Soft computing techniques
Artificial neural networks (ANN) are data modelling tools that are increasingly used in civil and geotechnical engineering because of their ability to model complex relationships between inputs and outputs without a theoretical model.
The aim of this project was to study the capability of ANN to simulate the output of a problem where it was possible to compare the results with an existing numerical model. The undrained capacity T-max of suction anchors in soft clay was chosen as a suitable problem.
The input parameters used in this study are a fixed anchor diameter of 5 m, a fixed linearly increasing static direct simple shear strength (suDSS = 5 + 1.5 x z), the height of the anchor, ratios between anisotropic cyclic shear strengths and the static direct simple shear strength, the effective unit weight of the soil, the interface roughness at the outer skirt wall and the loading angle at the pad-eye. The pad-eye is assumed to be located at the optimum position. The data base with relationships between the input data and the capacity were generated using HV-Cap. The relationships were in this case not perturbed by random variations in the capacities.
The main goal of the exercise was to asses ANN¿s ability to ¿learn¿ and ¿intelligently replicate¿ the complex multi-parameter model represented by HVCap. In geotechnical engineering problems, available data sets from measurements are most often limited in size. It was therefore of interest to evaluate the performance for varying size of the dataset used to train the network. Based on this a procedure to obtain good training sets of any desired size, was conducted.
The planning and calibration of a neural network analysis involves a considerable number of choices and subjective assessments by the user. Operationally, the exercise comprised a number of sequential phases, namely:
- definition of the reference problem
- compilation of reference database
- ANN configuration and training
- ANN simulations
- assessment of results.
The report (20071084-1) was structured consistently with the above described operational phases. The left figure above illustrates an example output of ANN and HVCap results; the adherence to the equality line attests for the very good performance of the network in replicating HV-Cap calculations. The right figure provides examples of the variation of the anchor capacity T-max with varying anchor penetration depth H and for different values of the other input parameters.