Artificial neural network for bioprocess monitoring based on fluorescence measurements: Training without offline measurements

Abstract

The feasibility of using a feed-forward neural network in combination with 2D fluorescence spectroscopy to monitor the state of Saccharomyces cerevisiae fermentation was investigated. The main point is that for the backpropagation training of the neural network, no offline measurement value was used, which is the ordinary approach. Instead, a theoretical model of the process has been applied to simulate the process state (biomass, glucose, and ethanol concentration) at any given time. However, the kinetic parameters of the simulation model are unknown at the beginning of the training. It will be demonstrated that the kinetic parameters of the theoretical process model as well as the parameters of the feed-forward neural network to predict the process state from 2D fluorescence spectra can be acquired from the 2D fluorescence spectra alone. Offline measurements are not actually required. The resulting trained neural network can predict the process state as accurate as a conventionally (with offline measurements) trained neural network. The calculated parameters result in a simulation model that is at least as accurate as a model with parameters acquired by least squares fitting to the offline measurements.

Keywords: Bioprocess monitoring; Fluorescence spectroscopy; Neural network; Saccharomyces cerevisiae.

Figure 1

Structure of the applied neural network. The spectra consisted of 120 channels. Two neurons in the hidden and three neurons in the output layer were used to predict biomass (X), glucose concentration (G) and ethanol concentration (E) of the cultivation process.

Figure 2

Flowchart of the model‐based training procedure to get optimal model parameters as well as a trained neural network to predict state variables without the need of offline measurements.

Figure 3

Validation error (value of objective function) of the neural network trained by the simulation of the S. cerevisiae cultivation and two thirds of the spectra from the actual cultivation. The validation error is minimal if the assumed specific growth rates for the simulation are close to their correct values.

Figure 4

Neural net prediction of the three S. cerevisiae fermentations. The spectra from the second (middle) fermentation in combination with the simulation were used to train the neural network.