Computational Geosciences最新文献

Identifying the source of an oil spill is an essential step in environmental forensics. The Bayesian approach allows to estimate the source parameters of an oil spill from available observations. Sampling the posterior distribution, however, can be computationally prohibitive unless the forward model is replaced by an inexpensive surrogate. Yet the construction of globally accurate surrogates can be challenging when the forward model exhibits strong nonlinear variations. We present an iterative data-driven algorithm for the construction of polynomial chaos surrogates whose accuracy is localized in regions of high posterior probability. Two synthetic oil spill experiments, in which the construction of prior-based surrogates is not feasible, are conducted to assess the performance of the proposed algorithm in estimating five source parameters. The algorithm successfully provided a good approximation of the posterior distribution and accelerated the estimation of the oil spill source parameters and their uncertainties by an order of 100 folds.

This work proposes a novel meshing technique that is able to extract surfaces from processed seismic data and integrate surfaces that were constructed using other extraction techniques. Contrary to other existing methods, the process is fully automated and does not require any user intervention. The proposed system includes an approach for closing the gaps that arise from the different techniques used for surface extraction. The developed process is able to handle non-manifold domains that result from multiple surface intersections. Surface and volume meshing that comply with user specified mesh control techniques are implemented to ensure the desired mesh quality. The integrated procedures provide a unique facility to handle geotechnical models and accelerate the generation of quality meshes for geophysics modelling. The developed procedure enables the creation of meshes for complex reservoir models to be reduced from weeks to a few hours. Various industrial examples are shown to demonstrate the practicable use of the developed approach to handle real life data.

This work studies the performance of a novel preconditioner, designed for thermal reservoir simulation cases and recently introduced in Roy et al. (SIAM J. Sci. Comput. 42, 2020) and Cremon et al. (J. Comput. Phys. 418C, 2020), on large-scale thermal CO(_2) injection cases. For Carbon Capture and Sequestration (CCS) projects, injecting CO(_2) under supercritical conditions is typically tens of degrees colder than the reservoir temperature. Thermal effects can have a significant impact on the simulation results, but they also add many challenges for the solvers. More specifically, the usual combination of an iterative linear solver (such as GMRES) and the Constrained Pressure Residual (CPR) physics-based block-preconditioner is known to perform rather poorly or fail to converge when thermal effects play a significant role. The Constrained Pressure-Temperature Residual (CPTR) preconditioner retains the (2times 2) block structure (elliptic/hyperbolic) of CPR but includes the temperature in the elliptic subsystem. Doing so allows the solver to appropriately handle the long-range, elliptic part of the parabolic energy equation. The elliptic subsystem is now formed by two equations, and is dealt with by the system-solver of BoomerAMG (from the HYPRE library). Then a global smoother, ILU(0), is applied to the full system to handle the local, hyperbolic temperature fronts. We implemented CPTR in the multi-physics solver GEOS and present results on various large-scale thermal CCS simulation cases, including both Cartesian and fully unstructured meshes, up to tens of millions of degrees of freedom. The CPTR preconditioner severely reduces the number of GMRES iterations and the runtime, with cases timing out in 24h with CPR now requiring a few hours with CPTR. We present strong scaling results using hundreds of CPU cores for multiple cases, and show close to linear scaling. CPTR is also virtually insensitive to the thermal Péclet number (which compares advection and diffusion effects) and is suitable to any thermal regime.

We study linear models for the prediction of the initial guess for the nonlinear Newton-Raphson solver. These models use one or more of the previous simulation steps for prediction, and their parameters are estimated by the ordinary least-squares method. A key feature of the approach is that the parameter estimation is performed using data obtained directly during the simulation and the models are updated in real time. Thus we avoid the expensive process of dataset generation and the need for pre-trained models. We validate the workflow on a standard benchmark Egg dataset of two-phase flow in porous media and compare it to standard approaches for the estimation of initial guess. We demonstrate that the proposed approach leads to reduction in the number of iterations in the Newton-Raphson algorithm and speeds up simulation time. In particular, for the Egg dataset, we obtained a 30% reduction in the number of nonlinear iterations and a 20% reduction in the simulation time.

The hydrocarbon extraction process is complex and involves numerous design variables and mitigating risk. Numerous time-consuming simulations are required to maximize objective functions such as NPV from a particular field while contemplating a significant representation of uncertainty scenarios and various production strategies. Production strategies searches may result in a high-dimensional search space which can yield sub-optimal reservoir economical exploration. As a solution, appropriate optimization algorithms selection and tuning may provide good solutions with lesser simulations. This paper presents a methodology to calibrate, develop, and select optimization algorithms for oil production strategy applications while quantifying the dimension and optimum location effects. Global optimum location altered the best method to be selected. It presents a novel algorithm (ASLHC) and a modification of the Nelder-Mead method (NMNS) to improve its high dimensionality performance. Performances of six pre-calibrated techniques were compared using novel normalized mathematical functions. Optimizations were limited to a 500 evaluation functions computational budget. The PSO, ASLHC, NMNS, and IDLHC were selected and implemented to perform production strategy improvements regarding two parameterizations of the reservoir management variables for a real reservoir model with restricted platform. Results showed the implemented algorithms successfully improved NPV by at least 8% at each of the 24 real-case optimizations. After upscaling the selected techniques for a 115 variable parameterization, the NMNS and IDLHC demonstrated good resilience against local convergence and each technique kept improving during all iterations of the process. An optimization method recommendation chart is presented based on the computational budget of the application.

Despite the robustness of standard genetic algorithms in receiver functions inversion for crustal and uppermost mantle velocity-depth structure, one drawback is that towards the end of a ‘run’, only a few variations in solution ideas are explored. This may lead to the stagnation of the optimization process and can be a major drawback for large model dimensions. To mitigate this problem, we introduced a new selection method that retains the best features of explored models, with an extinction procedure that increases the exploration of the model space through the principle of self-organized criticality. We test the performance of the modified genetic algorithm technique by applying it to the inversion of synthetically generated receiver functions for crustal velocity structure and comparing the results with those obtained using a standard genetic algorithm. The test cases involve using 2 different objective functions, based on the L2 norm and cosine similarity, with 2 different model parameterizations of different model sizes. The results show that our modified genetic algorithm improves the inversion process by consistently obtaining best models with the lowest misfit values and a distribution of best models with less deviations from the true model values. With an improvement of computation time of up to 11.2%, the results suggest that the modified genetic algorithm is best suited to obtain higher accuracy results in shorter computation times which will be especially useful for higher dimension models needing larger pool sizes.

The classification of land use and land cover (LULC) from remotely sensed imagery in semi-arid Mediterranean areas is a challenging task due to the fragmentation of the landscape and the diversity of spatial patterns. Recently, the use of deep learning (DL) for image analysis has increased compared to commonly used machine learning (ML) methods. This paper compares the performance of four algorithms, Random Forest (RF), Support Vector Machine (SVM), Multilayer Perceptron (MLP) and Convolutional Network (CNN), using multi-source data, applying an exhaustive optimisation process of the hyperparameters. The usual approach in the optimisation process of a LULC classification model is to keep the best model in terms of accuracy without analysing the rest of the results. In this study, we have analysed such results, discovering noteworthy patterns in a space defined by the mean and standard deviation of the validation accuracy estimated in a 10-fold cross validation (CV). The point distributions in such a space do not appear to be completely random, but show clusters of points that facilitate the discovery of hyperparameter values that tend to increase the mean accuracy and decrease its standard deviation. RF is not the most accurate model, but it is the less sensitive to changes in hyperparameters. Neural Networks, tend to increase commission and omission errors of the less represented classes because their optimisation lead the model to learn better the most frequent classes. On the other hand, RF and MLP prediction layers are the most accurate from a general qualitative point of view.

Accurately analyzing the flow and transport behavior in a large discrete fracture network is computationally expensive. Fortunately, recent research shows that most of the flow and transport occurs within a small backbone in the network, and identifying the backbone to replace the original network can greatly reduce computational consumption. However, the existing machine learning based methods mainly focus on the features of the fracture itself to evaluate the importance of the fracture, the local structural information of the fracture network is not fully utilized. More importantly, these machine learning methods can neither control the identified backbone’s size nor ensure the backbone’s connectivity. To solve these problems, a deep learning model named multi-aggregator graph neural network (MA-GNN) is proposed for identifying the backbone of discrete fracture networks. Briefly, MA-GNN uses multiple aggregators to aggregate neighbors’ structural features and thus generates an inductive embedding to evaluate the criticality score of each node in the entire fracture network. Then, a greedy algorithm, which can control the backbone’s size and connectivity, is proposed to identify the backbone based on the criticality score. Experimental results demonstrate that the backbone identified by MA-GNN can recover the transport characteristics of the original network, outperforming state-of-the-art baselines. In addition, MA-GNN can identify influential fractures with higher Kendall’s (tau ) correlation coefficient and Jaccard similarity coefficient. With the ability of size control, our proposed MA-GNN can provide an effective balance between accuracy and computational efficiency by choosing a suitable backbone size.