Approaches for a New Generation of Fast-Computing Catalyst Models

In the present study, three different approaches to model a SCRF disregarding the filtration effects were developed and tested against each other in terms of accuracy and calculation time. A high-dimensional, state-of-the-art, physicochemical axisuite Ⓡ $^{{\circledR }}$ model was used as reference and for the generation of training data. The examined approaches were a low-dimensional physicochemical model, a stack ensemble of machine learning models and a neural network. The training data, produced on a virtual test bench, consisted of 39 simulations of synthetic driving cycles which comply with the current RDE legislation. The test data includes simulations of a synthetic and a non-synthetic RDE cycle, a WLTC, and a NEDC. The chosen cycles differ significantly in terms of similarity to the training data, making it possible to examine the generalization capability of the models and the dependency on the data. The results clearly show the superiority of the neural network approach in terms of accuracy, training time, and simulation time. The neural network approach reached an average explained variance score of 0.98 followed very closely by the ensemble (0.96) and the low-dimensional approach (0.76). The runtime was compared based on the computational time required to simulate one real time second. The reference model needed 23 ms. The ensemble model was not able to achieve a major speedup. The low-dimensional model was five times faster and the neural network 450 times faster than the reference model.