Semi-empirical models based on in situ geotechnical tests have been the standard-of-practice for predicting soil liquefaction since 1971. More recently, prediction models based on free, readily available data have grown in popularity. These “geospatial models” rely on satellite remote sensing to infer subsurface traits without in situ tests. While the concept of such an approach is not new, the recent models of Zhu et al. [2015; 2017] are arguably the most rigorously formulated and well-trained to date. The use of such models is appealing for a range of applications, but these models have not been evaluated using independent datasets, nor have they been tested against more established geotechnical methods. These independent evaluations are important for community acceptance and for identifying pathways to improve the models via future research. In other words, when the geospatial models perform poorly, what do they miss that geotechnical models do not? Moreover, the physical damage and monetary loss from liquefaction are arguably more important than the probability of liquefaction occurring. The extension of geospatial models to predict the consequences from liquefaction is both enticing and consistent with the objectives of PEER. Accordingly, the presented study has two main components.
First, using 15,222 liquefaction case histories from 24 earthquakes, the performance of 23 models based on geotechnical or geospatial data are assessed using standardized metrics. Uncertainty due to finite sampling of case histories is accounted for and used to establish statistical significance. Geotechnical predictions are found to be significantly more efficient worldwide, yet successive models proposed over the last twenty years show little or no demonstrable improvement. In addition, geospatial models outperform geotechnical models for large subsets of the data—a provocative finding given the relative time and cost requirements underlying these predictions. Comparisons between geotechnical predictions versus geospatial models provide key insights into improving geospatial models.
Second, while geospatial models have limitations that can and should be addressed via future research, their capacity for predicting liquefaction is promising, as demonstrated herein for certain events. Accordingly, functions are developed to extend the use of the Zhu et al. models [2015; 2017] to predict: (1) severity of liquefaction ejecta; (2) magnitude of ground settlement; and (3) infrastructure damage and loss. Each of these efforts utilizes a subset of data for which geospatial models performed well in the first part of the study (i.e., the locations studied are those where geospatial models, in general, correctly predicted the occurrence and non-occurrence of liquefaction). These analyses thus represent a best-case scenario for predicting liquefaction consequences using geospatial models.
With respect to infrastructure damage and loss, this study focuses on structures built atop shallow foundation systems, which are the most common worldwide. Utilizing damage-survey data and insurance loss assessments for 62,000 such assets, functions are developed to predict liquefaction-induced damage conditioned on the Zhu et al. models. It is shown that while geospatial models are relatively useful for predicting some modes of damage (e.g., global settlement of foundations), they appear not to capture other significant and very costly modes (e.g., stretching, twisting, and separation of foundations). These modes of failure are presumably dependent on asset- and site-specific details that geospatial models do not consider. As an example, differential settlements cause structural distortion and are typically very costly, but such settlements are likely correlated to subsurface variability, which geospatial models cannot predict. Due to these limitations (i.e., the inability to predict all modes of damage), currently geospatial models are poor predictors of monetary loss at the site and neighborhood scales. Efforts to predict damage and loss at coarser scales (e.g., on a per-earthquake basis) may be more fruitful.
The most salient conclusions of this study are summarized as follows:
- Geospatial models demonstrate provocative potential for predicting the occurrence and severity of surficial liquefaction manifestations in the free field. Moreover, these models outperform geotechnical models (which are far costlier and time-consuming to implement) for the large subsets of the data analyzed.
- However, geospatial models were significantly less efficient on a global scale (i.e., when considering case histories worldwide) and provided results that were much closer to random guessing as opposed to accurate predictions. This highlights the inherent difficulty of predicting what is below the ground using only information from above the ground. Efficient geospatial models may be developed for certain locales, but the development of a single model that efficiently predicts subsurface traits across various seismological, geological, geomorphic, and climatic settings is inherently challenging. Given these findings, the global “portability” of geospatial models must be improved and considered a future research priority. Results from the testing of geotechnical vs. geospatial models provide useful insights for model improvement. Specific lessons and pathways for achieving these improvements are presented herein.
- Functions to predict infrastructure damage and loss using geospatial models are presented. These functions are detailed within the report and were developed for several specific types of shallow foundation and for several specific modes of foundation failure. In addition, functions combining all foundation types and all modes of failure were developed. These broadly applicable functions do not require asset-specific information (i.e., the specific type of foundation) and attempt to predict the severity of damage independent of the associated failure mode. Although these functions may be more desirable for general, region-scale analyses, the performance/utility of the developed functions is generally poor, regardless of whether asset-specific information is available. This may be attributable to the fact that some failure modes appear to be strongly dependent on “meso-scale” details (e.g., building geometry, construction quality, subsurface variability, etc.) that are inadequately captured by “macro-scale” geospatial data. The functions developed herein to predict loss are especially ineffective, due to accurate loss prediction being strongly dependent on accurate damage prediction.
Lastly, considering (a) the relatively poor performance of geospatial models globally; (b) the relative inability of the models to predict the consequences of liquefaction (even when the models efficiently predict liquefaction); (c) the primary applications of the models (e.g., post-earthquake reconnaissance and response; regional simulations); and (d) the seminal state of geospatial model-development, it is the authors’ opinion that near-term investment should focus on model improvement rather than model extension. Research that improves the capacity of geospatial models to predict liquefaction (e.g., via development of new models or modification of existing models) is likely to be more impactful than research to adapt existing models for prediction of downstream consequences. Although geospatial liquefaction models have demonstrated surprising and provocative potential, there remains significant room for improvement.
Full List of PEER Reports click here.