Description

Module 1

In this module, a neural network is trained involving the following steps:

• Defining the setup for a multi-step outflow-experiment and deriving the initial and boundary conditions for numerically simulating this experiment.

• Executing a large number of numerical flow experiments for a broad range of soil-hydraulic parameters for the Mualem/van Genuchten soil-model.

• Building the database for training the network. The training set consists of pairs of corresponding input and output. In RIAN, the input is formed by the simulated flow pattern (time series of water contents, pressure heads and/or cumulated outflow); the corresponding output is built by the soil-parameter combination which was used to generate the flow pattern (input part).

• Condensing the training database by eliminating redundant information

• Training the network.

• Verifying the network.

Modul 2

This module controls the routine applications of the trained network. The flow experiments must be carried out for the same experimental setup as used for training the network in Module 1. The measured time series of water contents, pressure heads and/or cumulated outflow are then presented to the trained network, which in turn identifies the corresponding soil-hydraulic parameters.