Bayesian Supervised Learning of Site-Specific Geotechnical Spatial Variability from Sparse Measurements

Non-parametric machine learning methods for interpolation of spatially varying non-stationary and non-Gaussian geotechnical properties

Subsurface profile prediction using AI tools has yet to be widely adopted in mainstream geotechnical engineering practice. A key gap is a lack of research efforts to validate AI-based predictions of subsurface conditions using field data. Geosetta, an innovative initiative in North American geotechnical practice, has developed a comprehensive geotechnical data repository with data contributed from various US State Departments of Transportation, Canadian Transportation Ministries, and several consulting firms across North America. These data are used to develop a machine learning (ML) model for subsurface predictions in data-sparse areas. Geosetta predictions are provided in the form of boring logs with ranges of Standard Penetration Test (SPT) N-values and soil type. While not a substitute for thorough geotechnical site investigations, ML predictions offer a valuable initial understanding of expected subsurface conditions and may help ground-truth geophysical surveys that are conducted prior to (or without) invasive geotechnical testing. Consequently, improving accuracy and assessing the efficacy of these ML predictions is of great importance. In this study, we evaluate the accuracy of Geosetta predictions by comparing them with high-quality geophysical data collected from Multichannel Analysis of Surface Waves (MASW) testing and traditional geotechnical boring data collected on the campus of Pennsylvania State University. These predictions are evaluated against traditional boring logs and MASW data through (1) direct comparison of N-values and strata delineation and (2) computation of Site Class based on VS30. Our results highlight the potential of Geosetta predictions as a practical bridge between desktop studies and site investigations, providing engineers with valuable insights into expected subsurface conditions and aiding preparation for site characterization.

In this study, we applied the geostatistical modeling to analyze an oil field. The reservoir properties, thickness, porosity and permeability, were studied. Data analysis tools, such as histogram, scatter plot, variogram and cross variogram modeling, were employed to capture the interpretable spatial structure and provide the desired input parameters for further estimation. SK (simple kriging), OK (ordinary kriging), Sgism (Sequential Gaussian Simulation), SC (simple cokriging), OC (ordinary cokriging) and MM2 (Markov model 2) methods were applied to estimate reservoir properties. Estimation difference maps were generated to compare the results of each method, providing more straightforward realizations in a visual way. For thickness, results indicated that anisotropic variogram could provide better interpretations for the spatial relationships than isotropic variogram. Both SK and OK could provide better estimates. In comparison to the conventional estimation techniques, the simulation method could well reflect the reservoir's intrinsical characteristics in terms of the associated extreme values. OOIP (Original Oil In Place) was calculated later with the parameters attained before, including thickness and porosity. Estimation difference maps showed that there was no obvious difference in SK vs. OK and SC vs. OC for the study of permeability. However, OC was slightly different from OK, and there were significant discrepancies between the estimates of OC and MM2 at the unsampled locations. In addition, OC estimates were closest to the sample data of permeability with the minimum variance. Geostatistical modeling is an effective way for thickness, porosity, and permeability prediction.

Enhancing spatial estimates of metal pollutants in raw wastewater irrigated fields using a topsoil organic carbon map predicted from aerial photography

“Concept drift” and class imbalance are two challenges for supervised classifiers. “Concept drift” (or non-stationarity) is changes in the underlying function being learnt, and class imbalance is a vast difference between the numbers of instances in different classes of data. Class imbalance is an obstacle for the efficiency of most classifiers. Previous methods for classifying non-stationary and imbalanced data streams mainly focus on batch solutions, in which the classification model is trained using a chunk of data. Here, we propose an online Neural Network (NN) model. The NN model, is composed of two different parts for handling concept drift and class imbalance. Concept drift is handled with a forgetting function and class imbalance is handled with a specific error function which assigns different importance to error in separate classes. The proposed method is evaluated on 3 synthetic and 8 real world datasets. The results show statistically significant improvement to previous online NN methods.

A novel edge detection of image based on window empirical mode decomposition (WEMD) was proposed in this paper. WEMD was inherited all advantages from EMD and can decompose any nonlinear and non-stationary data into a number of intrinsic mode functions (IMFs). In WEMD, window mean function is employed to get the mean surface in the decomposition process instead of the surface interpolation, which enables fast decomposition. Meanwhile, the drawback of gray spots in IMF images has been avoided. Since the first IMF provides the highest local spatial variations and/or scales of the image, this IMF is then processed for obtaining the edge. 2D-Hilbert transform and non-maxima suppression, hysteresis threshold in canny method are applied to the first BIMF to achieve the desired edge map. The proposed method is compared with Canny and wavelet edge detection techniques. Simulation results with the real images demonstrate the efficacy of the proposed algorithm for edge detection.

Derivation of site-specific soil-water characteristic curve (SWCC) from extremely sparse experimental data by hierarchical Bayesian method with consideration of geotechnical sites similarity

Modeling heterogeneous reservoirs is cumbersome as it requires a great effort to determine the variation of properties with respect to direction, while the lack of adequate data makes this a difficult task. Generating static models for heterogeneous reservoirs remains an important challenge in petroleum engineering applications which requires more investigations. Some heterogeneous reservoirs, such as thin bed reservoirs, may be divided into some homogeneous subzones where characterization of these homogeneous sub-reservoirs and their integration can represent the properties of the heterogeneous reservoir. To investigate this concept, in this paper, three exemplar reservoirs (ER) were generated. The heterogeneity in the data is increased from ER1 to ER2 and ER3. In the first step, each reservoir was studied as one single zone, so the results can be compared with the proposed method in this work. Ordinary kriging (OK) and multilayer perceptron neural network (MLP) were used for modeling of these exemplar reservoirs. This study showed that OK cannot model reservoir characteristics, whereas the MLP yielded reasonably acceptable results. In the next step, a hybrid clustering classification-based method was applied to divide the reservoir to homogeneous subzones. Each reservoir was modeled in terms of its homogeneous subzones. The homogeneous subzones were modeled using OK and MLP. The results showed that the developed model was successful in modeling the heterogeneity at a reasonable CPU processing time. Also, it was seen that in case of using the simple modeling techniques, MLP neural network yields more reasonable results, compared with OK.

Comparison between geostatistical and machine learning models as predictors of topsoil organic carbon with a focus on local uncertainty estimation

Large topsoil organic carbon variability is controlled by Andisol properties and effectively assessed by VNIR spectroscopy in a coffee agroforestry system of Costa Rica

Various approaches have been used to estimate soil salinity (EC e ) at unsampled locations. Some of these approaches are briefly discussed. Of these, geostatistical methods such as ordinary kriging (OK1 and OK2), regression kriging (RK), three‐dimensional kriging (3‐DK), and cokriging (Co‐K), provide best linear unbiased estimates (BLUE). These methods were tested with a raw electromagnetic induction instrument (Type EM38) in a soil electrical conductivity (EC a ) survey, and calibrated EC e data were obtained from an irrigated cotton ( Gossypium hirsutum L.)‐growing area in the Edgeroi district of the lower Namoi valley, northern New South Wales, Australia. We compared these methods, on the basis of precision and bias of estimation, and found that RK was the best performer. This is because of the incorporation of regression residuals within the kriging system. Mean and standard deviation of ranks (SDR) showed that Co‐K performed best against these criteria.

A spatial statistics approach to the quantification of finite strain variation in penetratively deformed thrust sheets: an example from the Sheeprock Thrust Sheet, Sevier Fold-and-Thrust belt, Utah

Issue Topics - updated

A new approach to the Ordinary Kriging interpolation method based on the combination of local interpolation and variogram modelling with Adaptive Neuro-Fuzzy Inference System (ANFIS) is proposed for scattered data interpolation. In this method, the experimental variogram is modelled by ANFIS and this model is used to interpolate the unknown values of specific points in a new local manner. In this local way, all the unknown points are grouped based on each reference point. The study data obtained from mathematical functions are used. The tests show that the proposed method provides better performances for all data sets in comparison to the well known and highly approved interpolation methods; Ordinary Kriging, Triangle Based Cubic and Radial Basis Function-Multiquadric. Moreover, by the proposed method the computational complexity impressively decreases compared to the global ordinary Kriging.

Deterministic construction of sparse binary matrices via incremental integer optimization