Enhancing computational fluid dynamics modelling using intelligent systems technology

Computational Fluid Dynamic simulations take a long time to run, yet the results that these simulations produce is required quickly in user interfaces and optimisation routines. One solution to this is preprocessing the simulations and using this as the basis of some interpolation methodology, such as neural networks. However creating such a system is not a simple task. A trade off needs to be made between the quality of the simulations run and thus the speed at which simulations can be generated; the number of simulations run and the time allocated to gathering data; the amount of inputs to interpolate over and the complexity of the system. This is complicated by the fact that simulations are difficult to run and monitor and thus can be hard to generally automate. Furthermore, merely running more simulations is not necessarily useful: the parameters used to run a simulation are critical to improve the interpolation ability of the prediction methodology, especially when dealing with more than a small number of inputs.

This thesis proposes a system, CFDLearner, which is a framework that attempts to break this problem down into a number of modules that then can be independently solved. It deals with the execution, monitoring and extraction of data from simulations, the generation of parameters with which a simulation is run and the learning of the results generated to create an ``expert surveyor' which is capable of quickly estimating the results gained from the simulations. Furthermore a visualisation tool is generated which provides a means of assessing the confidence of the interpolated results and allows for the optimisation of outputs. To accomplish this task, new techniques are proposed in the fields of neural networks, parameter selection and knowledge based systems.

CFDLearner has been used with the Fire Dynamics Simulator to simulate single compartment house fires. The advantages of the novel neural network methodologies used and the adaptive parameter selection scheme are presented. Furthermore the utility of CFDLearner is demonstrated in its graphical ability to interactively allow users to simply navigate through hundreds of simulation results with multiple input and output parameters. This allows non-expert users to simply obtain useful information about complex simulations without having to understand the underlying complexity that it takes to generate and collate such data.