Date of Award
Doctor of Philosophy (PhD)
Ronald D Andrus
Soil liquefaction is a process that saturated soils lose stiffness and strength due to the generation of pore water pressure under rapid earthquake loading and behave like a liquid. Earthquake-induced liquefaction could generate two interrelated hazards on liquefiable sites (belonging to NEHRP F class): excessive ground deformations and unfavorable ground motions. Liquefaction can significantly modify the surface ground motions through changing the material properties of subsurface soils, and the modified ground motions can affect the seismic response of the liquefied soils. Moreover, the site response on liquefiable sites may change from non-liquefied to liquefied response, and be subjected to environmental changes which can affect the seismic properties of shallow liquefiable soils. A critical step to improve seismic hazard analysis on liquefiable sites is to incorporate the liquefaction effects on site response to predict the liquefaction-induced deformation and liquefaction-affected ground motions.
The main objective of this dissertation is to approach an accurate characterization of site responses in liquefiable sites and to develop and validate novel methodologies for liquefaction triggering assessment and ground motion prediction using high-quality ground motion data. Revolving around this main objective, four studies are presented: (1) An accelerogram-based liquefaction triggering assessment; (2) Quantification of nonlinear site responses in liquefiable sites; (3) Coseismic and temporal changes of seismic responses in liquefiable sites; (4) Non-ergodic ground motion prediction using machine learning techniques.
The first study (Chapter 3) presents an accelerogram-based framework for the quick assessment of liquefaction occurrence based on ground motion records only. To separate the ground motions from liquefied and non-liquefied stations, two frequency-related ground motion indices, termed RL and MIFr, are defined and extracted from accelerograms using signal processing techniques. RL and MIFr indicate the richness of the low-frequency components and the temporal variation rate of the mean instantaneous frequency in the ground motion records, respectively. A new liquefaction database consisting of paired ground motions and liquefaction observations is compiled. Logistic regression is used to develop a new liquefaction classification model that takes RL and MIFr as inputs and calculates a liquefaction indicator (LQI) that can be used to assess liquefaction occurrence. The proposed method shows superior performance than other accelerogram-based approaches and has promising potentials for applications in real-time disaster mitigation systems. A follow-up study (Chapter 4) compares the effects of liquefaction and other source, path, and site effects on the RL and MIFr of site-specific ground motions. The results indicate subsoil liquefaction has generally dominant effects on the RL and MIFr values of surface ground motions. But special cautions need to be paid when applying the proposed accelerogram-based method for either one of the following two conditions: 1) far-field (g 200 km) strong ground motions that may have high RL values because of increasing amount of long-period surface waves due to the basin effects; 2) moderate (5 < M < 6) earthquake induced ground motions that may have high MIFr values because of the relatively fast fault rupture process.
In the second study (Chapter 5), paired surface and downhole ground motions from four selected liquefiable vertical arrays are processed to isolate the site effect from other source and path effects. The borehole spectral ratios computed using a large amount of weak ground motions are used to represent the linear site response with consideration of the inter-event variability for each array. Two parameters representing the overall effects of nonlinear soil behaviors on site response, the percentage of nonlinearity (PNL) and the frequency shift parameter (fsp), are used to quantify the modification of site response of strong events to that of weak events. The results indicate the resonance frequency shifts to lower frequency and the amplification decreases for the strong but non-liquefied event, showing a general nonlinear site response. Soil liquefaction can significantly change the shape of the borehole spectral ratio curves and the liquefaction effect varies site by site. The liquefaction on loose sandy site tends to amplify the low-frequency ground motions and de-amplify the high-frequency ground motions. The dense sandy site can amplify both the low-frequency and high-frequency ground motions while de-amplify the intermediate frequency ground motions. These different frequency-related amplification characteristics also lead to different amplification on PGA.
The third study (Chapter 6) investigates the temporal changes of site response characteristics of a liquefiable site during the 25-year-long instrument period. A slidingwindow approach is applied to compute the spectral ratio changes during and among the 1,238 earthquake events, which enables time resolution of the site response analysis and seismic property quantification to as small as two seconds (i.e. the window sliding step length) and as large as 25 years (i.e. the instrument period). The results show that the resonance frequency can drop by about 70% due to the soil softening effects while momently increases by about 30% due to cyclic mobility behavior of dense sand. The large drop of resonance frequency is recovered following a two-stage logarithmic trends. The first stage is a rapid recovery to 80% of the reference value within about 1.5 mins, and the second stage is a slow recovery of about 30 mins, which implies the mechanism of soil stiffness recovery process post liquefaction is dominated by the re-consolidation of liquefied soils. Besides, a periodic change of resonance frequency is observed, suggesting the seasonal changes of ground water table may affect site response of shallow soil sites.
In the fourth study (Chapter 7), machine-learning based ground motion models are proposed for ground motion predictions. Among the machine learning models tested, the gradient boosting model is found to have an overall superior performance. The results show that machine learning models integrating with additional explanatory variables can improve the performance of non-ergodic ground motion models even for the problematic site (e.g. liquefiable sites). The proposed method will be particularly useful for long-term seismic hazard analysis for critical projects with seismic observations.
Zhan, Weiwei, "Data-Driven Assessment of Site Responses at Liquefiable Sites" (2020). All Dissertations. 2747.