Geo-Risk 2023: Innovation in Data and Analysis Methods

The human exploration of the Moon will require the knowledge of physical and mechanical properties of the regolith on the Moon’s surface through terrestrial geotechnical testing methods such as cone penetration tests. This paper presents a study to investigate the effects of the porosity of lunar regolith simulants of JSC-1A on the interaction between regolith and cone penetrometer using the three-dimensional discrete element modeling (DEM). The particles of lunar regolith simulants are modeled using clumps in DEM with contact parameters calibrated to replicate the realistic stress-strain behavior of regolith simulants using triaxial tests. An assessment of the tip resistance and sleeve friction in relation to depth increase during the penetration into the lunar regolith is evaluated. The discrete element models with different regolith porosity settings are employed to develop a correlation between the porosity and the corresponding tip resistance and sleeve friction during the cone penetration process. The results from this study can provide useful references to assess the in situ porosity of lunar regolith on the Moon’s surface based on the tip resistance and sleeve friction during the cone penetration process.

The uniaxial compressive strength (UCS) of rocks is used widely in tunneling, rock, and mining engineering. Direct methods for testing UCS of rocks often require well-prepared samples of high quality and therefore are often relatively time-consuming and expensive. In this case, indirect approaches such as empirical equations and machine learning methods are proposed for UCS prediction. Due to the data-driven and non-parametric characteristics of machine learning methods, UCS data of rocks are often estimated through these methods. In this study, one of the machine learning methods, entitled a fully Bayesian Gaussian process regression (fB-GPR), is applied to predict UCS as well as uncertainty quantification associated with the prediction. In the meanwhile, the optimal model for predicting UCS is also determined and examined through a systematical manner. Results show that the fB-GPR method can be well used for accurately predicting UCS with reasonably quantified uncertainty through developing the optimal model.

This study investigated the performance of two data-driven site characterization (DDSC) methods, Glasso and Glasso-BFs, through a benchmark example based on a real case history [labeled as RG1(Adelaide)]. This study focused on root-mean square errors (RMSEs) of q_t depth profiles and runtime required for both of model training and validation, and these performance metrics of Glasso and Glasso-BFs were compared. Although two methods can predict q_t profiles well, it was difficult to conclude that which method has “higher” performance in the benchmarking. Based on the comparisons, we found that (1) if abrupt changes such as layer boundaries need to be captured, Glasso is recommended; and (2) if 3D subsurface models need to be estimated with reasonable time, it is recommended to use Glasso-BFs.

This paper examines assimilation of meteorological and hydrological monitoring data into landslide prediction models and landslide early warning systems. The monitoring data are collected with an Internet of Things−based monitoring system. Basic information on the instrumented case study area and the implemented sensor setup are provided. The sensor setup consisted of volumetric water content and suction sensors with a weather station for meteorological monitoring. The sensor data, collected from August 2020 to October 2021, are presented and interpreted highlight main trends in hydrological conditions characteristic for seasonally cold climate in Norway. Several types of water-induced landslide early warning systems are examined with the sensor-based, empirical, and physical-based landslide early warning systems. Details of the implementation of the physical-based early warning system based on a physical-based landslide prediction model are presented.

The landslide is a major natural hazard that can cause catastrophic economic and life losses. Previous studies mainly focus on the evaluation of spatial variability of soil parameters on slope stability, and the effects of stratigraphic variations on slope failure mechanisms have often been overlooked. In this study, a Bayesian supervised learning algorithm, called multiple point statistics (MPS), is adopted to determine the spatial distribution of stratigraphic boundaries from sparse site-specific data and a single training image. Multiple generated realizations of slope stratigraphy serve as input for slope stability analysis. Subsequently, a unified indicator is defined for the landslide risk assessment considering both the exceedance probability and sliding mass. A synthetic slope example is simulated to illustrate the proposed framework. Results indicate that stratigraphic uncertainty has a significant effect on the slope risk, and the calculated risk may be significantly underestimated when only sparse site investigation data are available.

The tunnel projects are incredibly significant for the safety and performance of the aboveground structure. The tunnel deformational performance is significantly affected by both geological uncertainty and spatial variability of soil properties. Many efforts have been devoted to analyzing the effect of geological uncertainty and spatial variability, respectively. Limited studies dealt with both geological uncertainty and spatial variability in a specific geological model. In this paper, the influence of these two uncertainties will be considered simultaneously. Geological uncertainty is modeled by improved coupled Markov chain theory based on spare limited boreholes. Spatial variability is characterized using random field theory. This method can effectively evaluate the tunnel structural response considering these two types of uncertainty. The results indicate that calculations only considering geological uncertainty may be overestimated if the spatial variability is ignored. Moreover, subset simulation is used to efficiently calculate the reliability index of tunnel deformational performance.

Characterization of the spatial variability of soil properties is a prerequisite for risk and reliability analysis of geotechnical structures. The spatial variability of soil properties is usually characterized by random field parameters, including mean value, standard deviation, and autocorrelation distance. Due to the limited and sparse data usually available in geotechnical practice, the estimated random field parameters usually suffer from large uncertainty. To address this issue, this paper proposes a Bayesian estimation method that uses a novel prior probability distribution derived from soil data and the likelihood ratio test statistics. The unique feature of the proposed probability distribution is that it has an analytical form, making it able to be readily used in Bayesian updating. The effectiveness of the proposed method is demonstrated using simulated soil data. The results show that the Bayesian estimation with this prior information significantly reduces the uncertainty of the random field parameters, compared with the estimation of the maximum likelihood method and that of a Bayesian method using a uniform prior distribution.

Identifying soil constitutive model parameters is essential for geotechnical engineering analysis. Soil constitutive model parameters are usually calibrated from laboratory or field test data, which can be formulated as an inverse analysis problem. Nevertheless, the test data or test costs in engineering are limited, so that available data may be insufficient to identify the unknown model parameters. The identifiability, that is, whether the data provide enough information for parametric identification, is rarely discussed for inverse analysis in geotechnical engineering. This paper aims to explore the identifiability of the modified Cam clay (MCC) model parameters based on the triaxial test data by the Bayesian method. In this study, the sequential attribute of triaxial data is considered by modelling the data using the probabilistic state space model. Results show that MCC model parameters might be unidentifiable using consolidated undrained triaxial data, while it is globally identifiable for consolidated drained test data. In addition, the maximum a posterior (MAP) estimate and mean values obtained from the joint posterior distribution are close to the true values when MCC model parameters are globally identifiable. However, this is not the case for unidentifiable cases.

This study presents a method for data-driven site characterization (DDSC) based on a Markov random field (MRF) model called Potts model. A Potts model is a graphical model of interacting spins on a lattice and has been successfully applied in image processing such as image de-noising, restoration, and inpainting. Existing Potts model-based methods have been generally designed for image processing, and they do not fit the setting of subsurface modeling in which the available observation data is very limited. In the proposed method, an anisotropic Potts model under fixed external magnetic fields is used as an MRF model, and the hyperparameters are trained using leave-one borehole-out cross-validation. The proposed method is simple and requires only a few borehole data on soil types to train the Potts model. The proposed method was demonstrated through numerical tests, and the effects of the number of training set on estimations were investigated. The proposed method can quantify the uncertainty/accuracy of the estimation, and spatial distributions of estimation uncertainty are also shown in this paper.

Borehole drilling has been regarded as a traditional way of obtaining geological information in site investigation, while the borehole locations are difficult to be determined. Since increasing the number of boreholes requires additional expenditure, the optimization of borehole locations considering stratigraphic uncertainty has been crucial for sustainable development. In this study, a borehole planning strategy is proposed considering updated 3D stratigraphic uncertainty distribution, which is determined by a stratum-informed random field-based method, site division approach, and information entropy. Several sub-sites are generated based on the neighboring boreholes, enabling the site-specific uncertainty to be evaluated and compared. The optimized borehole location is the center of the sub-site presenting the highest average entropy. The effectiveness of the proposed strategy is verified through a three-layer geological model and compared with the traditional borehole planning strategy. The results have shown that the proposed strategy effectively decreases the information entropy and increases the accuracy.

Since soil properties are generally measured at limited locations, it is preferable to use all available measured data including indirect data effectively to estimate the spatial distribution in a rational manner. A more accurate estimation can be achieved by considering the indirect measurement data that are highly correlated with the soil property in interest. In this study, we use GPR with multiple random fields. The random and trend components are estimated using the Whittle-Matérn autocorrelation function. The accuracy of estimation is compared between the cases with and without consideration of indirect data for the random component. Even if the direct data is not observed around the target point, a detailed spatial distribution is obtained by considering the surrounding indirect data when they have strong correlation.

Cone penetration test (CPT) is one of the most widely used in situ methods for characterizing the spatial variability of site characterization, due to its rapidity, affordability, and repeatability. It is often encountered in practice that some CPTs probe deeper than others, and that some CPT soundings may contain missing data due to presence of gravel-sized particles or intentional bypassing of gravelly soil layers. Furthermore, the number of CPT soundings in a cross-section is often sparse, although the number of data points along the depth is adequate. In these cases, it is difficult and challenging to estimate spatially varying soil property at locations without CPT soundings, especially for non-stationary CPT within multi-layers. While certain methods have been proposed hoping to address these concerns, they are frequently constrained by either stationary assumption of data, autocorrelation function forms, or computational issues. This study presents a Bayesian compressive sensing-based machine learning methods for this issue, with both numerical and real-world CPT data for validation and demonstration. Results show that the proposed method performs reasonably well in simulating 2D non-stationary CPT profiles from incomplete data set.

Data points in geotechnical site investigation data (i.e., CPT data) may be missing sometimes due to various reasons. This study proposed to use a semi-parametric Gaussian process regression (GPR) method for predicting missing data in geotechnical testing results. Semi-parametric GPR divides the spatial data into the trend function, spatial residual, and measurement errors. Compared with conventional GPR method, semi-parametric GPR enhances model interpretability and accuracy. However, this involves challenges in estimating the parameters in the model. Conventional GPR applications infer the model parameters based on maximum a posteriori (MAP) estimation. However, this method can only provide a point estimation of the model parameters. Point estimation may be trapped by a local optimum result. This study utilizes the Hamiltonian Monte Carlo (HMC) method to get the full posterior distribution of the model parameters. MAP and HMC methods are both applied to infer the model parameters based on a synthetic CPT data set. The performances of both methods are compared with the true model values. The results show that the model parameters estimated from the HMC are more reliable.

In situ dynamic cone penetration tests (CPTu) are performed to efficiently characterize near-surface marine sediments. Dynamic data require a strain-rate correction according to strain-rate model and law due to the nonlinear penetration rate of the cone probe insertion. This requires paired (i.e., closely located in space and time) quasi-static and dynamic CPTu measurements, increasing cost and reducing time efficiency of the dynamic tests. Two Bayesian methodologies for dynamic cone tip resistance correction are exploited, which make use of limited unpaired quasi-static CPTu observations acquired proximal to dynamic test’s location. Unpaired dynamic cone tip resistance profiles are fitted to discrete cosine transform functions within two different Markov chain Monte Carlo simulation processes: a transient (TMCMC) and not transient one (MCMC). Results show the effectiveness of the transient Markov chain scheme over the not transient one to predict strain-rate coefficients in agreement with those obtained from the conversion of dynamic and paired quasi-static CPTu.

The within-site variability of seismic site response describes the randomness in site response from earthquake to earthquake at a single site. Although the within-site variability of seismic site response can be quantified for seismic stations with certain number of earthquake recordings, it remains challenging to estimate the within-site variability of site response across a large region. This paper aims to provide a geospatial modeling framework for predicting the within-site variability using inexpensive statistical learning methods and geospatial proxies. We obtain the within-site variability data using a large number of earthquake recordings from 114 KiK-net vertical array sites. We examine the predictive power of 18 geospatial proxies associated with site stiffness, topography, basin, and saturation conditions in estimating the within-site variability of seismic site response. The results suggest that geospatial modeling shows promise for evaluating site response across broad regions.

Soil boundary delineation is an important task in geotechnical site characterization. It can be achieved by extracting borehole samples, conducting laboratory tests, and classifying them according to soil classification system such as the Unified Soil Classification System (USCS). However, borehole samples are sparse in space, so the delineation of soil boundaries is challenging. In practice, soil boundary delineation is sometimes achieved by conducting multiple cone penetration test (CPT) soundings. Because CPT provides nearly continuous records in the vertical direction, the soil boundaries at the soundings may be identified from the I_c (soil behavior−type index) profiles. However, given the I_c value of a soil, its USCS classification is not 100% certain. It is unclear whether the boundaries identified by I_c reflect the actual boundaries defined by USCS classification. The purpose of this paper is to propose a novel framework of inferring the spatial variation of USCS classification (e.g., sand, silt, and clay) based on both CPT and borehole data. With past experiences on the I_c-USCS correlation from other sites, it is possible to combine site-specific CPT and borehole information to infer the site-specific spatial variation of USCS classification. Once the spatial variation of USCS classification is inferred, the soil boundaries can be identified. The uncertainties for the spatial variation of USCS classification as well as soil boundary identification are quantified through Bayesian analysis.

Geotechnical risk is an important part of managing large projects. This study developed a data-driven approach to tracking risk identification on major US transportation projects using textual analysis of historical project documents. In early phases of a project, the identification and assessment of risk is based largely on expert judgment. As a project moves through its life cycle, identified risks and their assessments evolve. Some risks are realized to become issues, some are mitigated, and some are retired as no longer important. The study investigated the comprehensiveness of early risk registers on 11 large transportation projects in comparison to how those risks changed as the projects progressed. Finite state automation methods similar to Markov chain models were used to track changes in risk attributes as the case-study projects matured. The objective was to be better able to anticipate how project risks change as projects mature and to be better able to forecast changes to risk registers through the project life cycle. Results suggest that fewer than 60% of initially identified risks, both geotechnical and other, ultimately occur in projects, while more than 40% are retired. Categorizing risk management styles illustrates that planning for geotechnical risks in the initial phase of the project is necessary but not sufficient for successful project delivery.

Rapid damage assessment following natural hazard events is critical to first responders, cognizant government agencies, and private organizations. This paper describes the use of crowdsourced social media data to complement more traditional sources of information in monitoring natural hazards and their effects. In a series of studies, the use of social media data, principally Twitter postings, has been investigated as a complementary source of information to more traditional sensor-based systems. The intent is to explore the value of social media data to provide rapid indications of damage following sudden-onset events. The study focused on testing whether social media users react to natural hazard events such that the resulting data can be used to make damage assessments. Text-based assessment and classification models were applied to identify damage levels from contemporary tweets. The study defined a damage assessment scale for earthquake damages adjusted from Modified Mercalli Intensity and developed a text classification model for assessing damages. The results were validated against USGS appraisals of the same event.

Landslide susceptibility mapping (LSM) predicts the possibility of future landslides and is critical for risk assessment, resource allocation, and land-use planning. To promote the prediction accuracy, generalization ability, practicability, and interpretability of LSM models, this paper makes the following novel contributions: (1) Two hybrid models based on the infinite slope stability model (ISSM) and different data-driven approaches for regional LSM are proposed. The adopted data-driven approaches are the logistic regression (LR) model and convolutional neural network (CNN); and (2) The LR model is compared with the LR-ISSM hybrid model to verify the important role of the physical module. The results reveal the necessity of considering the spatial correlation among grids/pixels for grid-based LSM models. While the CNN-ISSM hybrid model is slightly more accurate, the LR-ISSM hybrid model has better interpretability and produces promising prediction accuracy and generalization ability.

The main objective of this research is to investigate the use of horizontal-to-vertical spectral ratio (HVSR) measured from ambient vibrations for predicting site response using artificial neural networks (ANNs). The entire spectrum (amplitudes of all frequencies) of HVSR is used as a predictor variable in the model. Artificial neural network models are developed and then optimized through hyperparameter tuning. The developed models have been assessed based on the standard deviation of the residual terms and how these residuals compare with the corresponding values in ground motion models based on traditional regression approaches. It has been observed that the ANN models improve the prediction of site response as evidenced by the reduction in the uncertainty (about 10% at short periods) in the site-to-site variability.

This paper presents a deep learning-based spatial estimation of unconfined compression strength. First, we briefly describe the deep-learning method used in this study. Next, we explain the boring investigations and the topography of downtown Osaka City. A parametric study, with hidden layers and neurons as variable parameters, performed to obtain the optimal structure in the network is detailed. Then, the relationship between the unconfined compression strength and topography is discussed through the distribution of unconfined compression strength in both east–west and north–south directions. Finally, the availability of spatial estimation with deep learning was represented.

The pipe-roofing method is a new pre-supporting large-section excavation technique. Rapid and effective prediction of ground settlement for excavation adjustment and decision-making is significant to ensure safety during pipe-roofing excavation construction, especially in saturated soft soil areas. This paper develops an intelligent model mixing the physics mechanism and field data for forecasting the ground settlement caused by pipe-roofing excavation. To begin with, a data set containing features regarding physical mechanisms and field monitoring items is well prepared. For one thing, there are three kinds of physics-based features to be taken into account, including excavation and supporting condition, soil condition, and condition of steel pipe-roofing and supporting system. Based on the relationship between construction time sequence and intrinsic physical mechanism, a finite-element analysis model is introduced, which can transfer spatial structure features into data characters. For another, field monitoring data, including the distance of the excavation surface and ground settlement, also need pretreatment. After that, a proper machine learning algorithm is trained to fully learn the prepared data set where it contains the time sequence. To verify the effectiveness and practicability, the proposed model mixing physics mechanism and field data is retrospectively applied to a recent field monitoring case of Guiqiao Road Station on Shanghai Metro Line No. 14.

The purpose of this study is to apply ensemble methods to predict surface settlement induced by earth pressure balance tunnel boring machine. Random forest (RF) and Extreme Gradient Boosting (XGBoost) algorithms are applied on 1,101 settlement measurements collected from the Grand Paris Express project. The results are compared with the performance of the back-propagation artificial neural networks (BPNN). Finally, the results show that both ensemble methods XGBoost and RF are better than BPNN based on R² and RMSE indicators.

Recent technological developments in remote sensing and image analysis and their introduction into the mining industry have opened the door for updating risk analysis techniques. The R&D team of the Mining Automation Lab of the University of Nevada–Reno develops UAV-based photogrammetry and image analysis technologies to address the geotechnical risks of the tailings dam. All procedures are based on high-resolution imaging by a special UAV control software for terrain-following and photogrammetry for 3D modeling. The system can efficiently cover large areas and capture high-resolution images as input for novel image analysis methods. This approach’s biggest challenge is acquiring enough annotated samples to train a model that produces high-accuracy results in various rock types and environmental conditions. The results can be used for localized risk analysis of rockfall using available empirical models. Our team at UNR is working on including multispectral imaging and unifying all the data formats and analysis techniques under a digital twin framework.

Hydraulic conductivity is one of the most important mechanical properties of soil in geotechnical engineering—especially on seepage, consolidation, slope stability, and so on. In general, estimation of the hydraulic conductivity has been made using the empirical correlations with the relevant engineering characteristics such as porosity, soil textural class, grain-size distribution, organic content, specific surface, etc. This study proposes applications of self-organizing map (SOM) technique on the stochastic estimation of the hydraulic conductivity. The database consisting of fully saturated soils, primarily from the Florida area, is utilized for model training and verification. The result shows that data clustering using SOM can promote the more accurate statistical estimation than the conventional empirical correlation formulas.

Sidehill and through cuts are utilized in the construction of transportation corridors to optimize routes through mountainous and hilly terrain. Although the use of cuts increases the capabilities of transportation systems, their use comes with considerable risk. Primarily, rights-of-way sections that utilize cuts are susceptible to damage from mass wasting events such as shallow-rapid landslides and rockfalls. This study aimed to create a new slope stability risk framework for railroad cut sections by processing digital images of railroad rights-of-way recorded by inspection vehicles and related geophysical data. A geohazard-affected track section along the Harrisburg Line was used as the study area. Computer vision techniques were used to identify and quantify geohazard features that indicated slope instability. An object detection model based on deep learning was trained to detect these slope instability indicators and generate risk scores from rights-of-way inspection videos. Moreover, a landslide inventory was compiled, and a landslide susceptibility model was developed for the study area based on available geophysical data. The object detection model and the landslide susceptibility model were combined using a relative risk assessment framework to determine which sections were most at-risk of landslide, and results were compared with the railroad-identified geohazard sections across the study area.

Landslides comprise about 42% of the entire natural hazards in Indonesia, claiming one-third of the annual economic losses caused by natural disasters. To mitigate their severe impacts, a landslide susceptibility map (LSM) with actual, continuous, and accurate information about landslide occurrences and their likelihood in a particular area is urgently needed. Therefore, this research presents a reliable landslide susceptibility mapping with combined interferometry synthetic aperture radar (In-SAR) and machine learning (ML) techniques. A new framework of LSM by using In-SAR and ML applied to road infrastructure was proposed. The framework begins with the acquisition of a satellite imaging-based digital elevation model (DEM) of a particular road corridor, in which the landslide contributing factors (slopes, natural drainage networks, lithology, and rainfall) were rendered and overlaid by using GIS. Then, the ML was used to rate those contributing factors to the actual landslide occurrences. In a parallel way, the In-SAR was employed to obtain the ground movements in the road corridor from a series of SAR images derived from the Sentinel-1. Interferograms were then generated to produce ground movement maps. By combining the ground movement maps from the In-SAR and landslide ratings maps from the ML, a landslide susceptibility map was created. The applicability of the framework was investigated through a case study of landslide susceptibility in West Sulawesi, Indonesia.

Building vulnerability under landslide impacts is a critical component in the quantitative risk assessment (QRA) of landslides. This paper proposes a probabilistic building vulnerability evaluation framework under landslide impacts. In the proposed framework, a unified numerical model combining a dynamic run-out model (used for simulating run-out behaviors of landslides) and a finite element method (used for simulating the building capacities) is constructed to simulate the damage degree of buildings under landslide impacts. Considering the uncertainties of both the landslide run-out and the building structures, a probability density evolution method (PDEM) is used to estimate the building vulnerability under landslide impacts. The proposed methodology is illustrated in a scenario where a loess-flow landslide impacts a five-story RC frame building. The results show that the computational efficiency of PDEM is significantly improved compared with the Monte Carlo simulation. Moreover, the vulnerability results can be more closely related to the QRA of individual landslides.

The slope system unavoidably contains many uncertainties. These uncertainties can be effectively taken into account by probabilistic analysis methods. In the present paper, a quantile-based first-order second-moment method (QFOSM) is applied to efficient slope reliability analysis. The QFOSM uses the quantile values of uncertain parameters and partial derivatives at the quantile to estimate the slope reliability index. It shares similar simplicity and efficiency with first-order second-moment method (FOSM) and simultaneously has similar accuracy and robustness with first-order reliability method (FORM). The effectiveness of the applied method is demonstrated by two homogenous soil slopes. The adopted method is compared with the FOSM, FORM, and MCS. The results show that the method can accurately estimate the slope reliability with several times of evaluations of slope finite element model and does not need to be performed in the standard normal space. It may serve as a simple and efficient analysis tool for a geotechnical engineer to estimate the slope reliability quickly when subjected to tight schedule.

Previous studies show that the failure probabilities estimated using existing reliability methods are often overestimated more than an order of magnitude than the observed frequency, and this has been deemed as a challenging problem in geotechnical risk and reliability community. The actual failure probability is often quite small. For example, the observed annual rates of slope failure in Hong Kong were 0.041% and 0.329% for engineered and non-engineered slopes in 2020. To address this challenge, this study proposes to explicitly model landslide triggering mechanism (e.g., rainfall infiltration) and utilize the fact that an existing slope stayed stable during past rainfall events to update uncertain soil parameters for predicting the time-variant slope failure probability under a future target rainstorm. Results showed slope survival records could be utilized to reduce uncertainties in soil parameters and enhance prediction of slope failure probability in both its absolute magnitude and variation over time.

Many methods based on Monte Carlo simulation (MCS) have been studied and developed for calculating small limit state probabilities efficiently. Recently reliability estimation methods with surrogate model have attracted attention due to their low calculation cost. Since Echard et al. proposed active learning method combining kriging and MCS (AK-MCS) in 2011, many papers have been published on the improvement of this method. AK-MCS uses surrogate models by kriging to evaluate not only the estimated values but also their uncertainties. In this study, we use the Gaussian process regression with multiple random as a surrogate model and importance sampling without design points to calculate the limit state probability as the reliability estimation method. Finally, as an example of the application, we show the results of estimating the limit state probability using the proposed method for an eight-dimensional consolidation settlement problem.

Insufficient knowledge of the ground condition may lead to design failure or issues during/after construction, which is particularly critical for dam safety. Nevertheless, performing a comprehensive site investigation may not be practical due to limitations about time, budget, or accessibility. To better understand and further reduce the uncertainties and risks, the rational approach is combining multiple methods, such as site investigation, quantitative analysis, modeling, and instrumentation and monitoring, along with applying observational method throughout project. The paper introduces a project of constructing a new bridge adjacent to a more than 100-years reservoir dam, while keeping dam operational during construction. In the project, a combination of multiple means was applied to efficiently reduce the uncertainty and risk when very limited information was available. The paper presents the measures that have been taken to increase understanding about the site conditions. The uncertainty and probability of failure before and after acquiring more information were quantitively evaluated. The results show that the uncertainty in seepage and slope stability have been greatly reduced with the coefficient of variation values have been reduced from 0.62 to 0.41 and from 0.23 to 0.08, respectively, so as the probabilities of failure have also been reduced.

This study aims to establish a city-scale 3D geological model for Hong Kong offshore to advance geotechnical practitioners’ understanding of the subsurface soil conditions and facilitate more rational geotechnical planning and design. A large number of marine borehole logs and two marine geophysical survey maps including the seabed level and Quaternary base level are integrated in the 3D modeling. The Markov random field method is applied to assess the 3D distribution of major geological units from existing borehole logs at a horizontal interval of 50 m and a vertical interval of 0.5 m over 70 km × 46 km × 100 m area. Based on the 3D model, the probability distribution of geologic units can be obtained, and the most probable formation at each spatial location is generated. The dominant soil and rock surfaces are identified for use in geotechnical engineering designs.

This work aimed to propose a new framework for regional slope stability based on probabilistic analysis by combining a hydromechanical model, that is, transient rainfall infiltration and grid-based regional slope-stability model (TRIGRS) and the reliability method. A user-friendly software based on geographic information system (GIS) platform called the improved GIS-TRIGRS-FORM landslide prediction toolbox applying the Python programming language for considering the possible uncertainties of model parameters as well as rainfall conditions was developed. To incorporate the probability information of the hydromechanical model incorporating the infinite slope, the first-order reliability method (FORM) is firstly embraced in the analysis. The proposed approach can effectively provide regional hazard distribution maps subject to several indexes in different time phases, for example, the factor of safety (FoS), reliability index (RI), and failure probability (P_f), which can achieve a regional-scaled landslides susceptibility analysis using physically based modelling.

Due to the continuous development of urban areas, excavation has become a crucial part of the building’s construction. From the geotechnical point of view, excavation safety must be considered. Nail walls are an economical solution, widely used to support excavations. In this study, the construction of a nailed wall is simulated using a limit equilibrium method (LEM) to study the excavation stages effect on the factor of safety (FS). The variability of the soil properties and nails’ bonding strength are considered to capture the nailed wall reliability index (RI). The Janbu corrected method coupled with the Latin Hyper Cube is used to estimate the FS and RI. The impacts of the correlation length between the soil parameters are investigated on the RI.

This study explored the effect of learning function on reliability analysis of geotechnical engineering system using adaptive Bayesian compressive sensing (ABCS) and Monte Carlo simulation (MCS) (ABCS-MCS). The ABCS-MCS method can provide both response prediction at an unsampled point and quantify explicitly the associated prediction uncertainty. It relies on a learning function to adaptively determine the minimum number of sampling points and corresponding sampling locations for achieving a target accuracy of reliability analysis. Therefore, learning function plays an important role in ABCS-MCS, and its learning criteria and stopping condition directly affect the accuracy and efficiency of reliability analysis. Four different learning functions are investigated together with ABCS-MCS. A comparative study using these four learning functions in ABCS-MCS is illustrated using a two-layered cohesive slope reliability analysis problem. Results show that ABCS-MCS combined with U-learning function has the highest accuracy, efficiency, and robustness for reliability analysis.

Aiming at the problems that the life evaluation results of most lining segments are too conservative and the tunnel life prediction is inaccurate, the existing monitoring data is used to complete the dynamic update of the service life of the lining segments. The results show that in the environment of chloride ion erosion, the theoretical calculation and simulation calculation show that the lining segment will reach the end of life due to insufficient reliability in the 92nd year; using the existing observation data, the theoretical data is dynamically updated according to the Bayesian formula, 88. In the early years, the reliability of the lining segments failed. Tunnel lifespan reduced by 4%. In practical applications, it is necessary to strengthen the protection in advance to ensure that the tunnel can complete the set construction period.

Risk assessment of structures subjected to flow-like natural hazards, such as mud/debris flows, lahars, or avalanches, are commonly based on propagation models. Numerical models require significant computational effort to predict output quantities, such as run-off lengths, velocities, or impact forces. In this study, Bayesian inversion is applied to a well-documented case in order to calibrate the governing parameters of the propagation model on kinematic measurements. As this technique is heavy in terms of model evaluations, it is performed with the help of a surrogate model. To ensure the accuracy of this surrogate model, various meta-modeling techniques are compared in the case of a small experimental design. Finally, a propagation model is built on the best point estimate resulting from the calibration. The results of this model are compared to the actual measurements of the studied case in order to demonstrate the applicability of the presented approach.

This study presents a simple consolidation analysis method to predict surface settlement using Bayesian back analysis for embankments built over soft soil. Simplification of the geotechnical model and parameters is common in engineering practice; however, review of the performance to verify and update the prediction is seldom. Oversimplification however may not capture the appropriate conditions for reliable settlement prediction. Bayesian back analysis provides a way to update the adopted prior parameters using monitored data. Parameters such as the compression ratio, recompression ratio, creep strain rate, and the coefficient of compressibility were treated stochastically as random variables. Both a nine-layered and three-layered model were analysed for comparison. The analysis used simplified models to show the surface settlement was well predicted utilising about 117 to 215 days of observed data. The settlement data was used to update the selected parameters through Bayesian back analysis to fit the entire time-settlement history.

The asphalt core wall dam has good adaptability to deformation and anti-seepage ability, but it still faces the risk of collapse when floods exceed 10 times or even 100 times of the design, posing a serious threat to the lives of downstream residents. Therefore, it is very important to quantitatively analyze the loss of life caused by flood overtopping. Based on the actual case of Sheyue Gully in Xinjiang, according to the logical structure of the loss of life mechanism, this paper considers the uncertainty and interrelationship between the five main influencing factors of alarm time, flood severity, risk population, evacuation, and understanding of flood, and 12 secondary influencing factors. A model based on Bayesian network quantitative analysis of the loss of life caused by outburst is proposed. According to the statistical data, the probability of network nodes (factors) and arcs (interrelationships) is quantified, and the loss of life caused by outburst is quantitatively estimated. By comparing the recorded life loss with the analysis results of other model methods, the correctness and applicability of the Bayesian life loss model are verified. Finally, through the sensitivity analysis of the influencing factors of life loss, the most important influencing factors leading to life loss are revealed. The model is of great significance for the quantitative analysis of the risk of life loss caused by excessive flood fracture and also provides important technical support for similar accidents.

Predicting the probability of slope failure caused by rainfall is pivotal to landslide risk assessment and mitigation. Soil properties are inherently spatially variable and often exhibit uncertainties due to a lack of site investigation data, which can lead to inaccurate estimation of probability of slope failure. Recent studies showed that slope performance records (i.e., the slope stays stable or fails after a heavy rainfall) can be used as a valuable source of information to reduce the uncertainties of soil parameters and enhance the estimate of probability of slope failure. This study proposes to successively update the distributions of soil parameters and probability of slope failure by integrating three pieces of slope performance records (i.e., slope keeping stable before rainfall, slope keeping stable after 57 days of weak rainfall, and slope instability after continuous 3 days of heavy rainfall). Bayesian Updating with Subset simulation (BUS) approach is adopted to enhance the computational efficiency of Bayesian analysis. The updated results are then used to predict the probability of slope failure under a future rainfall. The results indicate that the sequential updating greatly reduces the uncertainties in soil parameters and contributes to accurately predict the probability of slope failure under a future rainfall.

Quick clay landslide poses a significant hazard to lives, infrastructure, and the environment in many countries, including Norway. Detecting and mapping quick clay is the most important measure to deal with such hazards in Norway. The rotary sounding method is one of the most effective techniques for detecting quick clay in Norway. There are however inherent uncertainties associated with this approach arising from its interpretation and the number of soundings available for an area of interest. This study introduces a two-step approach for quantifying the uncertainty of quick clay detection using rotary soundings. First, an automated procedure is developed to interpret the rotary sounding data with depth. The interpreted data are subsequently used to infer the probability of quick clay occurrence away from the sounding locations through a hierarchical Bayesian approach. The efficacity of the proposed method is demonstrated based on rotary soundings available at a quick clay test site in Norway.

The simplicity and efficiency of the cone penetration test (CPT) for geotechnical investigation makes it one of the most commonly accepted as well as one of the most employed in situ testing methods worldwide. To predict the most probable soil types and their corresponding spatial uncertainty from CPTs, two geostatistical methods, the Sequential Gaussian Simulation (SGSIM) and Sequential Indicator Simulation (SISIM), are applied in a case study. Hereby, three dimensional (3D) models of the soil behavior index are calculated from the cone penetration resistance and skin friction. The CPTs were carried out to investigate the foundation of a dam in a pre-alpine valley filled with alluvial gravelly to sandy and clayey sediments. The case study demonstrates the ability of the proposed approaches to estimate the spatially distributed uncertainty of soil classification based on CPTs.

This paper proposes to use a shape distribution to compare the soil layer identification ability in data-driven site characterization (DDSC). The proposed approach can represent the shape signature for a 2D/3D distribution of a soil layer as a probability distribution sampled from a shape function measuring geometric properties of the 3D model, and the probability distribution is called shape distribution. There are some options of shape functions, and this study uses D2 shape function (measures the distance between two random points on the surface) because it is easy to compute and robust. The applicability of the shape distribution as a performance metric in DDSC was demonstrated through a 2D simple example. Unlike existing accuracy measures for soil layer identification, a shape distribution can consider the 2D/3D shapes of a soil layer more precisely and can be a useful measure to compare the performance of DDSC methods.

Three-dimensional (3D) subsurface models for precise characterization of complex site conditions have attracted increasing attention recently. Various methods for 3D subsurface modelling have been developed. However, these methods are often validated using synthetic data, rather than real data, for example, cone penetration test (CPT) data. In addition, how to evaluate and compare the performance of these 3D methods in a fair and quantitative manner remains an open question. This may impede the application of 3D modelling methods in engineering practices. To address these issues, this paper presents a benchmarking study to evaluate the performance of 3D modelling methods on real site conditions using real CPT data from different sites. An in-house software called Analytics of Sparse Spatial Data by Bayesian compressive sampling/sensing (ASSD-BCS) is used in the benchmarking study for illustration. ASSD-BCS can directly generate 3D random field samples (RFSs) from sparse measurements. The performance of ASSD-BCS is evaluated in terms of accuracy, uncertainty, robustness, and computational efficiency. The evaluation results show that ASSD-BCS provides accurate prediction results with quantified uncertainty. More importantly, ASSD-BCS has remarkably high computational efficiency and performs robustly.

In this paper, a data-driven site characterization method called the sparse Bayesian learning (SBL) method previously proposed by the author is benchmarked by a set of virtual ground examples and a real ground example of cone penetration test (CPT) data. The SBL method assumes a zero-mean prior Gaussian random field model for the spatial trend modeled by sparse basis functions. The accuracy of the SBL method in predicting the cone tip resistance (q_t) of CPT is quantified by the root-mean-square prediction error (RMSE), whereas the accuracy in identifying soil layers is quantified by the identification rate (IR). The performance of SBL is compared with that of the GLasso method. It is found that SBL does not always outperform GLasso, and GLasso does not always outperform SBL, either. Nonetheless, SBL requires less computational cost than GLasso.