, dipping planar structures appear as sinusoids, often in incomplete form and in abundance. Their detection is a labour intensive and hence expensive task and as such is a significant bottleneck in data processing as companies may have hundreds of kilometres of logs to process each year. We present an image analysis system that harnesses the power of automated image analysis and provides an interactive user interface to support the analysis of televiewer images by users with different objectives. Our algorithm rapidly produces repeatable, objective results. We have embedded it in an interactive workflow to complement geologists' intuition and experience in interpreting data to improve efficiency and assist, rather than replace the geologist. The main contributions include a new image quality assessment technique for highlighting image areas most suited to automated structure detection and for detecting boundaries of geological zones, and a novel sinusoid detection algorithm for detecting and selecting sinusoids with given confidence levels. Further tools are provided to perform rapid analysis of and further detection of structures e.g. as limited to specific orientations.