Investigating inconsistent uncertainty quantification encountered in neural network modelling of nonlinear response of a laminar flame

Some investigations on neural network flame modelling report contradictory results. Conclusions on the performance and uncertainty are disparate, despite using an identical dataset for model derivation. Contrasting approaches to data preprocessing are determined as one of the major differences between the studies. Shuffling of the time series data is identified as a probable reason for sub-optimal results of some previously designed networks. Incorporating shuffling is shown to produce training, validation and test sets of extremely high similarity. It is illustrated that a close-to-linear dependence between examples from different sets is the cause for this effect. A'reference test set' is introduced to visualise that shuffling leads to a loss of overtraining indication from the default loss functions, increasing the chance of producing overfitting models. We conclude that for flame models, that rely on a history of velocity perturbations, shuffling before performing the data split into training, validation and tests sets is detrimental for the network design process.