ABSTRACT
Predictions are at the heart of rational design methods in geotechnical engineering. Because of the limited understanding of site conditions, bias from imperfect models, measurement error, and construction effects, the predicted behavior is expected to deviate from the measured behavior. A quantification of this deviation is useful for decision-making in analysis and design, even if it is approximate. This chapter presents an extensive review of the variability of predictions in geotechnics found in three common sources: (1) statistical analysis of prediction variability from a performance database with a consistent definition for the measured and predicted response; (2) lessons from prediction events reflecting the variability of participants’ judgment and experience in the selection of model and related parameter values; and (3) comparison of numerical modeling with field measurements that may be affected by site variability and error from the selection of model and parameter values. The variability arises from different causes in each data source. A comparison can provide deeper insights into how predictions are affected by the ground conditions, the installation of geotechnical structures, the modeling of ground–structure interactions, and the decisions made by the modeler (the engineer). The geotechnical structures covered in this chapter can be broadly divided into three groups based on the role of the soil in the predicted response: (1) foundation engineering (soil provides the resistance against external loading); (2) slope, excavation, and underground structure (soil acts as a load and a resistance); and (3) embankment constructed by soil (soil as a construction material). The model factor statistics presented in this chapter can be applied to satisfy Clause 2.4.16(P), EN 1997-1:2004 (CEN 2004) - “any calculation model shall be either accurate or err on the side of safety” in the probabilistic sense of ensuring Prob(model factor <1) < 0.05. The original deterministic intent of this clause cannot be fulfilled given the variability of predictions demonstrated in this chapter.
