Hybrid neural‐networks modeling of an enzymatic membrane reactor

Abstract

Many complex biochemical and chemical processes are suitable candidates for modeling via artificial neural‐networks (ANN) but the black‐box approach of ANN may limit the model's ability to extrapolate beyond its training data. In recent years, a major effort has been launched to develop hybrid artificial neural networks (HANN), in which a degree of deterministic first‐principle approach is integrated into the black‐box ANN to achieve the best of both. In this paper, two HANN models were developed for a steady‐state continuous‐flow enzymatic tubular membrane reactor used for saccharification of cellulose to glucose and cellobiose. These HANN models were found to perform much better than a pure ANN model with no first‐principle component. The improvements were particularly significant when extrapolations beyond the sets of training data were involved. It was also found that, in hybrid neural‐networks modeling of a bioreactor, incorporation of fundamental but flexible information on bio‐reaction rates beyond mass‐balance relations could lead to significant improvement in the performance of a HANN model.