Database classification for LRFD of spread foundations under uplift loading in cohesionless soils

The load and resistance factor design (LRFD) approach considers the uncertainties of foundation capacity predictions by assessing the probability distribution of the model factors (i.e. model statistics). The model statistics are dependent on the selected load test database, and specific design parameters of the foundation. While previous studies have focussed on the compilation of extensive databases, there has been limited research on the classification of the database in terms of the design parameters that can affect the model statistics. Thus, this study aims to use unsupervised clustering methods to classify the database for uplift load tests of spread foundations in cohesionless soils. Three different clustering algorithms (hard K-means, fuzzy C-means, and self-organising map) were used to categorise the database by inputting the following parameters: the embedment ratio of the foundation, the friction angle of the soil, the normalised foundation width, and the model factors obtained from the calculation models in existing design codes. This study considers an existing database, NUS/SpreadFound/919, along with additional load test data collected from the literature. First, the database was divided into seven data groups, which were classified in terms of the uplift failure behaviour of the foundation (shallow, intermediate, and deep) and the strength parameter of the soil (loose and dense). Second, an example LRFD calibration was carried out, and it was shown that higher resistance factors were yielded when the classified data groups were used, compared to when the whole database was used. This demonstrated that the use of clustering is feasible for classifying the database to obtain more economic and optimal LRFD solutions.