The target for this project is the development of NNs enhanced with chemical and physical information. Therefore, different model architectures were tested. Due to large number of hyperparameters for the training process and the model itself, parallel computation is mandatory to optimize the neural network. It was found that the usage of chemical constraints leads to more robust models. These constraints can include e.g. mass balance or energy conservation. A schematic framework of the models is shown in Figure 1. By including chemical information, it is possible to use smaller networks with only 3 layers containing each 25 nodes to represent complex correlations.
The graphs depicted in Figure 2 show automotive driving cycles before and after a Selective-Catalytic- Reduction (SCR) converter. The goal is to convert toxic species like nitrogen oxides (NOx) to harmless nitrogen (N2). The reduction is enabled by adding ammonia (AdBlue®) before the catalyst. The trained model represents several reactions taking place on the active sites including the surface coverage of ammonia, which is crucial for the conversion of the nitrogen oxides. The challenge for those driving cycles is the transient behavior. These transient cycles could be predicted with the newly developed models with deviations around 1% regarding the original training data.