Explicit thermodynamic properties using radial basis functions neural networks

Gas turbine design, development, monitoring and maintenance are widely based on numerical simulations of the steady and transient engine performance. Most of the equations that are solved in the simulation programs involve the thermodynamic properties of the fluid flowing through the engine. These properties depend on temperature, pressure, humidity and fuel dosage. As the solution of chemical equilibrium is not compatible with real-time computations, a chemical solver is used off-line to generate a large database which neural networks are trained on. These networks are built on radial basis functions such as multiquadrics. A forward selection approach is used to select data points from the training set as the centers of the transfer functions. The selection stops when the prediction error starts growing. The resulting networks for specific heat and enthalpy of the gas mixture are 3 orders of magnitude faster than the chemical solver. In order to further increase the efficiency and the generalization capabilities of the model, an external optimization solver has been used to tune the shape of the transfer functions. Several solutions are proposed and preliminary results are presented.