"A data-driven fuzzy approach to simulate the critical shear stress of mixed cohesive/non-cohesive sediments"The critical shear stress of cohesive and mixed cohesive/non-cohesive sediments is affected by multiple interacting physical, chemical and biological parameters. There are various mathematical approaches in the scientific literature for computing critical shear stress. However, processes that influence sediment stability are still not fully understood, and available formulas differ considerably. These discrepancies in the literature arise from random system behaviour (natural variability of the sediments), different definitions of the critical shear stress, different measurement techniques and different model frameworks (scope of the parameters, undisturbed versus artificial sediment samples). While analytical approaches fail to address the involved uncertainties, fuzzy logic-based models integrate uncertainty and imprecision.
Materials and methods:
With this in mind, a data-driven neuro-fuzzy model (ANFIS) was used to determine the critical shear stress based on sediment characteristics such as wet bulk density and grain size distribution. In order to select model predictors systematically, an automated stepwise regression algorithm was applied. The database for this analysis consisted of 447 measurements of the critical shear stress originating from 64 undisturbed sediment samples.
Results and discussion:
The study identified clay content as the primarily controlling variable for erosion resistance. Depending on the characteristics of the sampling location, the bulk density was also selected as a model predictor. In comparison to analytical models that are available in the scientific literature, the fuzzy model achieved higher correlation coefficients between measured and predicted data.
The neuro-fuzzy-model includes uncertainties of input variables and their interactions directly. Thus, it provides a reliable method for the prediction of erosion thresholds of cohesive/non-cohesive mixtures. It was also shown that this approach requires fewer measured variables as well as fewer assumptions than the models it was compared to.