Applied Computing and Geosciences最新文献

Understanding the variations in physical and mechanical behavior of rock materials due to progressive weathering is vital to carry on time and cost-effective engineering projects. Up to date, soft computing algorithms have been established to quantify the weathering degree (WD) of various rocks due to better prediction performance and problem-solving capability. However, the complexity of the weathering process does not allow the use of a single weathering quantification model for a wide range of rock types. Therefore, this study aims to provide a practical, quantitative, and effective framework for predicting the WD of andesitic rocks. To fulfill the aims of this study, a wide range of cases were collected from the previous studies to establish a predictive model based on dry unit weight (γ_d), effective porosity (n_e), and uniaxial compressive strength (UCS). Consequently, a combined application of fuzzy inference system (FIS) and artificial neural network (ANN) was introduced to assess the WD of the investigated andesitic rocks. The WD ratings were presented as four different weathering classes (from fresh (W₀) to highly weathered (W₃)). Since most soft computing algorithms are black-box models that cannot be efficiently utilized in any other study, an explicit neural network formulation was firstly developed for WD prediction in this study. As a result, the proposed formulation will provide a practical and straightforward assessment of WD for andesitic rocks. However, to improve the reliability and consistency of the proposed model, different datasets should be used in the explicit neural network formulation proposed.

Inferring underground porosity and evaluating its spatial distribution is of great significance in a wide range of Earth sciences and engineering, including hydrocarbon reservoir characterization and geothermal energy exploitation. Popular methods are largely based on the analysis of lithological cores, well logs, and seismic inversion. These methods are reliable, but they are still time-consuming, expensive, and difficult to conduct. In addition, seismic inversion confronts problems of nonlinearity and has multiple solutions. However, deep learning (DL) can provide a more flexible, efficient, and accurate capability, mapping directly from acoustic impedance and lithofacies data to porosity. To prove the point, in this paper, we trained an ensemble of DL models and then proposed a weight combination of every single trained model’s strength to improve the result. We evaluated the method's reliability using a number of metrics. Further, we compared it with traditional ones. The weighted ensemble resulted in a lower error than the simple ensemble and the single model. Its spatial distribution map showed the best connectivity with that of historical porosity. Finally, we tested our method's effectiveness using a dataset that was used in a previously published study. Our method improved the prediction of the latter.

Multiple-point statistics algorithms allow modeling spatial variability from training images. Among these techniques, the Direct Sampling (DS) algorithm has advanced capabilities, such as multivariate simulations, treatment of non-stationarity, multi-resolution capabilities, conditioning by inequality or connectivity data. However, finding the right trade-off between computing time and simulation quality requires tuning three main parameters, which can be complicated since simulation time and quality are affected by these parameters in a complex manner. To facilitate the parameter selection, we propose the Direct Sampling Best Candidate (DSBC) parametrization approach. It consists in setting the distance threshold to 0. The two other parameters are kept (the number of neighbors and the scan fraction) as well as all the advantages of DS. We present three test cases that prove that the DSBC approach allows to identify efficiently parameters leading to comparable or better quality and computational time than the standard DS parametrization. We conclude that the DSBC approach could be used as a default mode when using DS, and that the standard parametrization should only be used when the DSBC approach is not sufficient.

A deeper understanding of the drivers of evapotranspiration and the modelling of its constituent parts (evaporation and transpiration) may be of significant importance to the monitoring and management of water resources globally over the coming decades. In this work a framework was developed to identify the best performing machine learning algorithm from a candidate set, select optimal predictive features and rank features in terms of their importance to predictive accuracy. The experiments conducted in this work used 3 separate feature sets across 4 wetland sites as input into 8 candidate machine learning algorithms, providing 96 sets of experimental configurations. Given this high number of parameters, our results show strong evidence that there is no singularly optimal machine learning algorithm or feature set across all of the wetland sites studied despite their similarities. At each of the sites at least one model was identified that improved on the predictive performance of our baseline. A key finding discovered when examining feature importance is that methane flux, a feature whose relationship with evapotranspiration is not generally examined, may contribute to further biophysical process understanding. This work demonstrates the applicability of a machine learning framework for evapotranspiration partitioning that is independent of domain knowledge, producing improved models for partitioning and identifying new and useful predictive features.

Rock glaciers (RG) are landforms that occur in high latitudes or elevations and — in their active state — consist of a mixture of rock debris and ice. Despite serving as a form of groundwater storage, they are an indicator for the occurrence of (former) permafrost and therefore carry significance in the research for the ongoing climate change. For these reasons, the past years have shown rising interest in the establishment of RG inventories to investigate the extent of permafrost and quantify water storages. Creating these inventories, however, usually involves manual, laborious, and subjective mapping of the landforms based on aerial image - and digital elevation model analysis. We propose an approach for RG mapping based on supervised machine learning which can help to increase the mapping efficiency and permits rapid RG mapping in vast and not yet covered areas. We found deep convolutional artificial neural networks (ANN) that are specifically designed for image segmentation (U-Net architecture) to be well suited for this classification problem. The general workflow consists of training the ANNs with orthophotos and slope maps of digital elevation models as input. The output (RG label-maps) is derived from a recently published RG inventory of the Austrian Alps that features 5769 individual RGs and was compiled manually by several scientists. To increase the generalization capabilities, we use live data augmentation during training. Based on this inventory, the ANNs have learned the average expert opinion and the RG map generated by the ANN can be used to increase the consistency and completeness of already existing RG inventories. Moreover, this ANN approach might be valuable for other landform mapping tasks beyond rock glaciers (e.g., other mass movements).

High resolution characterization of sub-surface geology is critical to improving the performance of reservoir models in fluid flow and reactive transport simulation studies in the fields of groundwater, CO₂ geo-sequestration and oil and gas research. The modern improvements in wireline logging technology allow for the deduction of depth continuous records of individual rock properties at cm-scale resolution. However, to the best of the authors’ knowledge, no method exists to obtain high-resolution lithotype logs. Traditional methods for rock typing are based on core sample analysis and are therefore limited to discrete depth. The Kimeleon colourlith log output is based on the combination of wireline logs and is probably one of the few tools which creates continuous lithotype logs at high resolution primarily for the purpose of visualization. There is currently no automated method to transform the colourlith images into quantitative rock type logs which can be used as an input for reservoir modelling software. Additionally, colourlith logs are based solely on wireline information and do not incorporate local lithological information from core samples. This study addresses these issues by combining discrete core sample data with continuous colourlith logs. A code (Irida) has been developed to enhance the usability of the Kimeleon software by transforming colourlith logs into high resolution lithotype logs of rock types identified in core plugs. The code takes the colourlith image log as an input along with the names and properties of site-specific rock types identified and measured in discrete core samples. The code then uses k-means clustering, an unsupervised machine learning method, to compute a high-resolution log of rock types and their properties which are continuous with depth. The properties of interest include porosity, anisotropic absolute and relative permeabilities and capillary pressure curves. The code significantly reduces the efforts required for the preparation of lithotype logs.

This presents a novel hybrid 24-h forecasting model of convective weather events based on numerical simulation and machine learning algorithms. To characterize the convective events, 13-year from 2008 up to 2020 of precipitation data from the main airport stations in Rio de Janeiro, Brazil, and atmospheric discharges from the surrounding area of around 150 km are investigated. The Weather Research and Forecasting (WRF) model was used to numerically simulate atmospheric conditions for every day in February, as it is the month with the greatest daily rate of atmospheric discharge for the data period. The p-value hypothesis test (with $\alpha =0.05$) was applied to each grid point of the numerically predicted variables (defined as an independent attribute) to find those most associated with convective events using the output of the 3-D WRF grid. This one identified 36 attributes (or predictors) that were used as input in the machine learning algorithms' training-test process in this study. Several cross-validation training and testing experiments were carried out using the nine-selected categorical machine learning algorithms and the 36 defined predictors. After applying the boosting technique to the nine previously trained-tested algorithms, the results of the 24-h predictions of convective occurrences were deemed satisfactory. The RandomForest method produced the best results, with statistics values close to perfection, such as POD = 1.00, FAR = 0.02, and CSI = 0.98. The 24-h hindcast utilizing the nine algorithms for the 28 days of February 2019 was very encouraging because it was able to almost recreate the maturation phase of events and their eventual failures were noted during the formation and dissipation phases. The best and worst 24-h hindcast had POD = 0.97 and 0.88, FAR = 0.02 and 0.12, and CSI = 0.94 and 0.78, respectively.

Machine learning (ML) has been a technique employed to build data-driven models that can map the relationship between the input and output data provided. ML-based data-driven models offer an alternative path to solving optimization problems, which are conventionally resolved by applying simulation models. Higher computational cost is induced if the simulation model is computationally intensive. Such a situation aptly applies to petroleum engineering, especially when different geological realizations of numerical reservoir simulation (NRS) models are considered for production optimization. Therefore, data-driven models are suggested as a substitute for NRS. In this work, we demonstrated how multilayer perceptron could be implemented to build data-driven models based on 10 realizations of the Egg Model. These models were then coupled with two nature-inspired algorithms, viz. particle swarm optimization and grey wolf optimizer to solve waterflooding optimization. These data-driven models were adaptively re-trained by applying a training database that was updated via the addition of extra samples retrieved from optimization with the proxy models. The details of the methodology will be divulged in the paper. According to the results obtained, we could deduce that the methodology generated reliable data-driven models to solve the optimization problem, as justified by the excellent performance of the ML-based proxy model (with a coefficient of determination, R² exceeding 0.98 in training, testing, and blind validation) and accurate optimization result (less than 1% error between the Expected Net Present Values optimized using NRS and proxy models). This study aids in an enhanced understanding of implementing adaptive training in tandem with optimization in ML-based proxy modeling.

Domains define the boundaries of mineralisation zones, within which the grade distribution of the target minerals can be quantified via an established mineral resource estimation procedure. Available domain modelling techniques include manual interpretation, implicit modelling and advanced geostatistical approaches. In mining applications, the most commonly used method is manual domaining, which is labour-intensive and prone to subjective judgement errors. In addition, the output is largely deterministic and ignores the significant uncertainty associated with the domain interpretation and boundary definitions. There is, therefore, a need for a more objective framework that can automatically define mineral domains and quantify the associated uncertainty. This paper describes such a framework, which consists of a hybrid approach based on simulated grade distributions and a machine learning (ML) classification technique, XGBoost, trained on lithological properties. Data from an Iron Oxide Copper Gold (IOCG) deposit are used as a case study to demonstrate the application of the proposed method. The study shows that the approach can handle complex multi-class problems with imbalanced features, and it can quantify the uncertainty of domain boundaries. A noise filtering method is used as a pre-processing step to improve the performance of the ML classification, especially in the case of problematic classes where domain boundaries are difficult to predict due to the similarity in geological characteristics.