Artificial neural networks for thermochemistry incorporating conservation laws

Abstract

The real time integration of the chemical kinetics in reacting flows consumes the majority of the computational time required when using many simulation techniques. Tabulation of the chemical kinetics using artificial neural networks (ANNs) has the potential to alleviate this computational bottleneck. ANN approaches should have high prediction accuracy but remain as general as possible in order to increase the range of applications and so the potential speed-up. Additionally, it is desirable for ANNs to be physically consistent with the chemical kinetics they replace.

In this thesis the generality of ANN approaches is advanced using a random data generation method, which takes data from canonical combustion problems and modifies it to expand the composition space covered. ANN accuracy is improved by using a quasi-second order training algorithm, an automatic approach to regularisation and by predicting the reaction behaviour of each species individually. The accuracy of the ANN tabulation method is demonstrated via its application to large eddy simulations using transported probability density function closure for the filtered reaction source terms, with numerical solution using stochastic fields.

Statistics obtained from two such simulations of a non-premixed \ce{CH4}/air jet flame with a thirty one species mechanism using conventional real time integration and the ANN tabulation method respectively, show excellent agreement for major species and good agreement for minor species, while the time required for reaction source term calculations is reduced eighteen fold. The range of applications of ANN thermochemistry tabulation is then expanded by simulating turbulent swirl-stabilised bluff-body \ce{CH4}/air premixed jet flames with the same thirty one species mechanism. Recirculation zones near the bluff-body in these flames require very high prediction accuracy to prevent the accumulation of prediction errors affecting downstream predictions, particularly for minor species. The required prediction accuracy is therefore obtained by automatically dividing the output ranges for each species by their magnitudes and using separate ANNs for each. Excellent agreement is observed between statistics collected from simulations using the ANN method and those obtained when using conventional integration, with a reaction source term calculation speed-up of fourteen times.

A method for ensuring physical consistency by enforcing the conservation laws when using ANN thermochemistry tabulation is then developed and applied to simulations of methane/air and propane/air flames. The results of these simulations show that enforcing the conservation laws positively impacts the accuracy of ANN predictions while incurring an insignificant extra computational cost. This conservation method can be readily appended to existing ANN architectures and also be applied to data generation and the prediction of reaction rates.