Two substantially different methods have become popular in building fast computing catalyst models: physico-chemical approaches focusing on dimensionality reduction and machine learning approaches. Data driven models are known to be very fast computing and to achieve high accuracy but they can lack of extrapolation capability. Physico-chemical models are usually slower and less accurate but superior regarding robustness. The robustness can even be reinforced by implementing an extended Kalman filter, which enables the model to adapt its states based on actual sensor values, even if the sensors are drifting.
The present study proposes a combination of both approaches into one hybrid model, keeping the robustness of the physico-chemical model in edge cases while also achieving the accuracy of the data based model in well-known regimes. The output of the hybrid model is controlled by an autoencoder, utilizing methods well known from the field of anomaly detection.
It was validated that the autoencoder has comparable extrapolation properties as the neural network representing the catalyst. Hence, the reconstruction error of the autoencoder can be used as an indicator for the data driven model’s performance. The physico-chemical model is applied when the reconstruction loss surpasses a chosen threshold and a low performance of the machine learning approach is anticipated due to extrapolation.
The present study proves that an increase of accuracy and robustness can be achieved by the combination of both modeling approaches.