(190e) Hybrid Modeling of Bioreactor Systems Using First Principles and Deep Neural Networks with Constraints of Validity Domain for Optimization of Feeding Strategy

Bioreactors in industry often grow genetically engineered cells and their operations often switch between different modes. Hence, identification of a reliable model can offer the potential for improving the economic performance and operational sustainability of the bioprocess. However, finding a proper model structure and identifying its parameters is an onerous task for practical applications because there is still a lack of fundamental understanding of the biological mechanisms governing the process. In addition, the existing growth kinetic models are mainly designed for wild-type cells and have a limited range of validity only to describe particular dynamics such as the exponential growth.

A framework combining first principles and data-based models to describe unknown kinetics has been studied as a remedy to the problems in the field of bioprocess optimization. In general, hybrid modeling approaches use first principles models, e.g., mass balances, as the main structure of the model, and data-based models to estimate kinetic parameters in net generation terms [1] [2]. With the component of first principles modeling, this hybrid model can be potentially exploited beyond the region of identification data and is amenable to physical interpretations. With ongoing advances in sensor technology, massive amounts of operational data are becoming available for industrial processes. The data sets often include different types of measurements such as optical density, light intensity, concentrations, etc. under different operating conditions from various strains of a cell. These characteristics make it difficult to analyze the bioreactor system using a simple nonlinear regression model like feed-forward neural networks.

This study proposes a hybrid modeling approach that combines a first principles model with deep neural networks (DNN) that can utilize the complex data and describe dynamic behaviors of a bioreactor system. The proposed approach first compresses the measurements using principal component analysis (PCA) and then feed the compressed data into the DNN model. The projected variables can reduce the dimension of input, remove the feature extraction steps, and make DNN learn faster. It is also important to identify valid domains for the model prediction if the hybrid model is to be used for other tasks such as optimization. The proposed approach includes criteria for checking “valid domain” of the hybrid model to assess whether the prediction is from undue extrapolation or not. The criteria are represented as convex hull and confidence interval [3] so that they can be readily integrated into real-time optimization techniques such as model predictive control. As an illustration of industrial operations, feed-rate optimization of a fed-batch bioreactor will be presented and show the efficacy of the proposed scheme.

[1] von Stosch, Moritz, et al. "Hybrid semi-parametric modeling in process systems engineering: Past, present and future." Computers & Chemical Engineering 60 (2014): 86-101.

[2] Oliveira, R. "Combining first principles modelling and artificial neural networks: a general framework." Computers & Chemical Engineering 28.5 (2004): 755-766.

[3] Kahrs, O., and W. Marquardt. "The validity domain of hybrid models and its application in process optimization." Chemical Engineering and Processing: Process Intensification 46.11 (2007): 1054-1066.