Evaluation and Development of Liquefaction Occurrence and Consequence Analytics Driven by Emerging Data and Technologies

Abstract

For 50 years, significant investments have been made to develop models that predict soil liquefaction. Many of the results may be described as either: (i) wholly empirical models that require only geologic or geospatial data, and which are accessible to a broad userbase; or (ii) semi-mechanistic “simplified stress-based” models that are based on in-situ tests, and which are generally limited to use by geoengineers, with cone-penetration-test (CPT) based models currently considered best. While these models are widely used, all to-date were trained on modest datasets (e.g., 250 datapoints) and in general, have not been rigorously tested against unbiased data, or against one other. In addition, relative to predicting liquefaction occurrence, less effort has historically been given to predicting the downstream consequences of liquefaction, such as damage and loss. As a result, decisions about whether, and how, to mitigate liquefaction must generally be made in a heuristic manner. Meanwhile, recent earthquakes in the age of improved satellite remote sensing and reconnaissance have significantly grown the data that could be used to model liquefaction and its effects, including ground settlement, foundation deformation, and insured losses. Moreover, the growth of remote sensing, in combination with emergent machine- and deep-learning algorithms, creates an opportunity to develop a new type of model that provides rapid, regional predictions of seismic impacts without the need to sample individual sites. Thus, the overarching goal of this work is to test and develop liquefaction occurrence and consequence analytics using emerging data and technologies. Namely, through ten objectives, this research aims to: (1) digitize and publicly compile all existing liquefaction case-histories where CPTs were performed, of which there are ~275, and which may be used to train or test CPT-based liquefaction models; and (2) grow the quantity of such cases to ~15,000 using data from earthquakes in New Zealand. These datasets will then be used to: (3) test existing liquefaction-prediction models based on either the CPT or on geospatial data, from which lessons for future improvement are identified; (4) test a new procedure for correcting “thin layer effects” on CPT data in the context of liquefaction model performance; (5) test existing CPT models that predict ground settlement induced by liquefaction; and (6) train new fragility functions that probabilistically predict liquefaction-induced ground failure. Using the data compiled in (2), in conjunction with additional foundation damage surveys and insurance loss assessments from New Zealand, a decision framework for mitigating liquefaction beneath buildings on shallow foundations is (7) proposed; and demonstrated. Using insights from the tests in (3), a next-generation geospatial modelling approach is (8) proposed, trained, tested, and (9) programmed wherein machine- and deep-learning are used to predict below-ground CPT data via above-ground geospatial information. Lastly, existing and newly proposed analytics in (1-7) are (10) programmed in Horizon, a program for analyzing the occurrence, consequence, and mitigation of liquefaction.